Data enhancement method, device, computing device, chip and computer storage medium

Abstract

The present application discloses a data enhancement method, device, computing equipment, chip and computer storage medium, which belongs to the field of computer vision technology. The method performs iterative data enhancement processing on the image according to the order of each business feature in the feature sequence, wherein, in any data enhancement process, based on the changing features in the data enhancement process and the fixed features in the data enhancement process, the image generated by the previous data enhancement process is data enhanced. Since each data enhancement process performs data enhancement on a changing feature in the image, when the iterative data enhancement process is completed, the enhanced image finally obtained is the result of data enhancement of the image based on multiple business features in the business scenario, so the enhanced image finally obtained can meet the business needs of the business scenario.

Description

Technical Field

[0001] The present application relates to the field of computer vision technology, and in particular to a data enhancement method, device, computing equipment, chip and computer storage medium for image recognition. Background Art

[0002] With the development of image recognition technology, a large number of images are needed as training data in the learning process of image recognition models. However, the existing training data (including public or private training data) cannot fully meet the needs of business scenarios in terms of quantity and quality (which can be reflected in multiple measurement indicators such as diversity, balance, consistency, distinguishability, and learnability of image data). Data enhancement is a solution to this problem, that is, using technical means to effectively change and expand the limited existing training data, increase the quantity and diversity of training data, and thus improve the accuracy and robustness of the trained model.

[0003] At present, there are many single-point data enhancement technologies that change specific features of images. Single-point data enhancement technology can change a specified feature of an existing image to obtain more images with the specified feature changed, thus achieving the purpose of data enhancement to a certain extent. For example, for a real face image, if the specified feature is age, the age of the face image can be changed through the single-point data enhancement technology of age change to obtain face images of different age groups of the same person.

[0004] Since complex business scenarios may require many features, but the computing device can only change one specified feature of the image, and cannot change multiple features of the image at the same time, the image with the specified feature changed cannot meet the needs of complex business scenarios. For example, a company or organization's access control face recognition scenario requires a large number of face data sets for face recognition model training. The set needs to meet the conditions of having a specific skin color (such as African American, Asian), an age range (such as 20-60 years old), different face angles (such as front face, 15 degrees and 30 degrees), different lighting conditions, different expressions, and a combination of features formed by different cross-matching between the multiple features; however, the single-point data enhancement technology can only meet the data enhancement of a single condition of the above-mentioned single feature (such as age), so that the face image obtained by the single-point data enhancement technology cannot directly meet the business needs of complex business scenarios. Summary of the invention

[0005] The embodiments of the present application provide a data enhancement method, apparatus, computing device, chip and computer storage medium, which can ensure that the image after data enhancement can meet the business requirements of complex business scenarios. The technical solution is as follows:

[0006] In a first aspect, a data enhancement method is provided, and the implementation process of the method may be:

[0007] A feature sequence of the business scenario can be obtained first, and then the image can be iteratively data enhanced according to the data enhancement processing order of each business feature, wherein the feature sequence is used to represent the data enhancement processing order of multiple business features under the business scenario, and one data enhancement processing process corresponds to a business feature in the feature sequence; wherein any data enhancement processing process can be: based on the changing features in the data enhancement processing process, data enhancement is performed on the image generated by the previous data enhancement processing, and the changing features are the business features corresponding to the data enhancement processing process.

[0008] The business scenario can be image recognition in any work scenario, such as face recognition in access control scenarios. A business feature is a feature of a real image in the business scenario. For example, in a face recognition business scenario, a business feature of a face image can include any one of skin color feature, age feature, side face angle feature, and expression feature.

[0009] The iterative data enhancement processing may include multiple data enhancement processing with a data enhancement processing sequence. Optionally, in the first data enhancement processing of the multiple data enhancement processing, data enhancement is performed on the image to obtain an image generated by the first data enhancement processing; then, according to the data enhancement processing sequence, the second data enhancement processing of the multiple data enhancement processing is performed, and in the second data enhancement processing, data enhancement is performed on the image generated by the first data enhancement processing to obtain an image generated by the second data enhancement processing; then, according to the data enhancement processing sequence, the third data enhancement processing of the multiple data enhancement processing is performed, and so on, until the last data enhancement of the multiple data enhancement processing is completed.

[0010] Each data enhancement process in the iterative data enhancement process corresponds to a business feature in the feature sequence, and the order of each business feature in the feature sequence can be regarded as the processing order of the data enhancement process corresponding to each business feature in multiple data enhancement processes. The last data enhancement process involved in any data enhancement process is also the data enhancement process corresponding to the target business feature, where the target business feature is the previous business feature of the business feature corresponding to any data enhancement process in the feature sequence. For example, for the Nth data enhancement process in N data enhancement processes with a data enhancement process sequence, the N-1th data enhancement process is also the data enhancement process before the Nth data enhancement process, where N is an integer greater than 1.

[0011] In a possible implementation, based on the change characteristics in the data enhancement process, performing data enhancement on the image generated by the previous data enhancement process includes:

[0012] The feature value of the changing feature in the image generated by the last data enhancement processing is changed, and the fixed feature in the image generated by the last data enhancement processing is kept unchanged, and the fixed feature is the business feature before the business feature corresponding to the data enhancement processing process in the feature sequence.

[0013] The above-mentioned data enhancement method performs iterative data enhancement processing on the image according to the order of each business feature in the feature sequence, wherein, in any data enhancement processing process, based on the changing features in the data enhancement processing process and the fixed features in the data enhancement processing process, the image generated by the previous data enhancement processing is data enhanced. Since each data enhancement processing process performs data enhancement on a changing feature in the image, when the iterative data enhancement processing is completed, the enhanced image finally obtained is the result of data enhancement of the image based on multiple business features in the business scenario, so the enhanced image finally obtained can meet the business needs of the business scenario.

[0014] In one possible implementation, the multiple business features can be sorted based on the number of values ​​that each business feature takes in the business scenario and the number of original values ​​of each business feature, and the sorted result is used as the feature sequence, thereby realizing the above-mentioned process of obtaining the feature sequence of the business scenario.

[0015] In a possible implementation, the process of sorting the multiple service features may be:

[0016] We can first calculate the ratio of the proportion of valid images in the image generated when a business feature is used as a changing feature for data enhancement, and then sort the business features in ascending order according to the size of the ratio corresponding to each business feature, where the ratio of each business feature is the ratio between the number of values ​​of each business feature in the business scenario and the number of original values ​​of each business feature.

[0017] In a possible implementation, before executing the step of sorting the multiple service features, the method may further execute the following steps:

[0018] receiving a data enhancement request and obtaining a comprehensive business feature of the business scenario carried in the data enhancement request, wherein the data enhancement request is used to instruct generation of an image that meets the requirements of the business scenario;

[0019] The comprehensive business characteristics of the business scenario are split and processed to obtain multiple business characteristics of the business scenario;

[0020] The values ​​of each business feature in the business scenario are determined based on the business requirements of the business scenario.

[0021] In a possible implementation, the process of iteratively performing data enhancement processing on the image according to the data enhancement processing order of each service feature may be:

[0022] Based on the order of each business feature in the feature sequence and each node on the data enhancement tree, iterative data enhancement processing is performed on the image. The data enhancement tree includes multiple nodes, and each node corresponds to a data enhancement processing process.

[0023] In a possible implementation, the process of iteratively performing data enhancement processing on the image based on the order of each service feature in the feature sequence and each node on the data enhancement tree may be:

[0024] Based on the order of each business feature in the feature sequence, a data enhancement tree consisting of multiple nodes is established, each business feature corresponds to at least one node, wherein a node corresponding to any business feature is a parent node of at least one target node among the multiple nodes, and the at least one target node corresponds to a business feature after the any business feature in the feature sequence;

[0025] According to the dependency relationship between the nodes in the data enhancement tree, the image is iteratively enhanced.

[0026] In a possible implementation, before executing the step of iteratively performing data enhancement processing on the image according to the dependency relationship between the nodes on the data enhancement tree, the method may further execute the following steps:

[0027] According to the dependency relationship between the nodes on the data enhancement tree, the change feature of each node is determined, so that when data enhancement is performed on each node, the feature value of the change feature of the to-be-processed image of each node is changed.

[0028] In a possible implementation, the process of determining the change feature of each node may be:

[0029] The business characteristics corresponding to each node are used as the change characteristics of each node.

[0030] In a possible implementation, before executing the step of iteratively performing data enhancement processing on the image according to the dependency relationship between the nodes on the data enhancement tree, the method may further execute the following steps:

[0031] According to the dependency relationship between nodes on the data enhancement tree, the changing features and fixed features of each node are determined.

[0032] In a possible implementation, determining the variable feature and the fixed feature of each node according to the dependency relationship between the nodes on the data enhancement tree includes:

[0033] The business characteristics corresponding to each node are used as the change characteristics of each node;

[0034] The fixed features and changing features of the parent node of each node are used as the fixed features of each node.

[0035] In a possible implementation, after performing the step of determining the fixed feature of each node and the variable feature of each node, the method may further perform the following steps:

[0036] A data tag is established for each initial node, and the data tag of a node includes fixed features of the node.

[0037] In a possible implementation, the process of establishing the data enhancement tree may be:

[0038] Based on the order of each business feature in the feature sequence and the deletion information corresponding to the business scenario, the initial nodes on the initial data enhancement tree are deleted to obtain the data enhancement tree, and each business feature corresponds to at least one initial node on the initial data enhancement tree.

[0039] Optionally, an initial data enhancement tree consisting of multiple initial nodes is first established according to the order of each business feature in the feature sequence, and then the initial nodes on the initial data enhancement tree are deleted according to the deletion information corresponding to the business scenario, that is, the data enhancement tree is pruned, so that the pruned initial data enhancement tree is also the data enhancement tree; wherein, each business feature corresponds to at least one initial node, an initial node corresponding to any business feature is the parent node of at least one target initial node among the multiple initial nodes, and the at least one target initial node corresponds to the next business feature of the business feature in the feature sequence, and the deletion information is used to indicate the initial node to be deleted on the initial data enhancement tree.

[0040] In a possible implementation, the method may also establish an initial data enhancement tree, wherein the process of establishing the initial data enhancement tree may be:

[0041] According to the order of each business feature in the feature sequence, at least one initial node corresponding to each business feature is generated in turn to obtain the initial data enhancement tree, and the initial node corresponding to a business feature is the parent node of the initial node corresponding to the next business feature of the business feature.

[0042] Optionally, a root node of the initial data enhancement tree is generated first, and the root node corresponds to the first business feature in the feature sequence; for any business feature other than the first business feature in the feature sequence, a first number of node sets of the initial data enhancement tree are generated according to the value of the business feature in the business scenario and the value of the previous feature of the business feature in the feature sequence in the business scenario, the first number is the number of values ​​of the previous business feature of any business feature in the feature sequence in the business scenario, each node set includes a second number of initial nodes, the nodes in each node set are child nodes of an initial node corresponding to the previous business feature, and the second data amount is the number of values ​​of any business feature in the business scenario.

[0043] In a possible implementation, after executing the step of deleting the initial node on the initial data enhancement tree according to the deletion information corresponding to the business scenario, for the initial node whose parent node is deleted, the deletion of the parent node of the initial node will cause the fixed feature of the initial node to change, therefore, the fixed feature of the initial node is updated to the fixed feature and the changed feature of the current parent node of the initial node, and the fixed feature on the data tag of the initial node is updated to the updated fixed feature of the initial node, wherein the current parent node of the initial node is the previous node of the initial node on the data enhancement path to which the initial node belongs on the data enhancement tree;

[0044] In a possible implementation, the process of iteratively performing data enhancement processing on the image may be:

[0045] For any node among the multiple nodes, data enhancement processing is performed on an image to be processed based on a data enhancement model corresponding to the node, and the image to be processed is an image generated when data enhancement is performed on a parent node of the node.

[0046] Optionally, for any node among the multiple nodes, the image to be processed is input into a first data enhancement model, and the first data enhancement model performs data enhancement on the input image based on the change characteristics of the node, and outputs an enhanced image, wherein the image to be processed is an image generated when data enhancement is performed on the parent node of the node;

[0047] For the child nodes of the node, the image output by the first data enhancement model is input into the second data enhancement model, and the second data enhancement model performs data enhancement on the input image based on the change characteristics of the child nodes.

[0048] In a possible implementation, after performing the step of iteratively performing data enhancement processing on the image, the method may further perform the following steps:

[0049] Based on the weight corresponding to each leaf node, partial images generated when data enhancement is performed on each leaf node are combined into a training set, and each leaf node is a node without child nodes.

[0050] Optionally, for leaf nodes without child nodes, a weight is set for each leaf node in the multiple nodes; and for any leaf node, a fourth number of images is extracted from the images generated by data augmentation at the leaf node based on the number of images required for the training set and the weight of the leaf, and the images obtained from each leaf node are combined into the training set, wherein the fourth number is the product of the number of images required for the training set and the weight of the leaf node.

[0051] In a possible implementation, after executing the step of combining the images obtained from each leaf node into the training set, the method may further execute the following steps:

[0052] Each image in the training set and the data label associated with each image are input into the initial image recognition model. The initial recognition model performs model training based on the input image and the input data label to obtain an image recognition model. The image recognition model is used to recognize images in the business scenario. The data label associated with each image is the data label of the node where each image is located.

[0053] In a possible implementation, after executing the step of obtaining the image recognition model, the method may further execute the following steps:

[0054] Multiple real images in the business scenario are input into the image recognition model, and the image recognition model performs image recognition on the input multiple real images, outputs recognition results for the multiple real images, and based on the recognition results for the multiple real images, determines the recognition effect of the image recognition model in recognizing the multiple real images; if the recognition effect reaches the preset recognition effect, the images in the training set are used as training data in the business scenario, otherwise, the weight of each leaf node is adjusted.

[0055] In a possible implementation, after performing the step of performing data enhancement on the image generated by the previous data enhancement process, the method may further perform the following steps:

[0056] The image generated by the data enhancement process can be filtered first, and then the filtered image that meets the first preset condition is used as the image to be processed in the next data enhancement process. The first preset condition includes at least one of the following: the authenticity of an image reaches a preset authenticity; the similarity between images with the same individual features is greater than or equal to the first similarity; the similarity between images with different individual features is less than or equal to the second similarity.

[0057] Based on the above possible implementation methods, by filtering the images generated by this data enhancement process through the first preset condition, it can be ensured that the images that meet the first preset condition have high authenticity, there is diversity between images of the same object, and there are differences between images of different objects, so that high-quality images that meet the first preset condition can be obtained.

[0058] In a possible implementation, the process of using the first preset image as the image to be processed in the next data enhancement process may be:

[0059] The image that meets the first preset condition is filtered again, and the image that meets the second preset condition that is filtered again is used as the image to be processed in the next data enhancement processing process, wherein the second preset condition includes a feature value of a change feature of an image during the data processing process as the value of the change feature in the business scenario.

[0060] Based on the above possible implementation methods, the images that meet the second preset condition and are obtained by filtering through the second preset condition can meet the requirements of the business scenario. Therefore, the images that meet both the first preset condition and the second preset condition are high-quality images that meet the requirements of the business scenario.

[0061] In a possible implementation, before performing the step of iteratively performing data enhancement processing on the image, the method may further perform the following steps:

[0062] The target number is input into the target data enhancement model, and the target data enhancement model generates the target number of source images based on the input target number; then the target number of source images are filtered, and the filtered source images that meet the third preset condition are used as the images to be processed in the first data enhancement process. Wherein, the third preset condition includes at least one of the following: the similarity between a source image and source images other than the source image in the target number of source images is less than or equal to the third similarity; the similarity between a source image and multiple real images is less than or equal to the fourth similarity, and the multiple real images are used to train the target data enhancement model; the authenticity of a source image reaches the preset authenticity.

[0063] Based on the possible implementation manners described above, a plurality of source images having a high degree of authenticity and being different from real images can be obtained through the third preset condition.

[0064] In a possible implementation, after performing the step of performing data enhancement on an image generated by a previous data enhancement process based on a changing feature in the data enhancement process and a fixed feature in the data enhancement process during any data enhancement process, the method may further perform the following steps:

[0065] The image generated by the data enhancement process is associated with the data label of the data enhancement process, and the data label of the data enhancement process includes the fixed features of the data enhancement process.

[0066] Based on the possible implementation methods described above, by associating the data labels corresponding to the data enhancement process with the images generated by the data enhancement process, it is possible to avoid manual labeling of the enhanced images by humans.

[0067] In a second aspect, a data enhancement device is provided, the device comprising:

[0068] An acquisition module, used to acquire a feature sequence of a business scenario, wherein the feature sequence is used to represent a data enhancement processing order of multiple business features in the business scenario;

[0069] A processing module, used for iteratively performing data enhancement processing on the image according to the order of each business feature in the feature sequence, wherein one data enhancement processing process corresponds to one business feature in the feature sequence;

[0070] Among them, in any data enhancement processing process, data enhancement is performed on the image generated by the previous data enhancement processing based on the change characteristics in the data enhancement processing process, and the change characteristics are the business characteristics corresponding to the data enhancement processing process.

[0071] In a possible implementation, the processing module is used to:

[0072] The feature value of the changing feature in the image generated by the last data enhancement processing is changed, and the fixed feature in the image generated by the last data enhancement processing is kept unchanged, and the fixed feature is the business feature before the business feature corresponding to the data enhancement processing process in the feature sequence.

[0073] In a possible implementation, the acquisition module includes:

[0074] The sorting unit is used to sort the multiple business features based on the number of values ​​of each business feature in the business scenario and the number of original values ​​of each business feature to obtain the feature sequence.

[0075] In a possible implementation, the sorting unit is used to:

[0076] Calculate the ratio between the number of values ​​of each business feature in the business scenario and the number of original values ​​of each business feature to obtain the ratio of each business feature, where the ratio of a business feature is the proportion of valid images in an image generated when the business feature is used as a change feature for data enhancement;

[0077] According to the size of the ratio of each business feature, the multiple business features are sorted in ascending order.

[0078] In a possible implementation manner, the device further includes:

[0079] A receiving module, used to receive a data enhancement request, where the data enhancement request is used to instruct to generate an image that meets the business scenario requirements;

[0080] A splitting module, used to split the comprehensive business features of the business scenario carried by the data enhancement request to obtain multiple business features of the business scenario;

[0081] The first determination module is used to determine various values ​​of each business feature in the business scenario based on the business requirements of the business scenario.

[0082] In a possible implementation, the processing module is used to:

[0083] Based on the order of each business feature in the feature sequence and each node on the data enhancement tree, iterative data enhancement processing is performed on the image. The data enhancement tree includes multiple nodes, and each node corresponds to a data enhancement processing process.

[0084] In a possible implementation, the processing module includes:

[0085] An establishing unit, configured to establish a data enhancement tree based on the order of each business feature in the feature sequence, wherein each business feature corresponds to at least one node, wherein a node corresponding to any business feature is a parent node of at least one target node among the multiple nodes, and the at least one target node corresponds to a business feature after any business feature in the feature sequence;

[0086] A processing unit is used to perform iterative data enhancement processing on the image based on the dependency relationship between the multiple nodes.

[0087] In a possible implementation manner, the device further includes:

[0088] The second determination module is used to determine the change characteristics of each node based on the dependency relationship between the multiple nodes.

[0089] In a possible implementation manner, the first determining unit is configured to:

[0090] The service feature corresponding to each node is determined as the change feature of each node.

[0091] In a possible implementation manner, the first determining unit is configured to determine the changing characteristics and the fixed characteristics of each node based on the dependency relationship between the multiple nodes.

[0092] In a possible implementation manner, the first determining unit is configured to:

[0093] Determining the service feature corresponding to each node as the change feature of each node;

[0094] The fixed features and the changing features of the parent node of each node are determined as the fixed features of each node.

[0095] In a possible implementation manner, the establishing unit is further configured to:

[0096] According to the fixed features of each node, a data tag of each initial node is established, and the data tag of a node includes the fixed features of the node.

[0097] In a possible implementation manner, the establishing unit is used to:

[0098] Based on the order of each business feature in the feature sequence and the deletion information corresponding to the business scenario, the initial nodes on the initial data enhancement tree are deleted to obtain the data enhancement tree, and each business feature corresponds to at least one initial node on the initial data enhancement tree.

[0099] In a possible implementation manner, the establishing unit includes:

[0100] Establishing a subunit, used to establish an initial data enhancement tree according to the order of each business feature in the feature sequence, wherein the initial data enhancement tree includes a plurality of initial nodes, each business feature corresponds to at least one initial node, an initial node corresponding to any business feature is a parent node of at least one target initial node among the plurality of initial nodes, and the at least one target initial node corresponds to the next business feature of the business feature in the feature sequence;

[0101] The deletion subunit is used to delete the initial nodes on the initial data enhancement tree according to the deletion information corresponding to the business scenario to obtain the data enhancement tree, wherein the deletion information is used to indicate the initial nodes to be deleted on the initial data enhancement tree.

[0102] In a possible implementation manner, the device further includes:

[0103] The first generation module is used to generate at least one initial node corresponding to each business feature in sequence according to the order of each business feature in the feature sequence to obtain the initial data enhancement tree, wherein the initial node corresponding to a business feature is the parent node of the initial node corresponding to the next business feature of the business feature.

[0104] In a possible implementation manner, the first generating module is used to:

[0105] For the first service feature in the feature sequence, generating a root node of the initial data enhancement tree;

[0106] For any business feature other than the first business feature in the feature sequence, a first number of node sets of the initial data enhancement tree are generated, the first number being the number of values ​​taken by the previous business feature of any business feature in the feature sequence in the business scenario, each node set including a second number of initial nodes, the nodes in each node set being child nodes of an initial node corresponding to the previous business feature, and the second data amount being the number of values ​​taken by any business feature in the business scenario.

[0107] In a possible implementation manner, the device further includes:

[0108] An updating module, for updating, for an initial node whose parent node is deleted, the fixed features of the initial node to the fixed features and the changing features of the current parent node of the initial node, wherein the current parent node of the initial node is the previous node of the initial node on the data enhancement path to which the initial node belongs on the data enhancement tree;

[0109] The updating module is further used to update the fixed features on the data tag of the initial node to the updated fixed features of the initial node.

[0110] In a possible implementation manner, the processing unit is used to:

[0111] For any node among the multiple nodes, data enhancement processing is performed on an image to be processed based on a data enhancement model corresponding to the node, and the image to be processed is an image generated when data enhancement is performed on a parent node of the node.

[0112] In a possible implementation manner, the processing unit is used to:

[0113] For any node among the multiple nodes, inputting the image to be processed into a first data enhancement model, and the first data enhancement model performs data enhancement on the input image based on the change characteristics of the node, and outputs an enhanced image, wherein the image to be processed is an image generated when data enhancement is performed on the parent node of the node;

[0114] For the child nodes of the node, the image output by the first data enhancement model is input into the second data enhancement model, and the second data enhancement model performs data enhancement on the input image based on the change characteristics of the child nodes.

[0115] In a possible implementation manner, the device further includes:

[0116] The combination module is used to combine partial images generated when data enhancement is performed on each leaf node into a training set based on the weight corresponding to each leaf node, and each leaf node is a node without child nodes.

[0117] In a possible implementation, the combination module is used to:

[0118] Setting a weight for each leaf node in the plurality of nodes;

[0119] For any leaf node, according to the number of images required for the training set and the weight of the leaf, obtaining a fourth number of images from the images generated by performing data augmentation on the leaf node, where the fourth number is the product of the number of images required for the training set and the weight of the leaf node;

[0120] The images obtained from each leaf node are combined into the training set.

[0121] In a possible implementation manner, the device further includes:

[0122] A training module is used to input each image in the training set and the data label associated with each image into an initial image recognition model, and the initial recognition model performs model training based on the input image and the input data label to obtain an image recognition model. The image recognition model is used to recognize images in the business scenario, and the data label associated with each image is the data label of the node where each image is located.

[0123] In a possible implementation, the training module is further used to:

[0124] Inputting a plurality of real images in the business scenario into the image recognition model, and having the image recognition model perform image recognition on the inputted plurality of real images, and outputting recognition results of the plurality of real images;

[0125] Based on the recognition results of the multiple real images, determining the recognition effect of the image recognition model in recognizing the multiple real images;

[0126] When the recognition effect reaches a preset recognition effect, the images in the training set are used as training data in the business scenario; otherwise, the weight of each leaf node is adjusted.

[0127] In a possible implementation manner, the device further includes:

[0128] A first filtering module, used for filtering the image generated by the data enhancement process to obtain an image that meets a first preset condition;

[0129] The third determination module is used to determine the image that meets the first preset condition as the image to be processed in the next data enhancement processing process.

[0130] In a possible implementation manner, the first preset condition includes at least one of the following:

[0131] The realism of an image reaches a preset realism;

[0132] The similarity between images having the same individual feature is greater than or equal to the first similarity;

[0133] The similarity between images of different individual features is less than or equal to the second similarity.

[0134] In a possible implementation manner, the second determining module is used to:

[0135] Filter the image that meets the first preset condition to obtain an image that meets the second preset condition, where the second preset condition includes a feature value of a change feature of an image during the data processing being a value of the change feature in the business scenario;

[0136] The image meeting the second preset condition is determined as the image to be processed in the next data enhancement processing process.

[0137] In a possible implementation manner, the device further includes:

[0138] A second generating module is used to input the target number into a target data enhancement model, and the target data enhancement model generates source images of the target number based on the input target number;

[0139] A second filtering module, used for filtering the target number of source images to obtain source images that meet a third preset condition;

[0140] The fourth determining module is used to use the source image that meets the third preset condition as the image to be processed in the first data enhancement processing process.

[0141] In a possible implementation manner, the third preset condition includes at least one of the following:

[0142] The similarities between a source image and source images other than the source image in the target number of source images are all less than or equal to a third similarity;

[0143] The similarities between a source image and a plurality of real images are all less than or equal to a fourth similarity, and the plurality of real images are used for training to obtain the target data enhancement model;

[0144] The realism of a source image reaches a preset realism.

[0145] In a possible implementation manner, the device further includes:

[0146] An association module is used to associate the image generated by the data enhancement process with the data label of the data enhancement process, and the data label of the data enhancement process includes the fixed features of the data enhancement process.

[0147] In a third aspect, a computing device is provided, which includes a processor and a memory, wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement operations performed by the above-mentioned data enhancement method.

[0148] In a fourth aspect, a computer storage medium is provided, wherein at least one instruction is stored in the storage medium, and the instruction is loaded and executed by a processor to implement the operations performed by the above-mentioned data enhancement method.

[0149] According to a fifth aspect, a computer program product comprising instructions is provided, which, when executed on a computing device, enables the computing device to execute the method of the first aspect.

[0150] In a sixth aspect, a chip is provided, comprising a processor and a data interface, wherein the processor reads instructions stored in a memory through the data interface to execute the method in the first aspect.

[0151] Optionally, as an implementation method, the chip may further include a memory, in which instructions are stored, and the processor is used to execute the instructions stored in the memory. When the instructions are executed, the processor is used to execute the method in the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0152] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0153] Figure 1 is a schematic diagram of a data enhancement system provided in an embodiment of the present application;

[0154] Figure 2 is a schematic diagram of a data enhancement system provided in an embodiment of the present application;

[0155] Figure 3 is a schematic diagram of the structure of a computing device provided in an embodiment of the present application;

[0156] Figure 4 is a flow chart of a data enhancement method provided in an embodiment of the present application;

[0157] Figure 5 is a flow chart of a data enhancement method provided in an embodiment of the present application;

[0158] Figure 6 is a schematic diagram of an initial data enhancement tree provided in an embodiment of the present application;

[0159] Figure 7 is a schematic diagram of a data enhancement tree provided in an embodiment of the present application;

[0160] Figure 8 is a schematic diagram of data fusion provided in an embodiment of the present application;

[0161] Fig. 9 is a flow chart of a data enhancement method provided in an embodiment of the present application;

[0162] Fig.10 is a structural schematic diagram of a data enhancement device provided in an embodiment of the present application;

[0163] Fig.11 It is a schematic diagram of the structure of a chip provided in an embodiment of the present application. DETAILED DESCRIPTION

[0164] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below in conjunction with the accompanying drawings.

[0165] Figure 1 is a schematic diagram of a data enhancement system provided in an embodiment of the present application, see Figure 1 , the data enhancement system 100 may include a data innovation layer, a data expansion layer, and a data fusion layer from the perspective of system architecture. The data innovation layer is used to generate highly realistic pseudo images based on real images, and the data expansion layer is used to perform data changes on different feature dimensions on the pseudo images generated by the data innovation layer to generate images with features required by different business sub-scenarios, wherein changing the feature value of a business feature of the pseudo image can be regarded as a data change on only one feature dimension of the pseudo image. The data fusion layer is used to combine the newly generated images of different categories after combining the feature dimensions according to a certain ratio to obtain a training data set that meets the requirements of the business scenario, so that the images in the training set can be used for model training in the future to obtain an image recognition model that can perform image recognition on the images in the business scenario. The image types of the images may include person types, animal types, text types, and other types. The images of the person types may be human face images and human body posture images, the images of the animal types may be animal face images and animal body posture images, and the images of the text types may be invoice images, contract images, and advertisement images, etc. Accordingly, the business scenario may be image recognition of person types, for example, face recognition in a door access control scenario. The business scenario can also be animal-type image recognition, for example, by identifying animal-type images, whether the lost animal is a registered animal. The business scenario can also be optical character recognition (OCR), where OCR refers to the process of analyzing and identifying images of text materials to obtain text and layout information. For example, text-type image recognition, such as identifying the amount and invoice date in the invoice image, identifying the text and trademark in the advertising image, etc., can also be identifying the vehicle model and license plate number in the vehicle image.

[0166] The data innovation layer may include a first computing device 101, and the first computing device 101 is used to generate multiple new source images with different individual features, so that the data expansion layer can perform data enhancement on the source image. Therefore, the source image can also be called "source data", and the data innovation layer is also called the source data generation layer. Each source image corresponds to a unique object, and each source image can display its corresponding object. The individual features of each object are certain, and the individual features of different objects are different. The individual features of an object are the most basic features of the object. Taking the object as a person as an example, the individual features of a person's face may include the outline of the person's face, the shape of the eyes, the shape of the mouth, the shape of the nose and other local features. Therefore, the individual features of multiple images of the same object are the same, while the individual features of images of different objects are different. For example, two facial images 1 and 2 of person A with different expressions, since these two images are the faces of the same person, the individual features displayed by facial images 1 and 2 are the same. Face image 3 of person B and face image 1 of person A are face images of two objects respectively, so the individual features displayed by face image 3 and face image 1 are different. The individual features of the source image are different from those of the real image, and the individuals in the source image also have different individual features, but the image type of the source image is the same as that of the real image, that is, a pseudo image of a highly realistic real image. Therefore, the source image generated by the first computing device 10 is an image of a different object, which is a virtual object and not an object that exists in reality, while the real image shows a target object, which is an actual physical object. Taking the source image as a facial image as an example, the facial image is not a facial image of an actual person, but is only a forged facial image. Taking the real image as a facial image as an example, the real facial image is a facial image of an actual person.

[0167] The first computing device 101 may include a source data generator (SG). When the data input interface in the first computing device 101 receives the required number of source images, the data input interface may input the received required number of source images to the SG, and the SG generates a plurality of source images based on the input number of source images, wherein the number of generated source images is the same as the required number of source images input into the SG, and the plurality of source images correspond to different objects, each source image corresponding to an object. Taking facial images as an example, 10 source images are facial images of 10 different objects, that is, 10 different faces.

[0168] In some cases, a small number of source images generated by SG may have quality problems. Taking face images as an example, a face displayed by a generated face image may not look like a face, that is, the face image is not realistic enough; it may also be the case that two or more generated face images are too similar to be distinguished as different individuals. In order to avoid this problem, the first computing device 101 can only retain source images with higher quality so that the data expansion layer can perform image enhancement processing on the source images with higher quality. In a possible implementation, the first computing device 101 may also include a data difference / diversity examiner (DE) and a source data quality examiner (SQE), wherein DE is used to filter out source images with different individual features compared with real images. SQE is used to filter out source images with higher authenticity from the source images filtered by DE, wherein authenticity is used to indicate the real situation of the source image. When the source image is a realistic face image, the authenticity of the source image is higher, otherwise the authenticity of the source image is lower. Of course, SQE can also first filter out source images with higher authenticity from the source images generated by SG, and DE can then filter out source images with different individual features from the real images used by SG from the source images filtered by SQE. The present embodiment of the application does not specifically limit the order in which the first computing device 101 uses DE and SQE. Finally, the first computing device 101 can obtain multiple source images with different individual features from the real images through filtering by DE and SQE, and the authenticity of these multiple source images is relatively high. For example Figure 2 As shown in the schematic diagram of a data enhancement system provided by an embodiment of the present application, SG can generate 100 face images (that is, source images). SQE filters the 100 face images based on their authenticity to obtain 80 face images with higher authenticity. DE then filters the 80 face images again based on the similarities between the 80 face images to obtain 60 face images. The similarity between these 60 face images and between the 60 generated images and the real face images in the training data set used by the face generation model is relatively low. It can also be considered that the 60 face images are brand new face images of 60 individual characters, so that the first computing device 101 can obtain 60 face images with higher authenticity and of different objects.

[0169] The data expansion layer may include a second computing device 102, and the second computing device 102 may determine multiple business features of the image in the business scenario according to the business requirements of the business scenario. For example, if the business scenario needs to generate a face of a black person aged 30-40, the skin color feature and the age feature are two business features of the image. The second computing device 102 may also perform iterative data enhancement processing on the source image finally output by the first computing device 101 based on the values ​​of the multiple business features in the business scenario. Each data enhancement processing process may be to change the feature value of a business feature of the image generated by the last data enhancement processing process to each value of the business feature in the business scenario, so that the image generated by the last data enhancement processing process may be changed in a feature dimension, so that the purpose of changing the source image in multiple business features can be achieved through the iterative data enhancement processing process to meet the business requirements of the business scenario. The second computing device 102 may include multiple computing nodes, and each computing node may implement a data enhancement processing process. Each computing node corresponds to a variable feature and at least one fixed feature. The variable feature is a business feature that changes when data enhancement is performed on an image, and the fixed feature is a business feature that remains unchanged when data enhancement is performed on an image. Thus, each computing node can keep the image generated by the previous computing node unchanged based on the fixed feature, and change the image generated by the previous computing node based on the variable feature to perform a data enhancement process on the image generated by the previous computing node. For example Figure 1In the computing nodes 1 and 2, computing node 1 uses business feature 1 as a fixed feature and business feature 2 as a variable feature. Computing node 1 keeps business feature 1 of the source image unchanged, and changes the source image according to business feature 2 of the source image to realize the first data enhancement process of the source image, and outputs the enhanced image to computing node 2. The enhanced image output by computing node 1 is the image after level 1 enhancement of the source image. The fixed features of computing node 2 are business features 1 and 2, and the variable feature is business feature 3. Then computing node 2 keeps business features 1 and 2 of the image output by computing node 1 unchanged, and changes business feature 3 of the image output by computing node 1 to realize the second data enhancement process of the source image, and obtains the image after level 2 enhancement of the source image. Therefore, computing nodes 1 and computing nodes 2 can realize data enhancement of the two feature dimensions of business feature 2 and business feature 3 for the source image. By analogy, the second computing device 102 can perform multi-dimensional dynamic data enhancement on the source image data output by the first computing device 101 to meet the final business needs. Therefore, the data expansion layer can also be regarded as a multi-dimensional data cultivation layer. The second computing device 102 can perform multiple sets of iterative data enhancement processing on the source image, so that the final iterative output image of each set can meet the business requirements of the next sub-scenario of the business scenario. Figure 1 The data expansion layer includes multiple data enhancement paths, each of which may include at least one computing node. The data enhancement process performed by a data enhancement path may be regarded as a set of iterative data enhancement processes. For example, computing nodes 1, 3, and 4 may form a data enhancement path 1. Data enhancement path 1 performs three data enhancement processes on the source image. These three data enhancement processes are also a set of iterative data enhancement processes.

[0170] Each computing node may include a conditional data generator (CG), which can perform data enhancement on the input image based on the fixed features and variable features corresponding to the computing node to obtain an enhanced image. Considering that different objects may have different facial expressions or body postures in this business scenario, and the same object may also have different facial expressions or body postures in this business scenario, when performing data enhancement, in order to ensure that the images of different objects are different and the images of the same object are diverse, each computing node may also include an identity and diversity examiner (IDE). The IDE is used to maintain diversity between images of the same object and differences between images of different objects, for example Figure 2The IDE in any computing node in the , the IED can filter the image generated by the computing node to obtain multiple images, in which the images of different objects have obvious differences, and the images of the same object have certain diversity. In order to make the enhanced image obtained by the computing node meet the business needs of the business scenario, each computing node can also include a conditional data enhancement quality examiner (CQE), which can filter out images with high authenticity and change characteristics that meet the needs of the business scenario from the enhanced image, such as Figure 2 The CQE in any computing node in the business scenario can filter the image generated by the computing node based on the value of the change feature corresponding to the computing node in the business scenario to obtain multiple images, and the feature value of the change feature corresponding to each image in the computing node is the value of the change feature in the business scenario. The embodiment of the present application does not specifically limit the order in which the computing nodes use IDE and CQE, and any computing node can output an image that has been double-filtered by IDE and CQE to the next computing node. It should be noted that the data expansion layer may include multiple second computing devices 102, and the multiple second computing devices 102 can jointly complete the iterative data processing process performed on the source image, that is, each second computing device 102 is used to complete the data enhancement processing process performed by at least one computing node to improve the data processing efficiency of the data expansion layer. The data expansion layer may include only one second computing device 102, and the data enhancement processing process performed by each computing node is completed by one second computing device 102.

[0171] The data fusion layer may include a third computing device 103, which is used to receive multiple images ultimately output by each data enhancement path sent by the second computing device 102, and obtain multiple groups of images, each group of images including multiple images ultimately output by a data enhancement path. Considering that each group of images can meet the business needs of a sub-scene in the business scenario, and the business scenario can have any sub-scene, the third computing device 103 can assign weights to each group of images, extract some images from each group of images according to the weight of each group of images, and then form a training set with the images extracted from each group, so that the images in the training set can meet the needs of each sub-scene in the business scenario, that is, can meet the comprehensive needs of the business scenario. Among them, the weights assigned to each group of images by the mergefusion (MF) in the third computing device 103 can be weights determined according to the user's prior knowledge, and can also be weights trained by the model. MF can obtain a training set according to the weight of each group of images, and the third computing device 103 can use the training set to perform model training to obtain a trained image recognition model. If the recognition effect of the image recognition model in the business scenario reaches the preset recognition effect, the current training set can be used to train each model applied to the business scenario. Otherwise, the third computing device 103 can dynamically adjust the weight of each group of images; the third computing device 103 obtains a new training set according to the adjusted weight, and then performs model training according to the images in the new training set to obtain a new image recognition model, and then determines whether to adjust the weight of each group of images again according to the recognition result of the new image recognition model for the images in the business scenario, until the image recognition model trained by the training set obtained according to the adjusted weight reaches the preset recognition effect for the images in the business scenario. Among them, the user can be a technician. It should be noted that the user in this application is a real person, the object in this application is a virtual object that does not exist in reality and is displayed by the source image, and the target object in this application is a real object that exists in reality and is displayed by the real image.

[0172] It should be noted that the processing process implemented by the first computing device 101, the second computing device 102 and the third computing device 103 can also be implemented by one computing device. The first computing device 101, the second computing device 102 and the third computing device 103 can also be cloud computing devices. The cloud computing device can receive a data enhancement request sent by the terminal. The data enhancement request can carry multiple business features in the business scenario and the various values ​​of each business feature in the business scenario. The cloud computing device can implement the image processing process performed by the first computing device 101, the second computing device 102 and the third computing device 103 according to the multiple business features carried by the data enhancement request and the various values ​​of each business feature in the business scenario, and return an optimal training set to the terminal so that the terminal can train the various models required in the business scenario based on the optimal training set. The data enhancement request can also carry the comprehensive features of the business scenario, so that the cloud computing device can split the comprehensive features carried by the data enhancement request to obtain multiple business features, and determine the value of each business feature in the business scenario according to business needs. Of course, the cloud computing device can also train the various models required in the business scenario based on the optimal training set, and return the trained various models required in the business scenario to the terminal. The first computing device 101, the second computing device 102, and the third computing device 103 in this application can be a server, a terminal, or a personal computer (PC). The embodiments of this application do not specifically limit the above computing devices.

[0173] To further illustrate the hardware architecture of the first computing device, the second computing device, and the third computing device, see Figure 3The schematic diagram of the structure of a computing device provided by an embodiment of the present application is shown, and the computing device 300 includes a relatively large difference due to different configurations or performances, and may include one or more processors 301 and one or more memories 302, wherein the memory 302 stores at least one instruction, and the at least one instruction is loaded and executed by the processor 301 to implement the method provided by each method embodiment described below. Of course, the computing device 300 may also have components such as a wired or wireless network interface, a keyboard, and an input and output interface for input and output, and the computing device 300 may also include other components for implementing device functions, which will not be described in detail here. The processor can be any processor such as a central processing unit (CPU), a graphics processing unit (GPU), a tensor processing unit (TPU), a neural network processing unit (NPU), a brain processing unit (BPU), a deep learning processing unit (DPU), a holographic processing unit (HPU), a vector processing unit (VPU) and an intelligence processing unit (IPU).

[0174] The processor 301 can adopt a general-purpose CPU, a microprocessor, an application-specific integrated circuit (ASIC), a GPU or one or more integrated circuits to execute relevant programs to implement the following data enhancement method.

[0175] The processor 301 may also be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the data enhancement method of the present application may be completed by an integrated logic circuit of hardware or software instructions in the processor 301. The above-mentioned processor 301 may also be a general-purpose processor, a digital signal processor (digital signal processing, DSP), ASI, a field programmable gate array (field programmable gate array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present application may be implemented or executed. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in the embodiments of the present application may be directly embodied as being executed by a hardware decoding processor, or may be executed by a combination of hardware and software modules in a decoding processor. The software module may be located in a mature storage medium in the art such as a random access memory (RAM), a flash memory, a read-only memory (ROM), a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 302, and the processor 301 reads the information in the memory 302, and combines its hardware to complete the functions required to be performed by the modules included in the data enhancement device of the embodiment of the present application, or executes the data enhancement method of the method embodiment of the present application.

[0176] In an exemplary embodiment, a computer storage medium is also provided, which may be a computer-readable storage medium, such as a memory including instructions, and the instructions may be executed by a processor in a terminal to complete the data enhancement method in the following embodiments. For example, the computer-readable storage medium may be a ROM, a RAM, a compact disc read-only memory (CD-ROM), a magnetic tape, a floppy disk, an optical data storage device, and the like.

[0177] In summary, the computing device can first generate source images of different objects based on real images, and then perform an iterative image data processing process on each source image according to the business characteristics of the business scenario. Finally, the computing device can select some images from the images generated by the final iterative image data processing process to form a training set, so as to train various models in the business scenario based on the images in the training set. To illustrate this process, see Figure 4 A flow chart of a data enhancement method provided in an embodiment of the present application is shown, and the flow of the method includes steps 401-406.

[0178] 401. A computing device obtains multiple business features in a business scenario and a value of each business feature in the business scenario.

[0179] The business scenario can be any scenario for image recognition business, and the computing device can generally disassemble the business requirements of the business scenario to obtain multiple demand information under the business scenario, each of which can include a business feature and the value of the business feature under the business scenario. In a possible implementation, this step 401 can be implemented by the process shown in the following steps 4011-4013.

[0180] Step 4011: The computing device obtains comprehensive business characteristics of the business scenario, where the comprehensive business characteristics are used to indicate business requirements of the business scenario.

[0181] The computing device may first obtain the business requirements of the business scenario, and then obtain the business comprehensive features of the business scenario based on the business requirements. The computing device may obtain the business requirements sent by the terminal, or may obtain the business requirements based on the user's operation. The user may enter the business requirements of the business scenario in the display interface of the computing device, and when it is detected that the user clicks the confirmation button in the display interface, the computing device is triggered to obtain the business requirements under the business scenario. After the computing device obtains the business requirements, the business requirements may be regarded as the business comprehensive features of the business scenario. For example, taking facial images as an example, the business requirements of the business scenario may be black facial images of people aged 30 to 40, black facial images of people aged 40 to 50, and black facial images require 15-degree side faces and 30-degree side faces, and black facial images have happy or melancholy expressions.

[0182] Of course, the computing device can also directly obtain the business comprehensive features from the data enhancement request sent by the terminal. In one possible implementation, the terminal sends a data enhancement request to the computing device, the data enhancement request is used to indicate the generation of an image that meets the needs of the business scenario, the data enhancement request carries the business comprehensive features of the business scenario, and the computing device receives the data enhancement request, so that the computing device can obtain the business comprehensive features from the data enhancement request, and perform the data enhancement process shown in steps 401-406 based on the indication of the data enhancement request. For example, when the computing device is a cloud server, the cloud server obtains the process of business features based on the data enhancement request.

[0183] Step 4012: The computing device splits and processes the comprehensive business characteristics to obtain multiple business characteristics of the business scenario.

[0184] Still taking the example in step 4011 as an example, the computing device can regard black as a skin color feature, 30 to 50 years old as an age feature, 15-degree side face and 30-degree side face as side face angle features, happy and sad expressions as expression features, and the skin color feature, age feature, side face angle feature and expression feature are each a business feature, so that the computing device can obtain multiple business features under the business scenario.

[0185] Step 4013: The computing device determines the values ​​of each business feature in the business scenario based on the business requirements of the business scenario.

[0186] In this business scenario, each business feature has a value range, and each value within the value range of each business feature is also a value of each business feature in this business scenario. Still taking the example in step 4011 as an example, the skin color feature is black, and black is regarded as the value of the business feature in this business scenario. In a possible implementation, the value range of a feature may also include multiple value intervals, and each value interval may be regarded as a value in this business scenario. For example, 30-50 years old may be regarded as the value range of the age feature in this business scenario, and the value range may include the value interval of 30-40 years old and the value interval of 40-50 years old. When the computing device regards a value interval as a value of a business feature, since the value range of the age feature includes 2 value intervals, the number of values ​​of the age feature in this business scenario is 2.

[0187] In one possible manner, the user may also input various business features under the business scenario and various values ​​of each business feature in the feature acquisition interface of the computing device. When the computing device detects that the user clicks the confirmation button in the feature acquisition interface, the computing device is triggered to acquire various business features under the business scenario and various values ​​of each business feature.

[0188] The user can also enter the various business features under the business scenario and the various values ​​of each business feature in the feature acquisition interface of the terminal. When the terminal detects that the user clicks the confirmation button in the feature acquisition interface, the terminal is triggered to send the various business features under the business scenario and the various values ​​of each business feature to the computing device, so that the computing device can obtain the various business features under the business scenario and the various values ​​of each business feature. For example, when the computing device is a cloud server, the cloud server can obtain the values ​​of the business features sent by the terminal.

[0189] 402. The computing device sorts the multiple business features based on the number of values ​​each business feature takes in the business scenario and the number of original values ​​of each business feature to obtain a feature sequence, which is used to represent the data enhancement processing order of the multiple business features in the business scenario.

[0190] The original value of each business feature includes all the feature values ​​that each business feature can reach. Taking the skin color feature as an example, the value of the skin color feature in this business scenario is black. The original value of the skin color feature can include four values: black, white, yellow and brown.

[0191] When the computing device performs data enhancement on the image by changing the original value of any business feature on the image, an enhanced image can be obtained. Based on the business requirements of the business scenario, only part of the enhanced image is a valid image, and the rest are invalid images. Among them, the value of any business feature on the valid image is the value of the business feature in the business scenario, and the image other than the valid image in the enhanced image is an invalid image. The value of the business feature on the invalid image does not meet the requirements of the business feature in the business scenario, and the invalid image is a wasted image. For example, if the computing device performs data enhancement on the face image based on the original 4 values ​​of the skin color feature, the obtained enhanced image may include face images of four skin colors, namely black, white, yellow and brown. Since only black skin color face images are required in the business scenario, 1 / 4 of the black skin color face images in the enhanced image are valid images, and the other 3 / 4 of the face images are invalid face images, resulting in 3 / 4 of the face images in the enhanced image being wasted.

[0192] The computing device may sort the multiple features based on the proportion of valid images in the image generated when each business feature is used as a variable feature for data enhancement. In a possible implementation, the process shown in step 402 may be implemented by the process shown in steps 4021-4022 below.

[0193] Step 4021: The computing device calculates the ratio between the number of values ​​of each business feature in the business scenario and the number of original values ​​of each business feature to obtain the ratio of each business feature. The ratio of a business feature is the proportion of valid images in the image generated when the business feature is used as a changing feature for data enhancement.

[0194] The change feature is a business feature that changes when data enhancement is performed on the image. Taking the age feature in step 4013 as an example, if the value of the age feature is a value interval within the value range in the business scenario, and the value interval of the age feature in the business scenario includes 30-40 years old and 40-50 years old, then the number of value intervals of the age feature in the business scenario is 2, and the original value interval of the age feature can include 7 value intervals such as under 10 years old, 10-20 years old, 20-30 years old, ..., over 60 years old, and the original number of values ​​of the age feature is 7, then the ratio of the age feature is 2 / 7.

[0195] Step 4022: The computing device sorts the multiple business features in ascending order according to the size of the ratio of each business feature to obtain the feature sequence.

[0196] The smaller the ratio of the business feature, the smaller the order of the business feature in the feature sequence, and the larger the ratio of the business feature, the larger the order of the business feature in the feature sequence. Therefore, the computing device can sort the various business features in order from small to large ratios to obtain the feature sequence.

[0197] Based on the above sorting, the business features corresponding to the smaller proportion of valid images are sorted first, and the business features corresponding to the larger proportion of valid images are sorted last. Since the subsequent computing device can iteratively perform data enhancement processing on the image according to the order of each business feature, the higher the invalid image, the more waste. When the computing device uses a business feature as a changing feature to process the image, a valid image is obtained instead of an invalid image. The lower the effective feature corresponding to the business feature, the more invalid images do not need to be generated. Therefore, giving priority to processing business features with low ratios can avoid generating a large number of invalid images.

[0198] The order of each feature in the feature sequence may also be affected by the weight. In a possible implementation, the computing device may determine the order of each service feature according to the size of the ratio of each service feature and the first weight of the service feature, wherein the first weight of each service feature is the weight assigned to each service feature by the user based on prior knowledge. The computing device may calculate the product of the ratio of each service feature and the first weight of each service feature to obtain the product corresponding to each service feature. The smaller the value of the product corresponding to any service feature, the smaller the order of the service feature in the feature sequence; otherwise, the greater the order of the service feature in the feature sequence.

[0199] 403. The computing device inputs the target number into the target data enhancement model, and the target data enhancement model generates the target number of source images based on the input target number.

[0200] This step 403 can be performed by the SG in the computing device. Different source images have different individual features. Since the number of images in the existing training data set is small, and the number of individuals corresponding to the images in the existing training data set is relatively small, the computing device can enhance the number of individuals in the images in the existing training data set through this step 403 to improve the diversity of the data individuals in the existing training data set.

[0201] The target data enhancement model can be trained by multiple real images and used to generate source images. The embodiment of the present application does not specifically limit the number of targets, and each real image corresponds to a target object. The target data enhancement model can be a style-based generative adversarial network (StyleGAN) model. The number of targets can be any value according to actual needs, and the embodiment of the present application does not specifically limit the number of targets.

[0202] Among them, the StyleGAN model is a type of generative adversarial network (GAN), which is a deep learning model. The model includes at least two modules: one module is a generative model, and the other module is a discriminative model. These two modules learn from each other through game, so as to produce better output. Both the generative model and the discriminative model can be neural networks, specifically deep neural networks or convolutional neural networks. The basic principle of GAN is as follows: Taking the GAN that generates pictures as an example, suppose there are two networks, G (generator) and D (discriminator), where G is a network that generates pictures. It receives a random noise z and generates pictures through this noise, which is recorded as G(z); D is a discriminative network, which is used to determine whether a picture is "real". Its input parameter is x, x represents a picture, and the output D(x) represents the probability that x is a real picture. If it is 1, it means that it is 100% a real picture, and if it is 0, it means that it cannot be a real picture. In the process of training the generative adversarial network, the goal of the generative network G is to generate realistic images as much as possible to deceive the discriminative network D, while the goal of the discriminative network D is to distinguish the images generated by G from the real images as much as possible. In this way, G and D constitute a dynamic "game" process, which is also the "adversary" in the "generative adversarial network". The final result of the game is that, under ideal conditions, G can generate images G(z) that are "real" enough, while D has difficulty determining whether the images generated by G are real, that is, D(G(z)) = 0.5. In this way, an excellent generative model G is obtained, which can be used to generate images.

[0203] 404. The computing device filters the target number of source images to obtain source images that meet a third preset condition.

[0204] The third preset condition includes at least one of the following: the similarity between a source image and source images other than the source image in the target number of source images is less than or equal to the third similarity; the similarity between a source image and multiple real images is less than or equal to the fourth similarity, and the multiple real images are used to train the target data enhancement model; the authenticity index of a source image reaches the preset authenticity. In some cases, each source image in the target number of source images can correspond to an object respectively, but if some source images are similar, it can be considered that similar source images are images of the same object, and for each object, the computing device only needs to retain one source image of the object, and in order to ensure the quality of the source image, the source image must also meet the authenticity requirements. Therefore, the computing device can obtain a source image that meets the requirements through the process shown in this step 404.

[0205] The similarity between two images can be represented by the cosine similarity between the feature vectors of the two images. The DE in the computing device can extract the feature vector of each source image based on the feature extraction algorithm. For any source image in the target number of source images, the DE can calculate the cosine similarity between the feature vector of the source image and the feature vectors of other source images in the target number of source images by the cosine similarity algorithm, and obtain multiple similarities between the source image and other source images, each similarity is the similarity between the source image and another other source image, and the other source images in the target number of source images are source images other than the source image in the target number of source images. When the multiple similarities are all less than or equal to the third similarity, and the third preset condition is that the similarities between a source image and source images other than the source image in the target number of source images are all less than or equal to the third similarity, it means that the source image is not similar to the other source images, then the source image meets the third preset condition, and the DE can retain the source image.

[0206] When the third preset condition is that the similarities between a source image and the source images other than the source image in the target number of source images are all less than or equal to the third similarity, and there is a similarity greater than the third similarity among the multiple similarities, it means that there is a source image similar to the source image among the target number of source images, then the source image does not meet the third preset condition, and the DE can delete the source image. After the judgment of the source image is completed, the DE can calculate the similarity between the next source image and the current remaining source image, and so on, the DE can obtain multiple source images that meet the third preset condition. The feature extraction algorithm can be an additive angular margin loss for deep face recognition (ArcFace) algorithm. The third similarity can be 0.2, of course, the third similarity can also be other values, and the third similarity can have different values ​​in different business scenarios. The embodiment of the present application does not specifically limit the third similarity.

[0207] For any source image, DE can calculate the similarity between the source image and multiple real images to obtain multiple similarities, each of which corresponds to a real image. When the multiple similarities are all less than the fourth similarity, and the third preset condition is that the similarity between a target source image and multiple real images is less than or equal to the fourth similarity, it means that the source image is not similar to each real image, and the source image meets the third preset condition, and DE can retain the source image. When there is a similarity greater than the fourth similarity among the multiple similarities, and the third preset condition is that the similarity between a source image and multiple real images is less than or equal to the fourth similarity, it means that there is a real image similar to the source image in the multiple real images. If, then the source image does not meet the third preset condition, and DE can delete the source image. After DE completes the judgment on the source image, the computing device can calculate the similarity between the next source image and multiple real images. By analogy, the DE can obtain multiple source images that meet the third preset condition. The fourth similarity may be the same as the third similarity or may be different. The fourth similarity may have different values ​​in different business scenarios. The embodiment of the present application does not specifically limit the fourth similarity.

[0208] The authenticity is used to indicate the degree of realism of the generated image. The higher the authenticity of the source image, the better the quality of the source image. Otherwise, the lower the quality of the source image. In order to ensure that the source image used is a high-quality image, the SQE in the computing device can be based on the authenticity of at least one source image among the target number of source images. When the authenticity of the at least one source image reaches the preset authenticity, the at least one source image is a high-quality image. If the third preset condition is that the authenticity index of a source image reaches the preset authenticity, it can be considered that the at least one source image meets the third preset condition. Otherwise, the at least one source image does not meet the third preset condition, and the SQE can delete the source image that does not meet the third preset condition. The authenticity of the at least one source image can be represented by the Fréchet inception distance (FID) indicator. When the index value of the FID index of the at least one source image reaches the preset index value, the authenticity of the at least one source image reaches the preset authenticity, wherein the preset index value is used to indicate the preset authenticity. The embodiment of the present application does not specifically limit the preset authenticity or preset index value.

[0209] When the third preset condition includes that the similarities between a source image and source images other than the source image in the target number of source images are less than or equal to the third similarity; the similarities between a source image and multiple real images are less than or equal to the fourth similarity; and the authenticity index of a source image reaches the preset authenticity, the computing device can use the same source image in the SQE and the source images retained in the DE as the source image that meets the third preset condition.

[0210] The computing device can also filter the source image with individual features based on the value of the first feature in the feature sequence in the business scenario to obtain at least one first source image, and the feature value of the first feature of each first source image is the value of the first feature in the business scenario; the computing device calculates the authenticity of the at least one first source image, and obtains a first source image that reaches a preset authenticity based on the authenticity of the at least one first source image. For example, the computing device may input the target number into SG, and SG uses the StyleGAN model to generate the target number of source images. The computing device may filter out multiple source images that meet the third preset condition from the source images of the target data volume based on the third similarity and the fourth similarity, and store the multiple source images. The multiple source images may be stored in the first database, and of course may be stored in the software storage space, the hardware storage space, or the cloud storage space. If the image is a face image, the first feature in the feature sequence is a skin color feature, and the skin color feature is black, then the computing device may screen out qualified black face images (first source images) from the multiple generated source images, and further perform quality screening on the black face images according to the authenticity of the images to obtain black face images with higher authenticity. The computing device may re-store the black face images with higher authenticity, or delete the previously stored black face images with lower authenticity.

[0211] It should be noted that the computing device may also execute steps 403-404 first, and then execute steps 401-402. The embodiment of the present application does not specifically limit the execution order of steps 403-404 and steps 401-402.

[0212] 405. The computing device sets an identity tag for each source image that meets the third preset condition. The identity tag of a source image includes an identity identifier of the source image.

[0213] Each source image corresponds to an individual feature, and each individual feature can represent an object. Therefore, when the computing device obtains a source image that meets the third preset condition, an identity (ID) can be set for each source image obtained. The ID of each source image is used to uniquely indicate the object corresponding to the individual feature of each source image, that is, each ID is used to uniquely indicate an object, so that the computing device can obtain source images of different objects, so as to perform multi-dimensional data enhancement on the source image of each object later. The ID can be a name or a number, and the embodiment of the present application does not specifically limit the representation form of the ID. The computing device can add the ID of each source image to the identity tag of each source image.

[0214] The computing device may also associate the identity tag of each source image with each source image, and the association process may be associated storage. The associated storage may be: the computing device stores the identity tag of each target source image and each source image in a storage space, or in the same storage path, or adds the data tag of each source image to the file name of each target source image, or uses an association table to associate each target source image with the identity tag of each target source image, thereby avoiding manual labeling.

[0215] 406. The computing device performs iterative data enhancement processing on the source image that meets the third preset condition according to the order of each business feature in the feature sequence, and one data enhancement processing process corresponds to one business feature in the feature sequence.

[0216] The process shown in step 406 is a process in which the computing device iteratively performs data enhancement processing on the image according to the order of each business feature in the feature sequence, wherein the computing device uses the source image that meets the third preset condition as the image to be processed in the first data enhancement processing process. If the computing device does not execute step 404, the computing device uses the target number of source images generated in step 403 as the image to be processed in the first data enhancement processing process. Each data enhancement processing process can be considered as a data enhancement processing process, wherein a data enhancement processing process is to perform a data enhancement operation on the image to be processed, and use the image after the data enhancement operation as the next data enhancement image to be processed. To further illustrate a data enhancement processing process, refer to the process shown in steps 4061-4065 below.

[0217] Step 4061: During any data enhancement process, the computing device determines the business feature corresponding to the data enhancement process as a change feature during the data enhancement process, and the change feature is a business feature that changes when data enhancement is performed on the image.

[0218] Step 4062: During any data enhancement process, the computing device determines the business feature preceding the business feature corresponding to the data enhancement process in the feature sequence as a fixed feature in the data enhancement process. The fixed feature is a business feature that remains unchanged when data enhancement is performed on the image.

[0219] During each data enhancement process, the computing device needs to keep the individual features of the image unchanged when performing data enhancement on the image. Therefore, the fixed features corresponding to the data enhancement process may include not only the business features before the business features corresponding to the data enhancement process in the feature sequence, but also the individual features of the image to be processed during the data enhancement process.

[0220] Step 4063: During any data enhancement process, the computing device performs data enhancement on the image generated by the previous data enhancement process based on the change characteristics during the data enhancement process, where the change characteristics are business characteristics that change when the image is data enhanced.

[0221] Since the fixed features in each data processing process are the business features before the business features corresponding to each data enhancement process in the feature sequence, and the business features corresponding to each data processing process are the changing features in each data processing process, the fixed features in this data enhancement process are the fixed features and the transformation features in the previous data enhancement process. Therefore, in this data enhancement process, the computing device can keep the fixed features and the changing features of the image to be processed in the previous data enhancement process unchanged, and only change the changing features of the image to be processed. That is, the feature value of the changing feature in the image generated by the previous data enhancement process is changed, and the fixed features in the image generated by the previous data enhancement process are kept unchanged. The fixed features are the business features before the business features corresponding to the data enhancement process in the feature sequence. In a possible implementation, the computing device can change the feature value of the changing feature on the image to be processed to a value of the changing feature in the business scenario, so that the data enhancement of the image to be processed can be realized in this data enhancement process, and the image to be processed in this data enhancement process is also the image generated by the previous data enhancement process. By analogy, the image generated by the computing device after the last data processing process is compared with the source image, which realizes data enhancement of multiple business features in the business scenario, so that the image generated by the last data processing process can meet the business needs of the business scenario.

[0222] The computing device can complete the data enhancement process through the fourth data enhancement model. In one possible implementation, the computing device inputs the image generated by the last data enhancement process into the fourth data enhancement model, and the fourth data enhancement model keeps the fixed features of the input image unchanged during the data enhancement process, and based on a value of the changing features in the data enhancement process in the business scenario, the input image is data enhanced, and the image generated by the data enhancement process is output. The fourth data enhancement model can be a star generative adversarial network (StarGAN) model. The embodiment of the present application does not specifically limit the fourth data enhancement model, as long as the fixed features of the input image can be kept unchanged, and the changing features change within the allowable range in the business scenario. The computing device can also associate the image generated by this data enhancement process with the identity tag of the corresponding source image, so that each generated image has an identity tag. For example, in the first data enhancement process, the computing device performs data enhancement processing on source images 1 and 2 to obtain 6 images, namely images 1-6, where images 1-3 are images generated by performing data enhancement on source image 1, and images 4-6 are images generated by performing data enhancement on source image 2. The computing device can use the identity tag of source image 1 as the identity tag of images 1-3, and the identity tag of source image 2 as the identity tag of images 4-6.

[0223] Step 4064: During any data enhancement process, the computing device uses the fixed features in the data enhancement process as the data label corresponding to the data enhancement process, and associates the data label corresponding to the data enhancement process with the image generated by the data enhancement process.

[0224] The computing device associates the data label corresponding to the data enhancement process with the image generated by the data enhancement process, thereby avoiding manual labeling of the enhanced image. The process of the computing device associating the data label with the image generated by the data enhancement process is similar to the process of the computing device associating the identity label of the source image with the source image in step 405. Here, the embodiment of the present application does not elaborate on the process of associating the data label with the image generated by the data enhancement process.

[0225] The fixed features in each data enhancement process can be the business features of any sub-scenario under the business scenario. The image generated by the data enhancement process can meet the business needs of the sub-scenario. Therefore, the computing device can also associate and store the data label corresponding to the data enhancement process with the image generated by the data enhancement process, so that when data processing is performed on the sub-scenario, an image that meets the business needs of the sub-scenario can be quickly obtained. The computing device can associate the identity label of the image to be processed with the feature label for subsequent search. The process shown in step 4064 is also the process of associating the image generated by the data enhancement process with the data label of the data enhancement process.

[0226] Step 4065: In any data enhancement process, the computing device determines the image generated by the data enhancement process as the image to be processed in the next data enhancement process of the data enhancement process.

[0227] In order to ensure the quality of the image to be processed in the next data enhancement process, the computing device can also filter the image generated by the data enhancement process to obtain an image that meets the first preset condition, and determine the image that meets the first preset condition as the image to be processed in the next data enhancement process of the data enhancement process, wherein the first preset condition can be at least one of the following: the authenticity of an image reaches a preset authenticity; the similarity between images with the same individual features is greater than or equal to the first similarity; the similarity between images with different individual features is less than or equal to the second similarity. By filtering the image generated by the data enhancement process through the first preset condition, a small number of images with low quality can be excluded.

[0228] If multiple images are generated in the data processing process, for any image among the multiple images, when the authenticity of the image reaches the preset authenticity, and the first preset condition is that the authenticity of an image reaches the preset authenticity, then the image meets the first preset condition, and the IDE in the computing device retains the image; otherwise, the image does not meet the first preset condition, and the IDE deletes the image, so that the IDE can obtain at least one image that reaches the preset authenticity. Each of the multiple images has an identity tag, and for at least one image with the same identity tag, it is considered that the individual features of the at least one image with the same identity tag are the same, and the IDE can calculate the similarity between the at least one image with the same identity tag. For any image among the at least one image with the same identity tag, when the similarity between the image and other images other than the image in the at least one image with the same identity tag is less than the first similarity, and the first preset condition is that the similarity between images with the same individual features is greater than or equal to the first similarity, then the image meets the first preset condition, and the IDE retains the image; otherwise, the image does not meet the first preset condition, and the IDE deletes the image, so that the IDE can obtain at least one image with a similarity greater than or equal to the first similarity and the same individual features. The first similarity may be 0.8 or other values. The first similarity may have different values ​​in different business scenarios. The embodiment of the present application does not specifically limit the first similarity.

[0229] For any image among the multiple images, the CQE in the computing device calculates the similarity between the image and an image with a different identity tag. When the similarity is less than or equal to the second similarity, and the first preset condition is that the similarity between images with different individual features is less than or equal to the second similarity, then the image meets the first preset condition, and the COE can retain the image. Otherwise, if the image does not meet the first preset condition, the IDE deletes the image, so that the IDE can obtain at least one image with a similarity less than or equal to the second similarity and different individual features. The first similarity can be 0.2 or other values. The second similarity can have different values ​​in different business scenarios. The embodiment of the present application does not specifically limit the second similarity.

[0230] When the first preset condition includes that the authenticity of an image reaches a preset authenticity; the similarity between images with the same individual features is greater than or equal to the first similarity; the similarity between images with different individual features is less than or equal to the second similarity, the computing device can take the same image from the images retained by the IED and the CQE, and use the same image as an image that meets the first preset condition. The calculation of the similarity between images is described above, and the process of calculating the similarity between images is not repeated in this embodiment of the present application.

[0231] In order to ensure that the image to be processed in the next data enhancement process can meet the business requirements of the business scenario, the computing device can also filter the image generated by the data enhancement process to obtain an image that meets the second preset condition, and use the image that meets the second preset condition as the image to be processed in the next data enhancement process of the data enhancement process. The second preset condition can be: the feature value of the change feature of an image in the data processing process is the value of the change feature in the business scenario.

[0232] Of course, the computing device can also filter the image that meets the first preset condition to obtain an image that meets the second preset condition (at this time the image also meets the first preset condition), and use the image that meets the second preset condition as the image to be processed in the next data enhancement processing process, so that the image to be processed in the next data enhancement processing process has high quality and meets the needs of the business scenario.

[0233] To further illustrate any of the data enhancement processing processes shown in 407-412, refer to the following data enhancement processing process: this process takes a face image as an example, and the image generated last time is a black face image. The computing device can obtain the black face image generated by the last data enhancement process from the stored image, and the computing device can use the fixed features in the last data enhancement process and the changing features in the last data enhancement process as the fixed features in this data enhancement process; the computing device can use the age feature corresponding to this data processing process as the changing feature, and a value of the age feature in this business scenario is 30-40 years old; the computing device sends the black face image, the value of the age feature (30-40 years old) and the fixed feature in this data enhancement process to the CG in the computing device, and the CG determines the StarGAN model to be used in this data enhancement process through the changing features and fixed features in this data enhancement process, and its In the data enhancement process, the StarGAN model to be used can keep the fixed features of the image unchanged during the data enhancement process, and the feature values ​​of the changing features of the image during the data enhancement process change; CG inputs the black face image and the value of the age feature (30-40 years old) into the StarGAN model, and the StarGAN model changes the feature value of the age feature of the input black face image between 30-40 years old, and outputs a face image of 30-40 years old (that is, the enhanced image) to complete the data enhancement process of the input black face image, and the enhanced image is the image generated by the data enhancement process; the computing device inputs the enhanced image into the filter, and the filter filters out an image with high quality that meets the requirements of the business scenario from the enhanced image, and the computing device associates and stores the filtered image and the data label corresponding to the data enhancement process for subsequent use.

[0234] The method provided in the embodiment of the present application performs iterative data enhancement processing on the image according to the order of each business feature in the feature sequence, wherein, in any data enhancement processing process, data enhancement is performed on the image generated by the previous data enhancement processing based on the changing features in the data enhancement processing process and the fixed features in the data enhancement processing process. Since each data enhancement processing process performs data enhancement on a changing feature in the image, when the iterative data enhancement processing is completed, the enhanced image finally obtained is the result of data enhancement of the image based on multiple business features in the business scenario, so the enhanced image finally obtained can meet the business needs of the business scenario.

[0235] The computing device can also perform iterative data enhancement processing on the image through a tree structure of a data augmentation tree (DA-Tree). The DA-Tree can include a root node and multiple common nodes. At least one of the multiple common nodes and the root node can form a data enhancement path. The same common nodes can exist between different data enhancement paths. Each data enhancement path is headed by a root node. The common nodes in each data enhancement path are sequentially connected to the root node, and the last common node in each data enhancement path can be regarded as a leaf node. For any node in a data enhancement path, the previous node of the node on the data enhancement path is the parent node of the node, and the next node of the node is the child node of the node. Therefore, there is a dependency relationship between two adjacent nodes on a data enhancement path. The computing device can perform a data enhancement process on each node of each data enhancement path, wherein a set of iterative data enhancement processing processes is completed on a data enhancement path, so that the computing device can perform multiple sets of iterative data enhancement processes on the image based on the tree structure of the DA-Tree. To illustrate this process, see Figure 5 The flowchart of a data enhancement method provided by an embodiment of the present application is shown, and the method flow may include 501-511.

[0236] 501. The computing device establishes an initial data enhancement tree according to the order of each business feature in the feature sequence, where the initial data enhancement tree includes multiple initial nodes, each business feature corresponds to at least one initial node, an initial node corresponding to any business feature is a parent node of at least one target initial node among the multiple initial nodes, and the at least one target initial node corresponds to the next business feature of the business feature in the feature sequence.

[0237] Since each business feature may have multiple values ​​in the business scenario, during a data enhancement process, the computing device may change the feature value of a business feature of the image to be processed into a value of the business feature in the business scenario. Considering that a data enhancement process is performed at each node, each business feature may correspond to at least one initial node, and each initial node corresponds to a value of a business feature. The computing device may establish an initial data enhancement tree based on the order of each business feature in the feature sequence and the values ​​of each business feature in the business scenario. In one possible implementation, the implementation process of this step 501 may be implemented by the process shown in the following steps 5011-5012.

[0238] Step 5011: For the first business feature in the feature sequence, the computing device generates a root node of the initial data enhancement tree.

[0239] The computing device can perform data enhancement on the source image at the root node according to various values ​​of the first business feature in the business scenario.

[0240] Step 5012: For any business feature other than the first business feature in the feature sequence, the computing device generates a first number of node sets of the initial data enhancement tree, the first number being the number of values ​​of the previous business feature of any business feature in the feature sequence in the business scenario, each node set including a second number of initial nodes, each node in the node set being a child node of an initial node corresponding to the previous business feature, and the second data volume being the number of values ​​of the business feature in the business scenario.

[0241] Since the previous business feature of any business feature can have at least one value in the business scenario, the computing device can subsequently change the image to be processed based on each value of the previous business feature in the business scenario at the node corresponding to the previous business feature, thereby obtaining multiple changed images, which are enhanced images generated after data enhancement of the image to be processed. Therefore, each data enhancement is for each business feature. For each image after the previous business feature changes, the service can perform data enhancement on the image after each change of the previous business feature based on the value of the business feature in the business scenario. Therefore, the computing device generates a first number of node sets, each of which corresponds to a value of the previous business feature of the business feature in the business scenario. For the business feature, the computing device can subsequently change the image after each change of the previous business feature based on each value of the business feature in the business scenario. Therefore, each node set includes an initial node of the second data volume, and an initial node in each node set corresponds to a value of the business feature in the business scenario.

[0242] The computing device can determine the second data volume according to each value of the business feature in the business scenario, generate the initial node of the second data volume, and form the initial nodes of the second data volume into a node set, each initial node in the node set corresponds to a value in the business scenario. The computing device can determine the first data volume according to each value of the previous business feature of the business feature in the business scenario, and generate a node set of the first data volume. The computing device uses the initial node in each node set as a child node of any node corresponding to the previous business feature of the business feature in the feature sequence, that is, any node corresponding to the previous business feature of the business feature in the feature sequence is the parent node of each node in the node set. If any business feature in step 501 is any business feature other than the first business feature in the feature sequence in step 5012, then an initial node corresponding to any business feature in step 501 is also the parent node of each initial node in a node set in step 5012, that is, a node set is composed of at least one target initial node.

[0243] For example Figure 6 The schematic diagram of an initial data enhancement tree provided by an embodiment of the present application is shown. Still taking a face image as an example, the feature sequence includes skin color feature, age feature, side face angle feature and expression feature, wherein the skin color feature is the first business feature in the feature sequence, and the value of the skin color feature in this business scenario is black, that is, there is one skin color feature, and the age feature has two values ​​in this business scenario, which are 30-40 years old and 40-50 years old, respectively. The side face angle feature has two values ​​in this business scenario, which are 30 degrees and 15 degrees, respectively. The expression feature has two values ​​in this business scenario, which are happy and sad, respectively. If the business feature is Figure 6 In the side face angle feature, since the previous feature of the side face angle feature in the feature sequence is the age feature, and there are 2 initial nodes corresponding to the age feature, the computing device can generate 2 node sets corresponding to the side face angle feature. Since the side face angle feature has 2 values ​​of 30 degrees and 15 degrees in this business scenario, each node set contains 2 initial nodes, wherein one initial node corresponds to the value of 30 degrees of the side face angle feature in the business scenario, and the other initial node corresponds to the value of 15 degrees of the side face angle feature in the business scenario.

[0244] It should be noted that the initial node corresponding to the last business feature in the feature sequence has only a parent node but no child nodes. Therefore, the initial node corresponding to the last business feature can be regarded as a leaf node.

[0245] 502. The computing device determines the fixed feature of each initial node and the variable feature of each initial node based on the order of each business feature in the feature sequence.

[0246] Each initial node corresponds to a data enhancement process, so each initial node has a changing feature and at least one fixed feature, wherein the changing feature of each initial node is the business feature corresponding to each initial node in the feature sequence. The fixed feature of each initial node is the feature before the changing feature of each initial node in the feature sequence. It should be noted that the individual feature of the image is the most basic business feature of the image, and the individual feature may not be located in the feature sequence. The individual feature is a business feature that remains unchanged during each data enhancement process. Therefore, the fixed feature of each initial node may include not only the features before the business feature corresponding to each initial node in the feature sequence, but also the individual feature. For example, Figure 6 Node 1 in is the root node. The first feature in the feature sequence corresponding to the root node is the skin color feature. Since there are no other business features before the first feature in the root node feature sequence, and the source image of each object needs to be expanded, the computing device can use the individual features of the image as the fixed features of the root node and the skin color feature as the variable feature of the root node.

[0247] The features before the changing features of each initial node in the feature sequence are the fixed features and changing features of the parent node of each initial node. In a possible implementation, after the computing device generates each initial node, the fixed features and changing features of the parent node of the initial node can be directly used as the fixed features of the initial node. For example, Figure 6 The parent node of node 2 in is node 1, and the fixed features of node 1 are individual features and variable features (skin color features). The fixed features of node 2 are individual features and skin color features. Since node 2 corresponds to the age feature in the feature sequence, the variable feature of node 2 is the age feature.

[0248] When generating each initial node, the computing device may assign a node identifier to each initial node. The node identifier of an initial node is used to uniquely indicate the initial node. The node identifier may be a node number, for example Figure 6 In the initial data tree, the computing device can set a number for each initial node respectively, according to the order in which the initial nodes are generated. Figure 6 The initial nodes in are numbered 1-15 respectively. Of course, the node identifier may also be in other representation forms, and the embodiment of the present application does not specifically limit the representation form of the node identifier.

[0249] The computing device may also store the node identification of each initial node in association with the fixed features and the variable features of each initial node for subsequent searching.

[0250] 503. The computing device establishes a data tag for each initial node according to the fixed features of each initial node, and the data tag of an initial node includes the fixed features of the initial node.

[0251] When the data tag of any initial node is established, the computing device can associate the data tag of the initial node with the node. In a possible implementation, the computing device can associate and store the data tag of each initial node with the node identifier of each node, so that the data tag of each initial node can have an associated relationship with each initial node.

[0252] 504. The computing device obtains deletion information corresponding to the business scenario, where the deletion information is used to indicate an initial node to be deleted on the initial data enhancement tree.

[0253] The deletion information may include the node identifier of the initial node to be deleted and the deletion identifier corresponding to the node identifier of the initial node to be deleted, and the deletion identifier is used to indicate the deletion of the node. The deletion information may also include the node identifier of the retained initial node and the retention identifier corresponding to the node identifier of the retained initial node, and the retention identifier is used to indicate the retained node, that is, the node that is not deleted. The deletion information may also include the scenario identifier of the business scenario, and the scenario identifier of the business scenario is used to uniquely indicate the business scenario. The embodiment of the present application does not specifically limit the representation method of the scenario identifier.

[0254] The fixed features of each node are the business features required in the next sub-scenario of the business scenario. Therefore, each node corresponds to a sub-scenario, but some sub-scenarios do not exist in the business scenario. The initial nodes corresponding to these non-existent sub-scenarios are invalid nodes. In order to avoid subsequent computing devices from performing data enhancement on invalid nodes, users can determine whether the fixed features of any initial node are the business features required by the next sub-scenario of the business scenario based on the fixed features of the initial node and business experience. If so, the initial node is a valid node, otherwise the initial node is an invalid node, so that users can determine multiple valid nodes and multiple invalid nodes, where the fixed features of the valid nodes are the business features required by the sub-scenario that may exist in the business scenario.

[0255] The user can enter the node identifier of each invalid node and the scene identifier of the business scenario in the node deletion interface of the computing device, and when the computing device detects that the user clicks the confirmation button in the node deletion interface, the computing device is triggered to obtain the deletion information corresponding to the business scenario. The user can also enter the node identifier of each invalid node and the scene identifier of the business scenario in the node deletion interface in the terminal, and when the terminal detects that the user clicks the confirmation button in the node deletion interface, the terminal is triggered to send the deletion information corresponding to the business scenario to the computing device, so that the computing device can receive the deletion information corresponding to the business scenario.

[0256] The deletion information can be represented by a binary variable, wherein the binary variable may include n elements, each element corresponding to an initial node. The user can set a value for each element in the binary variable based on business experience. When the initial node corresponding to any element is an invalid node, the value of the element is a first value, otherwise the value of the element is a second value, wherein the first value is used to indicate that the initial node corresponding to the element is an invalid node, and the first value is equivalent to a deletion mark, and the first value can be 0. The second value is used to indicate that the initial node corresponding to the element is a valid node, and the second value is equivalent to a retention mark, and the second value can be 1. The embodiment of the present application does not specifically limit the first value and the second value.

[0257] 505. The computing device deletes the initial node on the initial data enhancement tree according to the deletion information corresponding to the business scenario to obtain a data enhancement tree, wherein the data enhancement tree includes multiple nodes, each business feature corresponds to at least one node, wherein a node corresponding to any business feature is a parent node of at least one target node among the multiple nodes, and the at least one target node corresponds to a business feature after the any business feature in the feature sequence.

[0258] The computing device can obtain the node identifier of the initial node to be deleted from the deletion information, and delete the initial node to be deleted on the initial data enhancement tree according to the node identifier of the initial node to be deleted, to obtain the data enhancement tree. When any initial node is deleted, if the initial node has a child node, the child node is automatically connected to the parent node of the initial node, so that the parent node of the initial node becomes the parent node of the child node. For example, if the initial node to be deleted carried by the deletion information is Figure 6 Nodes 6-9 and 12-13 in the calculation device are deleted. Figure 6 Nodes 6-9 and 12-13 in the equation, we get Figure 7 The data enhancement tree in , where Figure 7 A schematic diagram of a data enhancement tree provided in an embodiment of the present application.

[0259] For the first node, the second node, and the third node with a dependency relationship on the initial data tree, where the first node is the parent node of the second node, and the second node is the parent node of the third node, if the second node is deleted, the first node is the parent node of the third node, and the business feature corresponding to the third node is not the next feature of the business feature corresponding to the first node in the feature sequence, but the next feature of the business feature corresponding to the first node in the feature sequence. Therefore, the business features corresponding to the two nodes with a dependency relationship in the data enhancement tree are not necessarily the two adjacent features in the feature sequence. Therefore, the at least one target node corresponds to a business feature after any business feature in the feature sequence, and is not necessarily the next business feature of any business feature in the feature sequence. For example Figure 6 Node 7 in is deleted, causing the parent node of node 15 to change from the original node 7 to node 3 (see Figure 7 ), and the business feature corresponding to node 3 is the age feature, and the business feature corresponding to node 15 is the expression feature, and the expression feature is the next business feature under the age feature in the feature sequence.

[0260] It should be noted that the process shown in steps 501-505 is the process in which the computing device deletes the initial nodes on the initial data enhancement tree based on the order of each business feature in the feature sequence and the deletion information corresponding to the business scenario to obtain the data enhancement tree, that is, the computing device establishes the data enhancement tree based on the order of each business feature in the feature sequence. Of course, the computing device can also directly use the initial data enhancement tree as the data enhancement tree without deleting the initial nodes on the initial data enhancement tree.

[0261] 506. For the initial node whose parent node is deleted, the computing device updates the fixed features of the initial node to the fixed features and variable features of the current parent node of the initial node, and the current parent node of the initial node is the previous node of the initial node on the data enhancement path to which the initial node belongs on the data enhancement tree.

[0262] For the initial node whose parent node is deleted, the parent node of the initial node has changed. Considering that the fixed feature of each initial node is the fixed feature and the changed feature of the parent node of each initial node, the computing device can update the fixed feature of the initial node to the fixed feature and the changed feature of the current parent node of the initial node.

[0263] It should be noted that when executing step 502, the computing device may only determine the changing features of each initial node, but not the fixed features of each initial node. Instead, after the data enhancement tree is established, the fixed features of each node are determined according to the dependencies between the nodes on the data enhancement tree and the changing features of each node. In one possible implementation, when there are no individual features in the feature sequence, the computing device may use the individual features as the fixed features of each root node in the data enhancement tree. For any node other than the root node, the computing device uses the changing features and fixed features of the parent node of the node as the fixed features of the node. In one possible implementation, for any node on any data enhancement path on the data enhancement tree, the computing device may determine the changing features and fixed features of the previous node of the node on the data enhancement path as the fixed features of the node. When there are no individual features in the feature sequence, the computing device may also add the individual features to the fixed features of each node.

[0264] 507. The computing device updates the fixed feature on the data tag of the initial node to the updated fixed feature of the initial node.

[0265] The computing device may also not set a data label for the initial node on the initial data enhancement tree in advance, that is, not executing 503, but after the computing device determines the fixed feature of any node on the data enhancement tree, directly uses the fixed feature of the node as the data label of the node. Figure 6 Node 7 in is deleted, and the parent node of node 15 is changed from node 7 to node 3. The fixed features of node 3 are individual features and skin color features, and the variable features are age features. Therefore, the current fixed features of node 3 include individual features, skin color features, and age features.

[0266] 508. For any node among the multiple nodes, the computing device inputs the image to be processed into the first data enhancement model, and the first data enhancement model performs data enhancement on the input image based on the change characteristics of the node, and outputs the enhanced image. The image to be processed is the image generated when data enhancement is performed on the parent node of the node.

[0267] Any node among the multiple nodes may be any node on any data enhancement path of the data enhancement tree, and the computing device may perform data enhancement on the image to be processed based on the fixed features and the changing features of the node. The computing device may configure a data enhancement model for each node on the data enhancement tree. For the image to be processed of each node, the data enhancement model of each node may keep the image to be processed of each node unchanged on the fixed features of each node, and change the image to be processed based on the value of the changing features corresponding to the node, so as to achieve data enhancement on the input image. The first data enhancement model is also the data enhancement model of any node. When the changing feature of the node is the side face angle feature, the first data enhancement model may be a face pose network (FPN) model, and when the changing feature of the node is the expression feature, the first data enhancement model may be an anatomically-aware facial animation from a single image (GANimation) algorithm model based on anatomically-aware facial animation from a single image. The first data enhancement model is determined by the changing features and fixed features of the node, and the embodiment of the present application does not specifically limit the first data enhancement model.

[0268] If the node is a root node on the data enhancement tree, the image input to the first data enhancement model is a source image, wherein the source image may be the source image that meets the third preset condition obtained in step 404. If the computing device does not filter the source image generated in step 403, the image to be processed at the root node is the image generated in step 403.

[0269] In order to ensure that the image output by the first data enhancement model is of high quality and meets the business scenario, the computing device can filter the image output by the first data enhancement model according to the first preset condition and the second preset condition to obtain an image that meets the first preset condition and the second preset condition, and use the image that meets the first preset condition and the second preset condition as the image to be processed as the child node of the node. Since each node corresponds to a data enhancement process, and the process of filtering the image based on the first preset condition and the second preset condition is described in the previous text, the process of filtering the image based on the first preset condition and the second preset condition is not described in detail in this embodiment of the present application.

[0270] 509. For any node among the multiple nodes, an image output by the data enhancement model corresponding to the node is associated with a data label of the node.

[0271] The data enhancement model corresponding to the node is also the first data enhancement model. The association processing can be an associated storage. In one possible implementation method, the computing device stores the data label of the node and the image output by the data enhancement model corresponding to the node in a storage space, or in the same storage path, or adds the data label of the node to the image output by the data enhancement model corresponding to the node, or uses an association table to associate the image output by the data enhancement model corresponding to the node with the data label of the node, thereby avoiding manual labeling.

[0272] 510. For the child nodes of the node, the computing device inputs the image output by the first data enhancement model into the second data enhancement model, and the second data enhancement model performs data enhancement on the input image based on the change characteristics of the child nodes.

[0273] The second data enhancement model is determined by the changing features and fixed features of the child nodes of the node. For the image input by the second data enhancement model, the second data enhancement model can keep the input image unchanged on the fixed features of the child node, and change the input image based on the value of the changing features corresponding to the child node to achieve data enhancement of the input image. The second data enhancement model is also the data enhancement model corresponding to the child node. When the changing feature of the child node is the side face angle feature, the second data enhancement model is an FPN model. When the changing feature of the child node of the node is the expression feature, the second data enhancement model can be a GANimation model. The embodiment of the present application does not specifically limit the second data enhancement model.

[0274] It should be noted that the process shown in steps 508-510 is a process in which the computing device performs data enhancement processing on the image to be processed based on the data enhancement model corresponding to the node, that is, a process in which the computing device performs iterative data enhancement processing on the image based on the dependency relationship between the multiple nodes, wherein the first data enhancement model is the data enhancement model corresponding to the node, and the second data enhancement model is the data enhancement model corresponding to the child node of the node.

[0275] 511. The computing device obtains multiple images from the images generated by data enhancement of each leaf node, and forms a training set with the multiple images, wherein each leaf node is a node without child nodes.

[0276] The computing device can extract some images from the images generated by data enhancement of each leaf node in a certain proportion to form a training set. In one possible implementation, the process shown in this step 511 can be implemented by the process shown in the following steps 5111-5116.

[0277] Step 5111: The computing device sets a weight for each leaf node in the multiple nodes.

[0278] The value of each leaf node, that is, the proportion of images generated by each leaf node in the training set in the training set. The user can determine the weight of each leaf node based on his or her own experience, and the computing device sets the weight of each leaf node determined by the user to the corresponding leaf node. The computing device can also dynamically adjust the weight of each leaf node according to the process shown in steps 5114-5116.

[0279] Step 5112: For any leaf node, the computing device obtains a fourth number of images from the images generated by performing data augmentation on the leaf node based on the number of images required for the training set and the weight of the leaf, where the fourth number is the product of the number of images required for the training set and the weight of the leaf node.

[0280] For example Figure 8 A schematic diagram of data fusion provided by an embodiment of the present application is shown in FIG. Figure 8 The child nodes of node x in include node af, where node af has no child nodes but only a parent node, then node af is a leaf node, and the weights set by the computing device for node af are 30%, 10%, 5%, 20%, 5% and 30% respectively. If the training set requires 1000 images, the computing device obtains 300 images, 100 images, 50 images, 200 images, 50 images and 300 images respectively from the images generated by data enhancement at node af. It should be noted that the data enhancement process performed on each node in the data enhancement tree is equivalent to the data enhancement process performed on each computing node in the previous text.

[0281] Step 5113: The computing device combines the images obtained from each leaf node into the training set.

[0282] The computing device combines the images obtained from each leaf node into a training set through MF. The computing device can obtain the changing features and fixed features of each leaf node, and input the changing features and fixed features of each leaf node into the weight matcher. The weight matcher assigns a weight to each leaf node according to the changing features and fixed features of each leaf node. The weight matcher obtains images from the images generated by each leaf node according to the weight of each leaf, and sends the obtained images to the deployment fusion device. The deployment fusion device combines the images sent by the weight matcher into a training set. The images in the training set can meet the business writing of the business scenario. The deployment fusion device can send the training set to the data computing device, and the data computing device associates and stores the business scenario and the training set.

[0283] After the computing device completes a set of iterative data enhancement processing on each data enhancement path, the leaf node of each data enhancement path can obtain the final processing result of a set of iterative data enhancement processing on the source image. For any node on any data enhancement path, the computing device changes a business feature of the image and keeps the fixed feature of the image at the node unchanged. After iteration, the image generated at the leaf node is compared with the source image, and data enhancement is performed on the changed features of each node of the data enhancement path, so that the image generated at each leaf node is an image generated after data enhancement based on multiple business features in the business scenario. Therefore, the image generated at each leaf node meets the needs of the business scenario. Since each node corresponds to a sub-scenario under the business scenario, the image generated at each leaf node can meet the business needs of a sub-scenario of the business scenario. Since each image in the training set meets the business needs of different sub-scenarios, the training set can meet the needs of the entire business scenario.

[0284] The computing device can directly use the training set generated in step 5113 as the final training set for the business scenario, or it can dynamically adjust the weights of the leaf nodes through the process shown in the following steps 5114-5116. The computing device reassembles a training set based on the final adjusted weights, and uses the final assembled training set as the final training set for the business scenario.

[0285] It should be noted that the process shown in steps 5111-5113 is a process in which the computing device combines the partial images generated when performing data enhancement on each leaf node into a training set based on the weight corresponding to each leaf node.

[0286] Step 5114: The computing device inputs each image in the training set and the data label associated with each image into an initial image recognition model. The initial recognition model performs model training based on the input image and the input data label to obtain an image recognition model, which is used to recognize images in the business scenario.

[0287] The initial recognition model uses each image in the training set and the data label associated with each image as training data of the initial recognition model. The initial recognition model performs model training based on the training data to obtain an image recognition model.

[0288] The computing device can also use the real images in the business scenario and the images in the training set to perform model training. In a possible implementation, the computing device uses the real images in the business scenario, the data labels of each real image in the business scenario, the images in the training set, and the labels of the images in the training set as training data, and inputs them into the initial recognition model. The initial recognition model performs model training based on the input training data to obtain an image recognition model. The data label of each real image in the business scenario can be a manually annotated label. For example, the computing device can obtain the images in the training set and the data labels of the images in the training set from the data computing device, and can also obtain the real images in the business scenario and the data labels of each real image from the data computing device. The computing device performs model training based on the acquired images and the data labels of each acquired image to obtain an image recognition model.

[0289] Step 5115: The computing device inputs multiple real images in the business scenario into the image recognition model. The image recognition model performs image recognition on the input multiple real images and outputs recognition results of the multiple real images. The computing device determines the recognition effect of the image recognition model in recognizing the multiple real images based on the recognition results of the multiple real images.

[0290] The recognition result of each real image may be the target object displayed by each real image. For example, for any facial image of target object A, the image recognition model performs image recognition on the facial image of target object A working in the business scenario, and outputs the recognition result of the facial image as target object A. The recognition result of each real image may also be a specific feature on each real image, and the embodiment of the present application does not specifically limit the recognition result.

[0291] The recognition effect of the image recognition model is used to indicate the recognition of the real image in the business scenario by the image recognition model. The recognition effect can be represented by the accuracy of the recognition result. Taking the recognition result as the target object displayed by the real image as an example, if the image recognition model recognizes 100 face images in the business scenario, the computing device compares the target object displayed by the 100 face images output by the image recognition model with the target object actually displayed by the 100 face images, and determines that the number of correct recognition results is 80, then the recognition effect is 80%. Of course, the recognition effect can also be represented by other indicators, and the embodiment of the present application does not specifically limit the representation method of the recognition effect.

[0292] The process shown in step 5115 is the process of the computing device testing the target recognition model based on the real image in the business scenario. The recognition effect obtained in step 5115 can be regarded as the test result of this test.

[0293] Step 5116: When the recognition effect reaches the preset recognition effect, the computing device uses the images in the training set as training data for the business scenario, otherwise, the computing device adjusts the weight of each leaf node.

[0294] Whenever the weight of a leaf node is adjusted, whether the computing device can perform the next test, that is, the computing device can jump to execute the process shown in 5112-5116, until the trained image recognition model can achieve the preset recognition effect for multiple real images in the business scenario. Therefore, the images in the training set finally obtained can better meet the needs of the business scenario, and the recognition effect of the image recognition model finally obtained is the best. Therefore, the image recognition model finally obtained can be used as the optimal image recognition model in the business scenario. The computing device can also associate and store the optimal image recognition model with the business scenario, so that the optimal image recognition model can be queried according to the business scenario later. For different business scenarios, the preset recognition effect can be different values, and the embodiment of the present application does not specifically limit the preset recognition effect.

[0295] The method provided in the embodiment of the present application establishes a data enhancement tree according to the order of each business feature in the feature sequence, and performs iterative data enhancement processing on the image according to the dependency relationship between each node on the data enhancement tree. At any node, the computing device performs data enhancement on the image generated by the data enhancement processing of the parent node of the node based on the changing features and fixed features of the node. Since each data enhancement process performs data enhancement on a changing feature in the image, when the iterative data enhancement processing is completed, the enhanced image finally obtained is also the result of data enhancement of the image based on multiple business features in the business scenario, so the enhanced image finally obtained can meet the business needs of the business scenario. In addition, after the computing device completes a data enhancement process at a node once, the image generated by the data enhancement process can be associated with the fixed features of the node, and the fixed features of the node are used as data features, so that the fixed features of each node can be used to realize automatic labeling of the enhanced image without manual labeling. In addition, after the computing device completes a data enhancement process at a node once, the image generated by the data enhancement can also be filtered to obtain an image that meets the first preset condition, wherein the image that meets the first preset condition can meet the business needs of the next sub-scenario of the business scenario. Furthermore, when generating source images, the source images may be filtered to obtain target source images that meet the third preset condition, wherein the target source has a high degree of authenticity, and the individual features of each target source image are different, thereby including the differences between different source images. Furthermore, after the computing device completes data enhancement once at a node, it may filter the image generated by the data enhancement to obtain an image with a high degree of authenticity, and images with the same individual features have diversity, then the images in the training set finally obtained also have a high degree of authenticity, images with the same individual features have diversity, and images with different individual features have differences, thereby meeting the requirements for training data.

[0296] pass Figure 6 and Figure 7 From the process shown, it can be seen that the computing device can obtain multiple business features based on the business needs of the business scenario, and then establish a data enhancement tree based on the obtained business features; the computing device then performs iterative data enhancement processing on the source image based on the dependency relationship between the nodes on the data enhancement tree, and combines the images generated at the leaf nodes into a training set. The computing device can implement the process of obtaining a training set based on the data enhancement tree through the following three parts of pseudo code. The input data of these three parts of pseudo code include: the list of features to be changed involved in the data enhancement process F = (f1,...,f i ,...,f l ), where l is the number of service features, 1≤i≤l, f iThe value range of the i-th business feature when making data enhancement changes, f i It can include all possible feature values ​​when the i-th business feature is changed for data enhancement (that is, f i The original values ​​of each); the feature list to be changed required by the business scenario R = (r1,...,r i ,...,r l ), where f i With r i One-to-one correspondence, R l is the value range of the i-th business feature in this business scenario, R i It may include the various feature values ​​required by the business scenario when the i-th business feature is subjected to data enhancement changes; the weight list W of each business feature defined according to business prior knowledge = (w1, ..., w i ,...,w l ), where w i is the first weight assigned to the i-th business feature; the list of binary variables E for pruning the constructed initial DA-Tree is (e1,...,e n ), the binary variable list is the deletion information, where n is the number of initial nodes on the data enhancement tree; the number of predefined data generated for each node of DA-Tree according to business needs is A=(a1,...,a n ), which is used to limit the number of nodes in the data enhancement tree.

[0297] The pseudo code of Part 1 includes lines 1 to 6 of pseudo code, which are used to implement the process of sorting multiple business feature lines to obtain feature sequences. Line 1 in Part 1 defines an empty feature list F' for the feature sorting result; Line 2 defines a sequence list O for the calculation result of the feature sorting; Lines 3 to 5 calculate the following formula (1) and store the calculation result of formula (1) in the sequence list O; Line 6 sorts F according to the sorting value of the sequence list O and the corresponding relationship between O and the elements in F, and saves it in F'.

[0298]

[0299] The pseudo code of Part 2 includes lines 7 to 25, which are used to build a data enhancement tree, where line 7 creates an empty DA-Tree (including a newly generated root node); line 8 obtains the root node of the data enhancement tree; lines 9 to 10 create two layer pointers, which are used to expand the DA-Tree layer by layer later; the loop from lines 11 to 23 expands the DA-Tree layer by layer, where lines 13 to 20 expand the child nodes of all nodes in the current layer one by one, and lines 14 to 18 expand the child nodes of the current node in the current layer. After completing the preliminary construction of the DA-Tree, line 24 prunes the tree accordingly according to the business prior and configures the number of data to be generated on each node, and line 25 updates the corresponding fixed features on each node according to the pruned DA-Tree (that is, updates the fixed features of each node according to the dependencies between nodes).

[0300] The pseudo code of Part 3 includes lines 26 to 31, which are used to implement data enhancement at each node, where line 26 is to re-acquire the root node of the generated DA-Tree, lines 27-30 are to traverse the DA-Tree according to the tree traversal algorithm (such as breadth-first or depth-first algorithm), and complete the corresponding data generation operation at each node; line 31 is to return the data (enhanced image) finally generated at the leaf node. It should be noted that the generated data of each leaf node is the data output of each data enhancement path for each sub-business scenario. Of course, if the business requires, the data generated on the non-leaf nodes on each data enhancement path can also be saved and used.

[0301]

[0302]

[0303] For further explanation of this process, see Fig. 9 A flowchart of a data enhancement method provided in an embodiment of the present application is shown, and the flow of the method includes the process shown in steps 901-914.

[0304] 901. The computing device obtains the business requirements of the business scenario, and decomposes the business requirements to obtain multiple business features.

[0305] The process shown in step 901 is the same as the process shown in step 401, and step 901 is not described in detail in this embodiment of the present application.

[0306] 902. The computing device sorts the multiple business features to obtain a feature sequence.

[0307] The process shown in step 902 is the same as the process shown in step 402, and step 902 is not described in detail in this embodiment of the present application.

[0308] 903. The computing device establishes a data enhancement tree according to the feature sequence, where each node on the data enhancement tree corresponds to a business feature in the feature sequence.

[0309] The process shown in step 903 is the same as the process shown in steps 501 and 503-505, and step 903 is not described in detail in this embodiment of the present application.

[0310] 904. The computing device determines the fixed features and the variable features of each node according to the dependency relationship between the nodes on the data enhancement tree.

[0311] The process shown in step 904 is the same as the process shown in step 502, and step 904 is not described in detail in this embodiment of the present application.

[0312] 905. The computing device determines a data enhancement model corresponding to each node based on the fixed features and variable features of each node.

[0313] The data enhancement model corresponding to any node is used to keep the fixed features of the input image at the node unchanged, change the changing features of the node, and output the enhanced image, so that the data enhancement model corresponding to each node can realize single-point data enhancement technology.

[0314] 906. The computing device generates a source image based on a target data enhancement model trained with real images.

[0315] The process shown in step 906 is the same as the process shown in steps 403-404, and step 906 is not described in detail in this embodiment of the present application.

[0316] 907. The computing device performs data enhancement processing on the image generated by the parent node of each node in turn at each node according to the dependency relationship between the nodes and the data enhancement model corresponding to each node.

[0317] The process shown in step 907 is the same as the process shown in steps 508-510, and step 907 is not described in detail in this embodiment of the present application.

[0318] 908. The computing device sets a weight for each leaf node according to the business requirements of the business scenario.

[0319] The process shown in step 908 is the same as the process shown in step 5111, and step 908 is not described in detail in this embodiment of the present application.

[0320] 909. The computing device combines the images generated by data enhancement at each leaf node into a training set according to the weight of each leaf node.

[0321] The process shown in step 909 is the same as the process shown in steps 5112-5113, and the present embodiment of the application does not repeat step 909. After the computing device obtains the training set, the training set can also be stored.

[0322] 910. The computing device obtains an image recognition model based on image training in the training set, and tests the image recognition model using multiple real images in the business scenario to obtain recognition effects of the image recognition model on the multiple real images.

[0323] The process shown in step 910 is the same as the process shown in steps 5114-5115, and step 910 is not described in detail in this embodiment of the present application.

[0324] 911. If the recognition effect reaches the preset recognition effect, the computing device publishes the training set, otherwise execute step 908.

[0325] In a practical application, the data enhancement method of the present application is applied to the business scenario of black face image recognition. It is well known in the industry that the accuracy of face recognition of dark-skinned people is relatively low, and the existing public data sets are mostly based on white people (also known as Caucasians), which brings challenges to face recognition of dark-skinned people. The following comparative experiments were conducted using the method of the present application. The specific 6 groups of comparative experiments are shown in Table 1. Among them, the first to fourth groups of experiments were conducted using Microsoft face data (Microsoft-Celebrity-1million, MS-Celeb-1M), Visual Geometry Group Face 2from University of Oxford, and the third group of experiments using Microsoft face data (Microsoft-Celebrity-1million, MS-Celeb-1M). Oxford, VGGFace2), MS-Celeb-1M and (VGGFace2+MS-Celeb-1M) are used as training data, and the experimental results of the first group of experiments are used as the baseline for comparison with the experimental results of other groups. Groups 5 and 6 are two groups of implementations using the data enhancement method in this application; the accuracy of the face recognition model trained by the training set of face images (face data enhanced by DA-Tree data) obtained by the data enhancement method of this application can be improved by 4.14% (from 88.56% in Experiment 1 to 92.7% in Experiment 6) compared with the original data set, and can be improved by 2.3% (from 90.40% in Experiment 3 to 92.70% in Experiment 6) compared with the black face data set in the original data set. For specific accuracy comparison, please refer to Table 1. In this application example, the recognition accuracy of the face recognition model obtained by the method of this application on dark races has exceeded the recognition accuracy of the most advanced face recognition algorithm model in the industry on white races.

[0326] Table 1

[0327]

[0328] Fig.10 : is a structural schematic diagram of a data enhancement device provided in an embodiment of the present application, the device comprising:

[0329] An acquisition module 1001 is used to acquire a feature sequence of a business scenario, where the feature sequence is used to represent a data enhancement processing order of multiple business features in the business scenario;

[0330] A processing module 1002 is used to perform iterative data enhancement processing on the image according to the order of each business feature in the feature sequence, and one data enhancement processing process corresponds to one business feature in the feature sequence;

[0331] Among them, in any data enhancement processing process, data enhancement is performed on the image generated by the previous data enhancement processing based on the change characteristics in the data enhancement processing process, and the change characteristics are the business characteristics corresponding to the data enhancement processing process.

[0332] Optionally, the processing module 1002 is used to:

[0333] The feature value of the changing feature in the image generated by the last data enhancement processing is changed, and the fixed feature in the image generated by the last data enhancement processing is kept unchanged, and the fixed feature is the business feature before the business feature corresponding to the data enhancement processing process in the feature sequence.

[0334] Optionally, the acquisition module 1001 includes:

[0335] The sorting unit is used to sort the multiple business features based on the number of values ​​of each business feature in the business scenario and the number of original values ​​of each business feature to obtain the feature sequence.

[0336] Optionally, the sorting unit is used to:

[0337] Calculate the ratio between the number of values ​​of each business feature in the business scenario and the number of original values ​​of each business feature to obtain the ratio of each business feature, where the ratio of a business feature is the proportion of valid images in an image generated when the business feature is used as a change feature for data enhancement;

[0338] According to the size of the ratio of each business feature, the multiple business features are sorted in ascending order.

[0339] Optionally, the device further comprises:

[0340] A receiving module, used to receive a data enhancement request, where the data enhancement request is used to instruct to generate an image that meets the business scenario requirements;

[0341] A splitting module, used to split the comprehensive business features of the business scenario carried by the data enhancement request to obtain multiple business features of the business scenario;

[0342] The first determination module is used to determine various values ​​of each business feature in the business scenario based on the business requirements of the business scenario.

[0343] Optionally, the processing module 1002 is used to:

[0344] Based on the order of each business feature in the feature sequence and each node on the data enhancement tree, iterative data enhancement processing is performed on the image. The data enhancement tree includes multiple nodes, and each node corresponds to a data enhancement processing process.

[0345] Optionally, the processing module 1002 includes:

[0346] An establishing unit, configured to establish a data enhancement tree based on the order of each business feature in the feature sequence, wherein each business feature corresponds to at least one node, wherein a node corresponding to any business feature is a parent node of at least one target node among the multiple nodes, and the at least one target node corresponds to a business feature after any business feature in the feature sequence;

[0347] A processing unit is used to perform iterative data enhancement processing on the image based on the dependency relationship between the multiple nodes.

[0348] Optionally, the device further comprises:

[0349] The second determination module is used to determine the change characteristics of each node based on the dependency relationship between the multiple nodes.

[0350] Optionally, the first determining unit is used to:

[0351] The service feature corresponding to each node is determined as the change feature of each node.

[0352] Optionally, the first determining unit is used to determine the changing characteristics of each node and the fixed characteristics of each node based on the dependency relationship between the multiple nodes.

[0353] In a possible implementation manner, the first determining unit is configured to:

[0354] Determining the service feature corresponding to each node as the change feature of each node;

[0355] The fixed features and the changing features of the parent node of each node are determined as the fixed features of each node.

[0356] Optionally, the establishing unit is further used for:

[0357] According to the fixed features of each node, a data tag of each initial node is established, and the data tag of a node includes the fixed features of the node.

[0358] Optionally, the establishing unit is used to:

[0359] Based on the order of each business feature in the feature sequence and the deletion information corresponding to the business scenario, the initial nodes on the initial data enhancement tree are deleted to obtain the data enhancement tree, and each business feature corresponds to at least one initial node on the initial data enhancement tree.

[0360] Optionally, the establishing unit includes:

[0361] Establishing a subunit, used to establish an initial data enhancement tree according to the order of each business feature in the feature sequence, wherein the initial data enhancement tree includes a plurality of initial nodes, each business feature corresponds to at least one initial node, an initial node corresponding to any business feature is a parent node of at least one target initial node among the plurality of initial nodes, and the at least one target initial node corresponds to the next business feature of the business feature in the feature sequence;

[0362] The deletion subunit is used to delete the initial nodes on the initial data enhancement tree according to the deletion information corresponding to the business scenario to obtain the data enhancement tree, wherein the deletion information is used to indicate the initial nodes to be deleted on the initial data enhancement tree.

[0363] Optionally, the device further comprises:

[0364] The first generation module is used to generate at least one initial node corresponding to each business feature in sequence according to the order of each business feature in the feature sequence to obtain the initial data enhancement tree, wherein the initial node corresponding to a business feature is the parent node of the initial node corresponding to the next business feature of the business feature.

[0365] Optionally, the first generating module is used to:

[0366] For the first service feature in the feature sequence, generating a root node of the initial data enhancement tree;

[0367] For any business feature other than the first business feature in the feature sequence, a first number of node sets of the initial data enhancement tree are generated, the first number being the number of values ​​taken by the previous business feature of any business feature in the feature sequence in the business scenario, each node set including a second number of initial nodes, the nodes in each node set being child nodes of an initial node corresponding to the previous business feature, and the second data amount being the number of values ​​taken by any business feature in the business scenario.

[0368] Optionally, the device further comprises:

[0369] An updating module, for updating, for an initial node whose parent node is deleted, the fixed features of the initial node to the fixed features and the changing features of the current parent node of the initial node, wherein the current parent node of the initial node is the previous node of the initial node on the data enhancement path to which the initial node belongs on the data enhancement tree;

[0370] The updating module is further used to update the fixed features on the data tag of the initial node to the updated fixed features of the initial node.

[0371] Optionally, the processing unit is used to:

[0372] For any node among the multiple nodes, data enhancement processing is performed on an image to be processed based on a data enhancement model corresponding to the node, and the image to be processed is an image generated when data enhancement is performed on a parent node of the node.

[0373] Optionally, the processing unit is used to:

[0374] For any node among the multiple nodes, inputting the image to be processed into a first data enhancement model, and the first data enhancement model performs data enhancement on the input image based on the change characteristics of the node, and outputs an enhanced image, wherein the image to be processed is an image generated when data enhancement is performed on the parent node of the node;

[0375] For the child nodes of the node, the image output by the first data enhancement model is input into the second data enhancement model, and the second data enhancement model performs data enhancement on the input image based on the change characteristics of the child nodes.

[0376] Optionally, the device further comprises:

[0377] The combination module is used to combine partial images generated when data enhancement is performed on each leaf node into a training set based on the weight corresponding to each leaf node, and each leaf node is a node without child nodes.

[0378] Optionally, the combined module is used for:

[0379] Setting a weight for each leaf node in the plurality of nodes;

[0380] For any leaf node, according to the number of images required for the training set and the weight of the leaf, obtaining a fourth number of images from the images generated by performing data augmentation on the leaf node, where the fourth number is the product of the number of images required for the training set and the weight of the leaf node;

[0381] The images obtained from each leaf node are combined into the training set.

[0382] Optionally, the device further comprises:

[0383] A training module is used to input each image in the training set and the data label associated with each image into an initial image recognition model, and the initial recognition model performs model training based on the input image and the input data label to obtain an image recognition model. The image recognition model is used to recognize images in the business scenario, and the data label associated with each image is the data label of the node where each image is located.

[0384] Optionally, the training module is further used to:

[0385] Inputting a plurality of real images in the business scenario into the image recognition model, and having the image recognition model perform image recognition on the inputted plurality of real images, and outputting recognition results of the plurality of real images;

[0386] Based on the recognition results of the multiple real images, determining the recognition effect of the image recognition model in recognizing the multiple real images;

[0387] When the recognition effect reaches a preset recognition effect, the images in the training set are used as training data in the business scenario; otherwise, the weight of each leaf node is adjusted.

[0388] Optionally, the device further comprises:

[0389] A first filtering module, used for filtering the image generated by the data enhancement process to obtain an image that meets a first preset condition;

[0390] The third determination module is used to determine the image that meets the first preset condition as the image to be processed in the next data enhancement processing process.

[0391] Optionally, the first preset condition includes at least one of the following:

[0392] The realism of an image reaches a preset realism;

[0393] The similarity between images having the same individual feature is greater than or equal to the first similarity;

[0394] The similarity between images of different individual features is less than or equal to the second similarity.

[0395] Optionally, the second determining module is used to:

[0396] Filter the image that meets the first preset condition to obtain an image that meets the second preset condition, where the second preset condition includes a feature value of a change feature of an image during the data processing being a value of the change feature in the business scenario;

[0397] The image meeting the second preset condition is determined as the image to be processed in the next data enhancement processing process.

[0398] Optionally, the device further comprises:

[0399] A second generating module is used to input the target number into a target data enhancement model, and the target data enhancement model generates source images of the target number based on the input target number;

[0400] A second filtering module, used for filtering the target number of source images to obtain source images that meet a third preset condition;

[0401] The fourth determining module is used to use the source image that meets the third preset condition as the image to be processed in the first data enhancement processing process.

[0402] Optionally, the third preset condition includes at least one of the following:

[0403] The similarities between a source image and source images other than the source image in the target number of source images are all less than or equal to a third similarity;

[0404] The similarities between a source image and a plurality of real images are all less than or equal to a fourth similarity, and the plurality of real images are used for training to obtain the target data enhancement model;

[0405] The realism of a source image reaches a preset realism.

[0406] Optionally, the device further comprises:

[0407] An association module is used to associate the image generated by the data enhancement process with the data label of the data enhancement process, and the data label of the data enhancement process includes the fixed features of the data enhancement process.

[0408] The data enhancement device performs iterative data enhancement processing on the image according to the order of each business feature in the feature sequence, wherein, in any data enhancement processing process, based on the changing features in the data enhancement processing process and the fixed features in the data enhancement processing process, the image generated by the previous data enhancement processing is data enhanced. Since each data enhancement processing process performs data enhancement on a changing feature in the image, when the iterative data enhancement processing is completed, the enhanced image finally obtained is the result of data enhancement of the image based on multiple business features in the business scenario, so the enhanced image finally obtained can meet the business needs of the business scenario.

[0409] All the above optional technical solutions can be arbitrarily combined to form optional embodiments of the present disclosure, and will not be described in detail here.

[0410] It should be noted that: the data enhancement device provided in the above embodiment only uses the division of the above functional modules as an example when performing data enhancement. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the data enhancement method embodiments provided in the above embodiment belong to the same concept, and the specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0411] In a possible example, the above data enhancement method can also be implemented in the form of a chip, which can be Fig.11 The chip shown, Fig.11 1 is a schematic diagram of the structure of a chip provided in an embodiment of the present application, and the chip includes a neural network processor 1100. The chip can be set in the computing device provided in the embodiment of the present application to complete the work performed by the computing device in the present application.

[0412] The neural network processor (NPU) 1100 can be mounted on the main CPU (host CPU) as a coprocessor, and the host CPU assigns tasks. The core part of the NPU is the operation circuit 1103, which is controlled by the controller 1104 to extract the matrix data in the memory (input memory 1101 or weight memory 1102) and perform multiplication operations. Among them, the input memory 1101 is used to store input data, which can be the number of real images required or the real images in the business scenario. The weight memory 1102 is used to store the weights in each data enhancement model, and is also used to store the weights of each leaf node, and is also used to store the first weight of each business feature.

[0413] In some implementations, the operation circuit 1103 includes multiple processing units (PEs) inside. In some implementations, the operation circuit 1103 is a two-dimensional systolic array. The operation circuit 1103 can also be a one-dimensional systolic array or other electronic circuits capable of performing mathematical operations such as multiplication and addition. In some implementations, the operation circuit 1103 can be a general-purpose matrix processor.

[0414] For example, assume that there is an input matrix A (which may be composed of image data in a training set), a weight matrix B (which may include weights in an initial image recognition model), and an output matrix C. The operation circuit 1103 takes the data in the matrix B from the weight memory 1102 and caches it on each PE in the operation circuit 1103. The operation circuit 1103 takes the data in the matrix A from the input memory 1101 and performs a matrix operation on the data in the matrix B, and the partial result or final result of the matrix is ​​stored in the accumulator 1108. The accumulator 1108 is used to sum the partial results of the matrix.

[0415] The vector calculation unit 1107 can further process the output of the operation circuit 1103, such as vector multiplication, vector addition, exponential operation, logarithmic operation, size comparison, etc. For example, the vector calculation unit 1107 can be used to implement network calculations when the initial image recognition model recognizes real images, such as pooling, batch normalization, local response normalization, etc.

[0416] In some implementations, the vector calculation unit 1107 may store the vector of processed outputs to the unified memory 1106. For example, the vector calculation unit 1107 may apply a nonlinear function to the output of the operation circuit 1103 and store it to the unified memory 1106, such as a vector of accumulated values ​​to generate an activation value. In some implementations, the vector calculation unit 1107 generates a normalized value, a merged value, or both. In some implementations, the vector of processed outputs of the vector calculation unit 1107 can be used as an activation input of the operation circuit 1103, for example, for use in a subsequent layer in a neural network.

[0417] The unified memory 1106 is used to store input data and output data.

[0418] The NPU can directly transfer the input data in the external memory to the input memory 1101 and / or the unified memory 1106 through the direct memory access controller 1105 (DMAC), and can also store the weight data in the external memory into the weight memory 1102, and can also store the data in the unified memory 1106 into the external memory. The input data can be an image in the training set, and the weight data can be the weight of the leaf node, and can also be the weight used to sort the service features.

[0419] The bus interface unit (BIU) 1110 is used to implement interaction between the main CPU, DMAC and instruction fetch memory 1109 through the bus.

[0420] An instruction fetch buffer 1109 connected to the controller 1104 is used to store instructions used by the controller 1104 .

[0421] The controller 1104 is used to call the instructions cached in the instruction fetch memory 1109 to implement the working process of the control operation accelerator.

[0422] Generally, the unified memory 1106, the input memory 1101, the weight memory 1102 and the instruction fetch memory 1109 are all on-chip memories, and the external memory is a memory outside the NPU, which can be a double data rate synchronous dynamic random access memory (DDR SDRAM), a high bandwidth memory (HBM) or other readable and writable memory.

[0423] It should be noted that the steps implemented by the above data enhancement method can be completed by the main CPU and the NPU in cooperation. Fig.11 The structure of the chip shown is only an illustrative example. Of course, the chip can also have other structures. The embodiments of the present application do not limit the specific structure of the chip.

[0424] In one possible implementation, an embodiment of the present application provides a chip, which includes a processor and a data interface. The processor reads instructions stored in a memory through the data interface to execute the above-mentioned data enhancement method.

[0425] Optionally, as an implementation method, the chip may further include a memory, in which instructions are stored, and the processor is used to execute the instructions stored in the memory. When the instructions are executed, the processor is used to execute the above-mentioned data enhancement method.

[0426] In one possible implementation, a computer program product including instructions is provided. When the computer program product is executed on a computing device, the computing device executes the data enhancement method provided in the method embodiment of the present application.

[0427] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0428] It should be noted that those skilled in the art can realize that the units, modules, chips and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to the function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0429] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system, device, module, unit or chip described above can refer to the corresponding process in the data enhancement method embodiment, and will not be repeated here.

[0430] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of modules within the device or units in the modules is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or units can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, or it can be an electrical, mechanical or other form of connection.

[0431] The modules or units described as separate components may or may not be physically separated, and the components displayed as modules or units may or may not be physical modules or physical units, that is, they may be located in one place, or may be distributed to multiple computing devices or chips. Some or all of the modules or units may be selected according to actual needs to achieve the purpose of the embodiments of the present invention.

[0432] In addition, each functional module or unit in each embodiment of the present invention may be integrated into a target processing module, or each module or unit may exist physically separately, or two or more modules or units may be integrated into a target processing module. The above-mentioned integrated modules or units may be implemented in the form of hardware or in the form of software functional units.

[0433] The above description is only an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A data enhancement method, characterized in that: The method comprises: Acquire a feature sequence of a business scenario, where the feature sequence is used to represent a data enhancement processing order of multiple business features in the business scenario; According to the order of each business feature in the feature sequence, iterative data enhancement processing is performed on the image, and one data enhancement processing process corresponds to one business feature in the feature sequence; Among them, in any data enhancement processing process, data enhancement is performed on the image generated by the previous data enhancement processing based on the change characteristics in the data enhancement processing process, and the change characteristics are the business characteristics corresponding to the data enhancement processing process.

2. The method according to claim 1, characterized in that Based on the change characteristics in the data enhancement process, data enhancement is performed on the image generated by the last data enhancement process, including: The feature value of the changing feature in the image generated by the last data enhancement processing is changed, and the fixed feature in the image generated by the last data enhancement processing is kept unchanged, and the fixed feature is the business feature before the business feature corresponding to the data enhancement processing process in the feature sequence.

3. The method according to claim 1, characterized in that: The feature sequence of obtaining the business scenario includes: Based on the number of values ​​that each business feature takes in the business scenario and the number of original values ​​that each business feature takes, the multiple business features are sorted to obtain the feature sequence.

4. The method according to claim 3, characterized in that The sorting of the multiple business features based on the number of values ​​of each business feature in the business scenario and the number of original values ​​of each business feature includes: Calculate the ratio between the number of values ​​of each business feature in the business scenario and the number of original values ​​of each business feature to obtain the ratio of each business feature, where the ratio of a business feature is the proportion of valid images in an image generated when the business feature is used as a change feature for data enhancement; According to the size of the ratio of each business feature, the multiple business features are sorted in ascending order.

5. The method according to claim 1, characterized in that Before acquiring the feature sequence of the business scenario, the method further includes: receiving a data enhancement request, wherein the data enhancement request is used to instruct generation of an image that meets the business scenario requirements; Splitting the comprehensive business features of the business scenario carried in the data enhancement request to obtain multiple business features of the business scenario; Based on the business requirements of the business scenario, respective values ​​of each business feature in the business scenario are determined.

6. The method according to claim 1, characterized in that The iterative data enhancement processing of the image according to the order of each business feature in the feature sequence includes: Based on the order of each business feature in the feature sequence and each node on the data enhancement tree, iterative data enhancement processing is performed on the image. The data enhancement tree includes multiple nodes, and each node corresponds to a data enhancement processing process.

7. The method according to claim 6, characterized in that The iterative data enhancement processing of the image based on the order of each business feature in the feature sequence and each node on the data enhancement tree includes: Based on the order of each business feature in the feature sequence, the data enhancement tree is established, each business feature corresponds to at least one node, wherein a node corresponding to any business feature is a parent node of at least one target node among the multiple nodes, and the at least one target node corresponds to a business feature after any business feature in the feature sequence; Based on the dependency relationship between the multiple nodes, iterative data enhancement processing is performed on the image.

8. The method according to claim 7, characterized in that Before performing iterative data enhancement processing on the image based on the dependency relationship between the multiple nodes, the method further includes: Based on the dependency relationship between the multiple nodes, a change characteristic of each node is determined.

9. The method according to claim 7, characterized in that: The step of establishing a data enhancement tree based on the order of each service feature in the feature sequence comprises: Based on the order of each business feature in the feature sequence and the deletion information corresponding to the business scenario, the initial nodes on the initial data enhancement tree are deleted to obtain the data enhancement tree, and each business feature corresponds to at least one initial node on the initial data enhancement tree.

10. The method according to claim 9, characterized in that The method further comprises: According to the order of each business feature in the feature sequence, at least one initial node corresponding to each business feature is generated in turn to obtain the initial data enhancement tree, and the initial node corresponding to a business feature is the parent node of the initial node corresponding to the next business feature of the business feature.

11. The method according to claim 7, characterized in that The iterative data enhancement processing of the image based on the dependency relationship between the multiple nodes includes: For any node among the multiple nodes, data enhancement processing is performed on an image to be processed based on a data enhancement model corresponding to the node, and the image to be processed is an image generated when data enhancement is performed on a parent node of the node.

12. The method according to claim 7, characterized in that After performing iterative data enhancement processing on the image based on the dependency relationship between the multiple nodes, the method further includes: Based on the weight corresponding to each leaf node, partial images generated when data enhancement is performed on each leaf node are combined into a training set, and each leaf node is a node without child nodes.

13. The method according to claim 1, characterized in that After performing data enhancement on the image generated by the last data enhancement process, the method further includes: Performing filtering processing on the image generated by the data enhancement processing process to obtain an image that meets the first preset condition; The image that meets the first preset condition is determined as the image to be processed in the next data enhancement processing process.

14. The method according to claim 13, characterized in that The step of determining the image meeting the first preset condition as the image to be processed in the next data enhancement process comprises: Filter the image that meets the first preset condition to obtain an image that meets the second preset condition, where the second preset condition includes a feature value of a change feature of an image during the data enhancement process being a value of the change feature in the business scenario; The image meeting the second preset condition is determined as the image to be processed in the next data enhancement processing process.

15. The method according to claim 1, characterized in that Before performing iterative data enhancement processing on the image according to the order of each business feature in the feature sequence, the method further includes: Inputting the target number into a target data enhancement model, and generating source images of the target number based on the input target number by the target data enhancement model; Performing filtering processing on the target number of source images to obtain source images that meet a third preset condition; The source image that meets the third preset condition is used as the image to be processed in the first data enhancement processing process.

16. The method according to claim 1, characterized in that In any data enhancement process, after performing data enhancement on the image generated by the previous data enhancement process based on the change characteristics in the data enhancement process, the method further includes: An image generated by the data enhancement process is associated with a data label of the data enhancement process, wherein the data label of the data enhancement process includes a fixed feature of the data enhancement process.

17. A data enhancement device, characterized in that: The device comprises: An acquisition module, used to acquire a feature sequence of a business scenario, wherein the feature sequence is used to represent a data enhancement processing order of multiple business features in the business scenario; A processing module, used for iteratively performing data enhancement processing on the image according to the order of each business feature in the feature sequence, wherein one data enhancement processing process corresponds to one business feature in the feature sequence; Among them, in any data enhancement processing process, data enhancement is performed on the image generated by the previous data enhancement processing based on the change characteristics in the data enhancement processing process, and the change characteristics are the business characteristics corresponding to the data enhancement processing process.

18. The device according to claim 17, characterized in that The processing module is also used for: The feature value of the changing feature in the image generated by the last data enhancement processing is changed, and the fixed feature in the image generated by the last data enhancement processing is kept unchanged, and the fixed feature is the business feature before the business feature corresponding to the data enhancement processing process in the feature sequence.

19. A computing device, characterized in that The computing device includes a processor and a memory, wherein the memory stores at least one instruction, and the instruction is loaded and executed by the processor to implement the operation performed by the data enhancement method according to any one of claims 1 to claim 16.

20. A computer storage medium, characterized in that The storage medium stores at least one instruction, which is loaded and executed by the processor to implement the operation performed by the data enhancement method according to any one of claims 1 to 16.

21. A chip, characterized in that: The chip includes a processor and a data interface, and the processor reads instructions stored in the memory through the data interface. The instructions are loaded and executed by the processor to implement the operations performed by the data enhancement method according to any one of claims 1 to claim 16.

Images