Image processing model generation method, device, equipment, medium and program product

Abstract

The present disclosure relates to a method, apparatus, device, medium and program product for generating an image processing model. The method comprises: searching, based on a preset search space, at least two candidate combinations for generating an image processing model; wherein the preset search space comprises at least two of a first sub-search space, a second sub-search space and a third sub-search space; the first sub-search space is used for searching data cleaning corresponding hyperparameters, wherein the data cleaning corresponding hyperparameters are used for data cleaning of a training image set; the second sub-search space is used for searching loss function corresponding hyperparameters; the third sub-search space is used for searching network structure corresponding hyperparameters; training based on the at least two candidate combinations respectively to obtain at least two candidate image processing models corresponding to the at least two candidate combinations one by one; and determining a target image processing model based on the at least two candidate image processing models.

Description

Technical Field

[0001] This disclosure relates to the field of computer vision technology, and in particular to a method for generating an image processing model, an apparatus for generating an image processing model, an electronic device, a storage medium, and a program product. Background Technology

[0002] In the field of computer vision technology, how to train image processing models on large-scale training image sets to enable them to learn robust image feature extraction capabilities and thus achieve accurate image processing is a pressing technical problem. For example, face recognition is a fundamental problem with significant practical value. How to train face recognition models on large-scale training image sets to enable them to learn robust image feature extraction capabilities and thus achieve accurate face recognition is a crucial topic in the field of computer vision technology. Summary of the Invention

[0003] This disclosure provides a technical solution for generating image processing models.

[0004] According to one aspect of this disclosure, a method for generating an image processing model is provided, comprising:

[0005] Based on a preset search space, at least two candidate combinations for generating an image processing model are obtained; wherein the preset search space includes at least two items from a first sub-search space, a second sub-search space, and a third sub-search space; the first sub-search space is used to search for hyperparameters corresponding to data cleaning, wherein the hyperparameters corresponding to data cleaning are used to clean the training image set; the second sub-search space is used to search for hyperparameters corresponding to the loss function; and the third sub-search space is used to search for hyperparameters corresponding to the network structure.

[0006] Training is performed based on the at least two candidate combinations to obtain at least two candidate image processing models that correspond one-to-one with the at least two candidate combinations.

[0007] Based on the at least two candidate image processing models, the target image processing model is determined.

[0008] By searching a preset search space, at least two candidate combinations for generating an image processing model are obtained. The preset search space includes at least two items from a first sub-search space, a second sub-search space, and a third sub-search space. The first sub-search space is used to search for hyperparameters corresponding to data cleaning, which are used to clean the training image set. The second sub-search space is used to search for hyperparameters corresponding to the loss function. The third sub-search space is used to search for hyperparameters corresponding to the network structure. Training is performed based on the at least two candidate combinations to obtain at least two candidate image processing models that correspond one-to-one with the at least two candidate combinations. Based on the at least two candidate image processing models, the target image processing model is determined. Thus, a neural network search method is used to jointly search for at least two items from the hyperparameters corresponding to data cleaning, the loss function, and the network structure, thereby generating a more robust image processing model for a specific image processing task.

[0009] In one possible implementation, the hyperparameters corresponding to the data cleaning include at least one of the following:

[0010] Confidence thresholds for filtering training images within a class, and similarity thresholds for merging training images from different classes.

[0011] In this implementation, by searching for at least one of the confidence threshold for filtering training images within a class and the similarity threshold for merging training images of different classes, a search-based data cleaning method is adopted, which can effectively alleviate the problem in related technologies of difficulty in determining whether training images with edges belong to complex positive samples or noise samples.

[0012] In one possible implementation, the hyperparameters corresponding to the loss function include at least one of the following:

[0013] The interval in the loss function, the magnitude corresponding to the positive sample in the loss function, and the magnitude corresponding to the negative sample in the loss function.

[0014] In this implementation, for any one of the at least two candidate combinations, the loss function corresponding to the candidate combination can be determined based on at least one of the interval in the loss function of the candidate combination, the magnitude corresponding to the positive sample in the loss function, and the magnitude corresponding to the negative sample in the loss function.

[0015] In one possible implementation, the hyperparameters corresponding to the network structure include at least one of the following:

[0016] The width and depth of the backbone network.

[0017] In this implementation, the performance of the image processing model can be improved by searching for at least one of the width and depth of the backbone network and training the image processing model based on the network structure obtained by the search.

[0018] In one possible implementation,

[0019] The process of searching for at least two candidate combinations for generating an image processing model based on a preset search space includes: using a search strategy through a reinforcement learning agent to search for at least two candidate combinations for generating an image processing model based on a preset search space.

[0020] The step of training based on the at least two candidate combinations to obtain at least two candidate image processing models corresponding to the at least two candidate combinations includes: training based on the at least two candidate combinations for a preset number of rounds to obtain at least two candidate image processing models corresponding to the at least two candidate combinations.

[0021] The method further includes: obtaining reward values ​​corresponding to the at least two candidate image processing models based on the verification image set; and updating the search strategy according to the reward values.

[0022] In this implementation, a reinforcement learning agent employs a search strategy to search for at least two candidate combinations for generating an image processing model based on a preset search space. The agent then performs a preset number of training rounds based on each of these at least two candidate combinations to obtain at least two candidate image processing models that correspond one-to-one with each combination. Based on a validation image set, reward values ​​corresponding to the at least two candidate image processing models are obtained, and the search strategy is updated according to these reward values. Thus, with training, the reinforcement learning agent learns a more robust search strategy, enabling it to search for higher-quality candidate combinations and generate higher-quality image processing models.

[0023] In one possible implementation, the reinforcement learning agent employs a search strategy based on a preset search space to obtain at least two candidate combinations for generating the image processing model, including:

[0024] By employing a search strategy through a reinforcement learning agent, a preset search space is searched based on a preset search order to obtain at least two candidate combinations for generating an image processing model. The preset search order refers to the order in which each sub-search space is searched in advance.

[0025] In this implementation, a reinforcement learning agent employs a search strategy to search a preset search space based on a preset search order, thereby obtaining at least two candidate combinations for generating an image processing model, which improves the speed of joint search.

[0026] In one possible implementation, obtaining the reward values ​​corresponding to the at least two candidate image processing models based on the verification image set includes:

[0027] For any one of the at least two candidate image processing models, the accuracy and computational cost of the candidate image processing model are determined based on the validation image set.

[0028] The reward value corresponding to the candidate image processing model is determined based on the accuracy and computational cost of the candidate image processing model.

[0029] In this implementation, for any one of the at least two candidate image processing models, the accuracy and computational cost of the candidate image processing model are determined based on the verification image set, and the reward value of the candidate image processing model is determined based on the accuracy and computational cost. This allows for a balance between accuracy and computational cost when updating the search strategy.

[0030] In one possible implementation, the number of candidate image processing models is at least three;

[0031] The step of determining the target image processing model based on the at least two candidate image processing models includes:

[0032] Based on the reward value, at least two alternative image processing models are determined from the at least three candidate image processing models, wherein the number of alternative image processing models is less than the number of candidate image processing models;

[0033] The at least two candidate image processing models are trained separately until the at least two candidate image processing models meet the preset training conditions.

[0034] Based on the performance information of at least two candidate image processing models that have been trained, a target image processing model is determined from the at least two candidate image processing models that have been trained.

[0035] In this implementation, at least two candidate image processing models are determined from the at least three candidate image processing models based on the reward value. The at least two candidate image processing models are trained separately until they meet the preset training conditions. Based on the performance information of the at least two trained candidate image processing models, the target image processing model is determined from the at least two trained candidate image processing models. This reduces the number of fully trained image processing models and thus improves the speed of generating the target image processing model.

[0036] According to one aspect of this disclosure, an apparatus for generating an image processing model is provided, comprising:

[0037] The search module is used to search for at least two candidate combinations for generating an image processing model based on a preset search space; wherein the preset search space includes at least two items from a first sub-search space, a second sub-search space, and a third sub-search space; the first sub-search space is used to search for hyperparameters corresponding to data cleaning, wherein the hyperparameters corresponding to data cleaning are used to clean the training image set; the second sub-search space is used to search for hyperparameters corresponding to the loss function; and the third sub-search space is used to search for hyperparameters corresponding to the network structure.

[0038] The training module is used to train on the at least two candidate combinations respectively to obtain at least two candidate image processing models that correspond one-to-one with the at least two candidate combinations.

[0039] The determination module is used to determine the target image processing model based on the at least two candidate image processing models.

[0040] In one possible implementation, the hyperparameters corresponding to the data cleaning include at least one of the following:

[0041] Confidence thresholds for filtering training images within a class, and similarity thresholds for merging training images from different classes.

[0042] In one possible implementation, the hyperparameters corresponding to the loss function include at least one of the following:

[0043] The interval in the loss function, the magnitude corresponding to the positive sample in the loss function, and the magnitude corresponding to the negative sample in the loss function.

[0044] In one possible implementation, the hyperparameters corresponding to the network structure include at least one of the following:

[0045] The width and depth of the backbone network.

[0046] In one possible implementation,

[0047] The search module is used to: employ a search strategy by a reinforcement learning agent, and search for at least two candidate combinations for generating an image processing model based on a preset search space;

[0048] The training module is used to: perform a preset number of training rounds based on the at least two candidate combinations to obtain at least two candidate image processing models that correspond one-to-one with the at least two candidate combinations;

[0049] The apparatus further includes an update module, configured to obtain reward values ​​corresponding to the at least two candidate image processing models based on the verification image set, and update the search strategy according to the reward values.

[0050] In one possible implementation, the search module is used for:

[0051] By employing a search strategy through a reinforcement learning agent, a preset search space is searched based on a preset search order to obtain at least two candidate combinations for generating an image processing model. The preset search order refers to the order in which each sub-search space is searched in advance.

[0052] In one possible implementation, the update module is used to:

[0053] For any one of the at least two candidate image processing models, the accuracy and computational cost of the candidate image processing model are determined based on the validation image set.

[0054] The reward value corresponding to the candidate image processing model is determined based on the accuracy and computational cost of the candidate image processing model.

[0055] In one possible implementation, the number of candidate image processing models is at least three;

[0056] The determining module is used for:

[0057] Based on the reward value, at least two alternative image processing models are determined from the at least three candidate image processing models, wherein the number of alternative image processing models is less than the number of candidate image processing models;

[0058] The at least two candidate image processing models are trained separately until the at least two candidate image processing models meet the preset training conditions.

[0059] Based on the performance information of at least two candidate image processing models that have been trained, a target image processing model is determined from the at least two candidate image processing models that have been trained.

[0060] According to one aspect of this disclosure, an electronic device is provided, comprising: one or more processors; a memory for storing executable instructions; wherein the one or more processors are configured to invoke the executable instructions stored in the memory to perform the method described above.

[0061] According to one aspect of this disclosure, a computer-readable storage medium is provided that stores computer program instructions thereon, which, when executed by a processor, implement the above-described method.

[0062] According to one aspect of this disclosure, a computer program product is provided, including computer-readable code, or a non-volatile computer-readable storage medium carrying computer-readable code, wherein when the computer-readable code is run in an electronic device, a processor in the electronic device performs the above-described method.

[0063] In this embodiment, at least two candidate combinations for generating an image processing model are obtained by searching a preset search space. The preset search space includes at least two items from a first sub-search space, a second sub-search space, and a third sub-search space. The first sub-search space is used to search for hyperparameters corresponding to data cleaning, which are used to clean the training image set. The second sub-search space is used to search for hyperparameters corresponding to the loss function. The third sub-search space is used to search for hyperparameters corresponding to the network structure. Training is performed based on the at least two candidate combinations to obtain at least two candidate image processing models that correspond one-to-one with the at least two candidate combinations. Based on the at least two candidate image processing models, a target image processing model is determined. Thus, a neural network search method is used to jointly search for at least two items from the hyperparameters corresponding to data cleaning, the loss function, and the network structure, thereby generating a more robust image processing model for a specific image processing task.

[0064] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure.

[0065] Other features and aspects of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0066] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the specification, serve to illustrate the technical solutions of this disclosure.

[0067] Figure 1 A flowchart illustrating the method for generating an image processing model provided in an embodiment of this disclosure is shown.

[0068] Figure 2 A block diagram of an image processing model generation apparatus provided in an embodiment of this disclosure is shown.

[0069] Figure 3 A block diagram of an electronic device 1900 provided in an embodiment of this disclosure is shown. Detailed Implementation

[0070] Various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0071] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0072] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0073] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0074] In related technologies, when generating image processing models, the three stages of data cleaning, loss function design, and network structure design are considered in a fragmented manner, and the design of these three stages is usually done manually.

[0075] This disclosure provides a method, apparatus, electronic device, storage medium, and program product for generating image processing models. By searching a preset search space, at least two candidate combinations for generating image processing models are obtained. The preset search space includes at least two items from a first sub-search space, a second sub-search space, and a third sub-search space. The first sub-search space searches for hyperparameters corresponding to data cleaning, used for cleaning the training image set. The second sub-search space searches for hyperparameters corresponding to a loss function. The third sub-search space searches for hyperparameters corresponding to a network structure. Training is performed based on the at least two candidate combinations to obtain at least two candidate image processing models corresponding one-to-one with each combination. A target image processing model is determined based on these at least two candidate image processing models. Thus, a neural network search method is used to jointly search for at least two items from the hyperparameters corresponding to data cleaning, the loss function, and the network structure, thereby enabling the generation of more robust image processing models for specific image processing tasks.

[0076] The method for generating the image processing model provided in the embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0077] Figure 1 A flowchart illustrating a method for generating an image processing model according to an embodiment of this disclosure is provided. In one possible implementation, the entity executing the method for generating the image processing model can be an image processing model generation device. For example, the method can be executed by a terminal device, a server, or other electronic equipment. The terminal device can be a user equipment (UE), mobile device, user terminal, terminal, cellular phone, cordless phone, personal digital assistant (PDA), handheld device, computing device, in-vehicle device, or wearable device, etc. In some possible implementations, the method for generating the image processing model can be implemented by a processor calling computer-readable instructions stored in memory. Figure 1 As shown, the method for generating the image processing model includes steps S11 to S13.

[0078] In step S11, based on a preset search space, at least two candidate combinations for generating an image processing model are obtained; wherein, the preset search space includes at least two items from a first sub-search space, a second sub-search space, and a third sub-search space; the first sub-search space is used to search for hyperparameters corresponding to data cleaning, wherein the hyperparameters corresponding to data cleaning are used to clean the training image set; the second sub-search space is used to search for hyperparameters corresponding to the loss function; and the third sub-search space is used to search for hyperparameters corresponding to the network structure.

[0079] In step S12, training is performed based on the at least two candidate combinations to obtain at least two candidate image processing models that correspond one-to-one with the at least two candidate combinations.

[0080] In step S13, a target image processing model is determined based on the at least two candidate image processing models.

[0081] The image processing model in this embodiment can be used to perform any specified type of image processing task. For example, the image processing model in this embodiment can be used to perform image processing tasks such as image classification, object detection, object recognition (e.g., face recognition), semantic segmentation, and instance segmentation.

[0082] In one possible implementation, the preset search space includes a first sub-search space, a second sub-search space, and a third sub-search space. In this implementation, any candidate combination obtained based on the preset search space may include hyperparameters corresponding to data cleaning, hyperparameters corresponding to the loss function, and hyperparameters corresponding to the network structure.

[0083] In another possible implementation, the preset search space includes a first sub-search space and a second sub-search space. In this implementation, the preset search space may not include a third sub-search space. In this implementation, any candidate combination obtained based on the preset search space may include hyperparameters corresponding to data cleaning and hyperparameters corresponding to the loss function.

[0084] In another possible implementation, the preset search space includes a first sub-search space and a third sub-search space. In this implementation, the preset search space may not include a second sub-search space. In this implementation, any candidate combination obtained based on the preset search space may include hyperparameters corresponding to data cleaning and hyperparameters corresponding to the network structure.

[0085] In another possible implementation, the preset search space includes a second sub-search space and a third sub-search space. In this implementation, the preset search space may not include the first sub-search space. In this implementation, any candidate combination obtained based on the preset search space may include hyperparameters corresponding to the loss function and hyperparameters corresponding to the network structure.

[0086] In the following text, we will take the example of a preset search space including a first sub-search space, a second sub-search space, and a third sub-search space, and any candidate combination including hyperparameters corresponding to data cleaning, hyperparameters corresponding to the loss function, and hyperparameters corresponding to the network structure.

[0087] In this embodiment of the disclosure, the hyperparameters corresponding to data cleaning can represent the hyperparameters used for data cleaning of the training image set. In this embodiment of the disclosure, for any candidate combination among the at least two candidate combinations, if the candidate combination includes hyperparameters corresponding to data cleaning, the training image set can be data cleaned according to the hyperparameters corresponding to data cleaning in the candidate combination to obtain the data-cleaned training image set corresponding to the candidate combination, and a candidate image processing model corresponding to the candidate combination can be obtained based on the data-cleaned training image set corresponding to the candidate combination.

[0088] In one possible implementation, the hyperparameters corresponding to the data cleaning include at least one of the following: a confidence threshold for filtering training images within a class, and a similarity threshold for merging training images from different classes. That is, the first sub-search space can be used to search for at least one of the following: a confidence threshold for filtering training images within a class, and a similarity threshold for merging training images from different classes.

[0089] As an example of this implementation, the first sub-search space can be used to search for: a confidence threshold for filtering training images within a class, and a similarity threshold for merging training images from different classes.

[0090] As another example of this implementation, the first sub-search space can be used to search for a confidence threshold for filtering training images within a class. In this example, the hyperparameters corresponding to the data cleaning may not include a similarity threshold for merging training images from different categories.

[0091] As another example of this implementation, the first sub-search space can be used to search for a similarity threshold for merging training images from different categories. In this example, the hyperparameters corresponding to the data cleaning may not include a confidence threshold for filtering training images within a class.

[0092] In this implementation, for any of the at least two candidate combinations, the training image set can be cleaned according to the confidence threshold used to filter training images within a class and / or the similarity threshold used to merge training images of different classes in the candidate combination, so as to obtain the data-cleaned training image set corresponding to the candidate combination.

[0093] As an example of this implementation, for any training image in the training image set, the training image can be retained if its corresponding class confidence is greater than or equal to a confidence threshold, and can be filtered (i.e., discarded) if its corresponding class confidence is less than the confidence threshold. Here, the class confidence of the training image represents the confidence that the training image belongs to the class to which it belongs. In this example, training images with a class confidence less than the confidence threshold can be identified as noise within their respective classes. By adopting this example, intra-class noise can be reduced.

[0094] In one example, Equation 1 can be used to determine the class confidence corresponding to training image i.

[0095]

[0096] Where K represents the total number of categories in the training image set, p represents the category to which training image i belongs, n represents the category that is not to which training image i belongs among the K categories (i.e., category n is different from category p), and dist ip Dist represents the similarity (e.g., cosine similarity) between the features of training image i and the features of the class center of class p. in This represents the similarity (e.g., cosine similarity) between the features of training image i and the features of the class center of category n.

[0097] In one example, the confidence threshold could be τ. intra Indicates. If Less than τ intra Then, training image i can be filtered; if Greater than or equal to τ intra Then the training image i can be retained.

[0098] As an example of this implementation, for any two categories in the training image set, the two categories can be merged if the similarity of the features of the class centers of the two categories is greater than a similarity threshold. This example reduces inter-class noise. In one example, the similarity threshold could be τ. inter express.

[0099] In the above implementation, by searching for at least one of the confidence threshold for filtering training images within a class and the similarity threshold for merging training images of different classes, a search-based data cleaning method is adopted, which can effectively alleviate the problem in related technologies of difficulty in determining whether training images with edges belong to complex positive samples or noise samples.

[0100] Of course, in other possible implementations, the types of hyperparameters corresponding to data cleaning in the first sub-search space can be flexibly designed for different types of image processing tasks or different application scenarios, and no restrictions are imposed here.

[0101] In this embodiment of the disclosure, the hyperparameters corresponding to the loss function can represent the hyperparameters used to determine the formula for the loss function. In this embodiment of the disclosure, for any of the at least two candidate combinations, if the candidate combination includes the hyperparameters corresponding to the loss function, the loss function corresponding to the candidate combination can be determined based on the hyperparameters corresponding to the loss function in the candidate combination, and the parameters of the candidate image processing model corresponding to the candidate combination can be updated based on the loss function corresponding to the candidate combination. If the candidate combination does not include the hyperparameters corresponding to the loss function, a preset loss function can be used as the loss function corresponding to the candidate combination.

[0102] In one possible implementation, the hyperparameters corresponding to the loss function include at least one of the following: the margin in the loss function, the magnitude of positive samples in the loss function, and the magnitude of negative samples in the loss function. That is, the second sub-search space can be used to search for at least one of the following: the margin in the loss function, the magnitude of positive samples in the loss function, and the magnitude of negative samples in the loss function. The margin in the loss function can be used to control inter-class distance and intra-class distance. In some application scenarios, the margin in the loss function can also be called the margin term, the number of margins, etc., and is not limited here. The magnitude of positive samples in the loss function can be used to control the importance of positive samples to the loss function; the magnitude of negative samples in the loss function can be used to control the importance of negative samples to the loss function.

[0103] As an example of this implementation, the second sub-search space can be used to search for: the interval in the loss function, the magnitude corresponding to the positive sample in the loss function, and the magnitude corresponding to the negative sample in the loss function.

[0104] As another example of this implementation, the second sub-search space can be used to search for: the interval in the loss function and the magnitude corresponding to the positive sample in the loss function.

[0105] As another example of this implementation, the second sub-search space can be used to search for: the interval in the loss function and the magnitude corresponding to the negative sample in the loss function.

[0106] As another example of this implementation, the second sub-search space can be used to search for the interval in the loss function.

[0107] As another example of this implementation, the second sub-search space can be used to search for the magnitude of positive samples in the loss function.

[0108] As another example of this implementation, the second sub-search space can be used to search for the magnitude of the negative samples in the loss function.

[0109] In this implementation, for any one of the at least two candidate combinations, the loss function corresponding to the candidate combination can be determined based on at least one of the interval in the loss function of the candidate combination, the magnitude corresponding to the positive sample in the loss function, and the magnitude corresponding to the negative sample in the loss function.

[0110] In one example, the loss function can be expressed as Equation 2:

[0111]

[0112] Where K represents the total number of categories in the training image set, y and k represent two distinct categories, and w y w represents the weight of category y. k w represents the weight of category k. y w k The features of the class centers of category y and category k are vectors of the same dimension. Indicates w y The angle between the feature and the class center of category y, Indicates w k The angle between the feature and the class center of category k.

[0113] Assumption Indicates w k The cosine similarity between the feature x and the class center of category k can then determine the interval function as shown in Equation 3:

[0114]

[0115] Transforming Equation 3, we obtain:

[0116]

[0117] Where m1, m2 and m3 represent intervals.

[0118] Additionally, in Equation 2, s p s represents the amplitude corresponding to the positive sample. n This indicates the magnitude corresponding to the negative sample.

[0119] In this implementation, the accuracy of the trained image processing model can be improved by searching for at least one of the interval, the magnitude corresponding to the positive sample, and the magnitude corresponding to the negative sample in the loss function, and then training the image processing model based on the loss function obtained by the search.

[0120] Of course, in other possible implementations, the type of hyperparameters corresponding to the loss function in the second sub-search space can be flexibly designed for different types of image processing tasks or different application scenarios, and no restrictions are imposed here.

[0121] In this embodiment, the hyperparameters corresponding to the network structure can represent the hyperparameters used to determine the network structure of the image processing model. In this embodiment, for any of the at least two candidate combinations, if the candidate combination includes hyperparameters corresponding to the network structure, the network structure of the candidate image processing model corresponding to the candidate combination can be determined based on the hyperparameters corresponding to the network structure in the candidate combination. If the candidate combination does not include hyperparameters corresponding to the network structure, a preset network structure can be used as the network structure of the candidate image processing model corresponding to the candidate combination.

[0122] In one possible implementation, the hyperparameters corresponding to the network structure include at least one of the following: the width of the backbone network and the depth of the backbone network. That is, the third sub-search space can be used to search for at least one of the following: the width of the backbone network and the depth of the backbone network.

[0123] As an example of this implementation, a third sub-search space can be used to search for: the width and depth of the backbone network.

[0124] As another example of this implementation, a third sub-search space can be used to search for the width of the backbone network.

[0125] As another example of this implementation, a third sub-search space can be used to search for the depth of the backbone network.

[0126] In this implementation, for any one of the at least two candidate combinations, the network structure corresponding to the candidate combination can be determined based on the width and / or depth of the backbone network in the candidate combination.

[0127] In this implementation, the performance of the image processing model can be improved by searching for at least one of the width and depth of the backbone network and training the image processing model based on the network structure obtained by the search.

[0128] Of course, in other possible implementations, the types of hyperparameters corresponding to the network structure in the third sub-search space can be flexibly designed for different types of image processing tasks or different application scenarios, and no restrictions are imposed here.

[0129] In one possible implementation, the step of searching for at least two candidate combinations for generating an image processing model based on a preset search space includes: using a reinforcement learning agent to employ a search strategy to search for at least two candidate combinations for generating an image processing model based on a preset search space; the step of training based on the at least two candidate combinations to obtain at least two candidate image processing models corresponding one-to-one with the at least two candidate combinations includes: performing a preset number of training rounds based on the at least two candidate combinations to obtain at least two candidate image processing models corresponding one-to-one with the at least two candidate combinations; the method further includes: obtaining reward values ​​corresponding to the at least two candidate image processing models based on a validation image set; and updating the search strategy based on the reward values.

[0130] In this implementation, the reinforcement learning agent can represent a reinforcement learning-based agent. The search strategy can represent the strategy by which the reinforcement learning agent searches the preset search space. That is, the reinforcement learning agent uses the search strategy when searching the preset search space.

[0131] In this implementation, for any one of the at least two candidate combinations, a preset number of training iterations can be performed to obtain the candidate image processing model corresponding to the candidate combination. The preset number of iterations can be a preset epoch or a preset iteration, and is not limited here. For example, for any one of the at least two candidate combinations, one training epoch can be performed to obtain the candidate image processing model corresponding to the candidate combination.

[0132] As an example of this implementation, Proximal Policy Optimization (PPO) can be used to maximize the expected reward value to update the parameters of the reinforcement learning agent, thereby updating the agent's search policy.

[0133] In this implementation, the parameters of the reinforcement learning agent can be continuously updated until the parameters of the reinforcement learning agent converge.

[0134] In this implementation, the step of using a reinforcement learning agent to employ a search strategy and search for at least two candidate combinations for generating an image processing model based on a preset search space includes: using an initial search strategy and a preset search space to search for candidate combinations for generating an image processing model; and responding to an update of the search strategy, using an updated search strategy and a preset search space to search for candidate combinations for generating an image processing model.

[0135] In one example, the number of training epochs for the reinforcement learning agent can be denoted as T. At any given epoch, the reinforcement learning agent can employ a state-of-the-art search strategy to find B candidate combinations, where B is an integer greater than or equal to 1. For example, the B candidate combinations can be denoted as C1,...C1...C2...C3... B .

[0136] In this implementation, a reinforcement learning agent employs a search strategy to search for at least two candidate combinations for generating an image processing model based on a preset search space. The agent then performs a preset number of training rounds based on each of these at least two candidate combinations to obtain at least two candidate image processing models that correspond one-to-one with each combination. Based on a validation image set, reward values ​​corresponding to the at least two candidate image processing models are obtained, and the search strategy is updated according to these reward values. Thus, with training, the reinforcement learning agent learns a more robust search strategy, enabling it to search for higher-quality candidate combinations and generate higher-quality image processing models.

[0137] As an example of this implementation, obtaining the reward value corresponding to the at least two candidate image processing models based on the verification image set includes: for any candidate image processing model among the at least two candidate image processing models, determining the accuracy and computational cost of the candidate image processing model based on the verification image set; and determining the reward value corresponding to the candidate image processing model based on the accuracy and computational cost. The accuracy of the candidate image processing model can be determined by processing the verification image set using the candidate image processing model. In this example, the reward value corresponding to the candidate image processing model is positively correlated with the accuracy of the candidate image processing model, and negatively correlated with the computational cost of the candidate image processing model. That is, the higher the accuracy of the candidate image processing model, the larger the reward value; the lower the accuracy of the candidate image processing model, the smaller the reward value. Similarly, the larger the computational cost of the candidate image processing model, the smaller the reward value; and the smaller the computational cost, the larger the reward value.

[0138] In one example, Equation 5 can be used to determine the reward value R(C) corresponding to the candidate image processing model:

[0139]

[0140] Where C represents the candidate combination corresponding to the candidate image processing model, ACC(C) represents the accuracy of the candidate image processing model, COST(C) represents the computational cost of the candidate image processing model, TAR represents the target computational cost, and α represents the weight factor. The computational cost can be measured using metrics such as FLOPs (Floating Point Operations Per second), and is not limited here.

[0141] In this example, for any one of the at least two candidate image processing models, the accuracy and computational cost of the candidate image processing model are determined based on the validation image set, and the reward value of the candidate image processing model is determined based on the accuracy and computational cost of the candidate image processing model. This allows for a balance between accuracy and computational cost when updating the search strategy.

[0142] As another example of this implementation, for any one of the at least two candidate image processing models, the reward value corresponding to the candidate image processing model can be determined based on the accuracy of the candidate image processing model. In this example, the computational cost of the candidate image processing model can be disregarded when determining the reward value. In this example, the reward value of the candidate image processing model is positively correlated with the accuracy of the candidate image processing model.

[0143] As an example of this implementation, the reinforcement learning agent employs a recurrent neural network (RNN). In this example, by using a recurrent neural network as the reinforcement learning agent to search for candidate combinations, a more cognitive and efficient joint search can be achieved, thereby improving the quality of the candidate combinations obtained.

[0144] As an example of this implementation, each hyperparameter in the preset search space can be tokenized, and the process of searching the preset search space can be transformed into a token sequence prediction process, thereby improving the speed and flexibility of joint search.

[0145] As an example of this implementation, the step of using a reinforcement learning agent to adopt a search strategy and search for at least two candidate combinations for generating an image processing model based on a preset search space includes: using a reinforcement learning agent to adopt a search strategy and search a preset search space based on a preset search order to obtain at least two candidate combinations for generating an image processing model, wherein the preset search order represents a pre-set order of searching each sub-search space.

[0146] In one example, the preset search order can be the first sub-search space - the second sub-search space - the third sub-search space. That is, the hyperparameters corresponding to data cleaning, the hyperparameters corresponding to the loss function, and the hyperparameters corresponding to the network structure can be searched in the order of the first sub-search space - the second sub-search space - the third sub-search space to obtain candidate combinations.

[0147] Of course, those skilled in the art can flexibly set the preset search order according to the actual application scenario requirements, and no limitation is made here.

[0148] In this example, a reinforcement learning agent employs a search strategy to search a preset search space based on a preset search order, thereby obtaining at least two candidate combinations for generating an image processing model, which improves the speed of joint search.

[0149] In another possible implementation, a reinforcement learning agent may not be used when searching for candidate combinations based on a predefined search space. For example, all candidate combinations can be searched based on a predefined search space, and then each candidate combination can be trained separately (e.g., trained until the parameters converge). Finally, the candidate image processing model with the best performance among all candidate image processing models corresponding to all candidate combinations is determined as the target image processing model.

[0150] In one possible implementation, the number of candidate image processing models is at least three; determining the target image processing model based on the at least two candidate image processing models includes: determining at least two alternative image processing models from the at least three candidate image processing models according to the reward value, wherein the number of alternative image processing models is less than the number of candidate image processing models; training the at least two alternative image processing models respectively until the at least two alternative image processing models meet preset training conditions; and determining the target image processing model from the at least two alternative image processing models based on the performance information of the trained at least two alternative image processing models.

[0151] In this implementation, the N candidate image processing models with the highest reward values ​​from all candidate image processing models are selected as alternative image processing models, where N is an integer greater than or equal to 2. In this implementation, the alternative image processing models are a subset of the candidate image processing models.

[0152] In this implementation, for any one of the at least two candidate image processing models, the candidate image processing model can be trained until a preset training condition is met. As an example of this implementation, the preset training condition can be parameter convergence. As another example of this implementation, the preset training condition can be reaching a preset number of epochs.

[0153] In this implementation, after the at least two candidate image processing models have been trained, the performance information of the at least two candidate image processing models can be determined, so that the candidate image processing model with the best performance can be determined as the target image processing model.

[0154] As an example of this implementation, the performance information of any one of the at least two candidate image processing models may include: the accuracy of the candidate image processing model and the computational cost of the candidate image processing model.

[0155] As an example of this implementation, the performance information of any one of the at least two candidate image processing models may include the accuracy of the candidate image processing model.

[0156] In this implementation, at least two candidate image processing models are determined from the at least three candidate image processing models based on the reward value. The at least two candidate image processing models are trained separately until they meet the preset training conditions. Based on the performance information of the at least two trained candidate image processing models, the target image processing model is determined from the at least two trained candidate image processing models. This reduces the number of fully trained image processing models and thus improves the speed of generating the target image processing model.

[0157] In one possible implementation, after determining the target image processing model, the method further includes: inputting the image to be processed into the target image processing model, performing image processing on the image to be processed through the target image processing model, and obtaining the image processing result corresponding to the image to be processed.

[0158] As an example of this implementation, the target image processing model is used for face recognition; the step of inputting the image to be processed into the target image processing model and performing image processing on the image to be processed through the target image processing model to obtain the image processing result corresponding to the image to be processed includes: inputting the image to be processed into the target image processing model and performing face recognition on the image to be processed through the target image processing model to obtain the face recognition result corresponding to the image to be processed.

[0159] The following describes the image processing model generation method provided in this disclosure through a specific application scenario. In this application scenario, the image processing model is used for face recognition.

[0160] In this application scenario, the preset search space includes a first sub-search space, a second sub-search space, and a third sub-search space; wherein, the first sub-search space is used to search for: the confidence threshold τ for filtering training images within a class. intra The similarity threshold τ used to merge training images from different categories inter The second sub-search space is used to search for: the intervals m1, m2, and m3 in the loss function, and the magnitude s corresponding to the positive samples in the loss function. p The magnitude s of the negative samples in the loss function n The third sub-search space is used to search for: the width of the backbone network. Backbone network depth

[0161] In this application scenario, a reinforcement learning agent can be used to search for candidate combinations. The reinforcement learning agent can employ a recurrent neural network. The agent can search a predefined search space based on the order of a first sub-search space – a second sub-search space – a third sub-search space. Furthermore, each candidate combination obtained by the reinforcement learning agent includes τ. intra τ inter m1, m2, m3, s p and s n .

[0162] A reinforcement learning agent can employ an initial search strategy to search a predefined search space, obtaining B candidate combinations, denoted as C1,...C1. B For these B candidate combinations, after training for one epoch, the reward value can be calculated using Equation 5 above. A proximal policy optimization approach can be used to maximize the expected reward value, thereby updating the parameters of the reinforcement learning agent and thus updating its search policy. The reinforcement learning agent can search for B candidate combinations after each search policy update until the agent's parameters converge.

[0163] From all candidate image processing models, the N candidate image processing models with the highest reward values ​​can be selected as alternative image processing models. For each alternative image processing model, the cleaned training image set corresponding to the alternative image processing model and the loss function corresponding to the alternative image processing model can be used to train the alternative image processing model until the parameters of the alternative image processing model converge.

[0164] After the parameters of each candidate image processing model converge, the accuracy of each candidate model can be determined based on the validation image set. Then, based on the accuracy and computational cost of each candidate model, the best-performing candidate image processing model can be selected as the target image processing model.

[0165] By adopting this application scenario, there is no need to manually design data cleaning strategies, loss functions, and network structures, nor is it necessary to manually tune the hyperparameters corresponding to data cleaning, loss functions, and network structures. It can search for the optimal combination of data cleaning strategies, loss functions, and network structures for the face recognition system, and the image processing model obtained from the search can improve the accuracy and robustness of face recognition.

[0166] The image processing model generation method provided in this disclosure can be customized for specific scenarios, such as specific mobile phone chips. This customized design process is fully automated, requiring no manual parameter tuning, thereby improving the speed and accuracy of image processing model interaction.

[0167] It is understood that the various method embodiments mentioned above in this disclosure can be combined with each other to form combined embodiments without violating the principle and logic. Due to space limitations, this disclosure will not elaborate further. Those skilled in the art will understand that in the above methods of specific implementation, the specific execution order of each step should be determined by its function and possible internal logic.

[0168] In addition, this disclosure also provides an image processing model generation apparatus, electronic device, computer-readable storage medium, and computer program product, all of which can be used to implement any of the image processing model generation methods provided in this disclosure. The corresponding technical solutions and effects can be found in the relevant descriptions in the method section, and will not be repeated here.

[0169] Figure 2 A block diagram of an apparatus for generating an image processing model according to an embodiment of this disclosure is shown. Figure 2 As shown, the image processing model generation device includes:

[0170] Search module 21 is used to search for at least two candidate combinations for generating an image processing model based on a preset search space; wherein the preset search space includes at least two items from a first sub-search space, a second sub-search space, and a third sub-search space; the first sub-search space is used to search for hyperparameters corresponding to data cleaning, wherein the hyperparameters corresponding to data cleaning are used to clean the training image set; the second sub-search space is used to search for hyperparameters corresponding to the loss function; and the third sub-search space is used to search for hyperparameters corresponding to the network structure.

[0171] Training module 22 is used to train based on the at least two candidate combinations respectively to obtain at least two candidate image processing models that correspond one-to-one with the at least two candidate combinations;

[0172] The determination module 23 is used to determine the target image processing model based on the at least two candidate image processing models.

[0173] In one possible implementation, the hyperparameters corresponding to the data cleaning include at least one of the following:

[0174] Confidence thresholds for filtering training images within a class, and similarity thresholds for merging training images from different classes.

[0175] In one possible implementation, the hyperparameters corresponding to the loss function include at least one of the following:

[0176] The interval in the loss function, the magnitude corresponding to the positive sample in the loss function, and the magnitude corresponding to the negative sample in the loss function.

[0177] In one possible implementation, the hyperparameters corresponding to the network structure include at least one of the following:

[0178] The width and depth of the backbone network.

[0179] In one possible implementation,

[0180] The search module 21 is used to: employ a search strategy by a reinforcement learning agent, and search for at least two candidate combinations for generating an image processing model based on a preset search space;

[0181] The training module 22 is used to: perform a preset number of training rounds based on the at least two candidate combinations to obtain at least two candidate image processing models that correspond one-to-one with the at least two candidate combinations;

[0182] The apparatus further includes an update module, configured to obtain reward values ​​corresponding to the at least two candidate image processing models based on the verification image set, and update the search strategy according to the reward values.

[0183] In one possible implementation, the search module 21 is used for:

[0184] By employing a search strategy through a reinforcement learning agent, a preset search space is searched based on a preset search order to obtain at least two candidate combinations for generating an image processing model. The preset search order refers to the order in which each sub-search space is searched in advance.

[0185] In one possible implementation, the update module is used to:

[0186] For any one of the at least two candidate image processing models, the accuracy and computational cost of the candidate image processing model are determined based on the validation image set.

[0187] The reward value corresponding to the candidate image processing model is determined based on the accuracy and computational cost of the candidate image processing model.

[0188] In one possible implementation, the number of candidate image processing models is at least three;

[0189] The determining module 23 is used for:

[0190] Based on the reward value, at least two alternative image processing models are determined from the at least three candidate image processing models, wherein the number of alternative image processing models is less than the number of candidate image processing models;

[0191] The at least two candidate image processing models are trained separately until the at least two candidate image processing models meet the preset training conditions.

[0192] Based on the performance information of at least two candidate image processing models that have been trained, a target image processing model is determined from the at least two candidate image processing models that have been trained.

[0193] In this embodiment, at least two candidate combinations for generating an image processing model are obtained by searching a preset search space. The preset search space includes at least two items from a first sub-search space, a second sub-search space, and a third sub-search space. The first sub-search space is used to search for hyperparameters corresponding to data cleaning, which are used to clean the training image set. The second sub-search space is used to search for hyperparameters corresponding to the loss function. The third sub-search space is used to search for hyperparameters corresponding to the network structure. Training is performed based on the at least two candidate combinations to obtain at least two candidate image processing models that correspond one-to-one with the at least two candidate combinations. Based on the at least two candidate image processing models, a target image processing model is determined. Thus, a neural network search method is used to jointly search for at least two items from the hyperparameters corresponding to data cleaning, the loss function, and the network structure, thereby generating a more robust image processing model for a specific image processing task.

[0194] In some embodiments, the functions or modules of the apparatus provided in this disclosure can be used to perform the methods described in the above method embodiments. The specific implementation and technical effects can be referred to the description of the above method embodiments. For the sake of brevity, they will not be repeated here.

[0195] This disclosure also provides a computer-readable storage medium storing computer program instructions thereon, which, when executed by a processor, implement the above-described method. The computer-readable storage medium may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium.

[0196] This disclosure also proposes a computer program including computer-readable code, wherein when the computer-readable code is run in an electronic device, a processor in the electronic device executes the above-described method.

[0197] This disclosure also provides a computer program product, including computer-readable code, or a non-volatile computer-readable storage medium carrying computer-readable code, wherein when the computer-readable code is run in an electronic device, the processor in the electronic device executes the above-described method.

[0198] This disclosure also provides an electronic device, including: one or more processors; a memory for storing executable instructions; wherein the one or more processors are configured to invoke the executable instructions stored in the memory to perform the above-described method.

[0199] Electronic devices can be provided as terminals, servers, or other forms of devices.

[0200] Figure 3 A block diagram of an electronic device 1900 provided in an embodiment of this disclosure is shown. For example, the electronic device 1900 may be provided as a server. (Refer to...) Figure 3 The electronic device 1900 includes a processing component 1922, which further includes one or more processors, and memory resources represented by memory 1932 for storing instructions, such as application programs, that can be executed by the processing component 1922. The application programs stored in memory 1932 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 1922 is configured to execute instructions to perform the methods described above.

[0201] Electronic device 1900 may also include a power supply component 1926 configured to perform power management of electronic device 1900, a wired or wireless network interface 1950 configured to connect electronic device 1900 to a network, and an input / output (I / O) interface 1958. Electronic device 1900 can operate on an operating system stored in memory 1932, such as Microsoft Server operating system (Windows Server). TM Apple's graphical user interface-based operating system (Mac OSX) TM ), a multi-user, multi-process computer operating system (Unix) TM Linux is a free and open-source Unix-like operating system. TM ), the open-source Unix-like operating system (FreeBSD) TM (or similar.)

[0202] In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, such as a memory 1932 including computer program instructions that can be executed by a processing component 1922 of an electronic device 1900 to perform the above-described method.

[0203] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0204] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0205] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0206] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0207] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0208] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0209] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0210] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0211] The computer program product can be implemented specifically through hardware, software, or a combination thereof. In one alternative embodiment, the computer program product is specifically embodied in a computer storage medium; in another alternative embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0212] The description of the various embodiments above tends to emphasize the differences between the various embodiments. The similarities or similarities between them can be referred to, and for the sake of brevity, they will not be repeated here.

[0213] If the technical solution of this disclosure involves personal information, the product applying the technical solution of this disclosure has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution of this disclosure involves sensitive personal information, the product applying the technical solution of this disclosure has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of "express consent". For example, at personal information collection devices such as cameras, clear and prominent signs are set up to indicate that the user has entered the scope of personal information collection and that personal information will be collected. If the user voluntarily enters the collection scope, it is deemed to have consented to the collection of their personal information; or on the personal information processing device, with clear signs / information informing the user of the personal information processing rules, authorization is obtained from the user through pop-up information or by asking the user to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

[0214] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for generating an image processing model, characterized in that, include: Based on a preset search space, at least two candidate combinations for generating an image processing model are obtained; wherein the preset search space includes at least two items from a first sub-search space, a second sub-search space, and a third sub-search space; the first sub-search space is used to search for hyperparameters corresponding to data cleaning, wherein the hyperparameters corresponding to data cleaning are used to clean the training image set; the second sub-search space is used to search for hyperparameters corresponding to the loss function; and the third sub-search space is used to search for hyperparameters corresponding to the network structure. Training is performed based on the at least two candidate combinations to obtain at least two candidate image processing models that correspond one-to-one with the at least two candidate combinations. Based on the at least two candidate image processing models, the target image processing model is determined.

2. The method according to claim 1, characterized in that, The hyperparameters corresponding to the data cleaning include at least one of the following: Confidence thresholds for filtering training images within a class, and similarity thresholds for merging training images from different classes.

3. The method according to claim 1 or 2, characterized in that, The hyperparameters corresponding to the loss function include at least one of the following: The interval in the loss function, the magnitude corresponding to the positive sample in the loss function, and the magnitude corresponding to the negative sample in the loss function.

4. The method according to any one of claims 1 to 2, characterized in that, The hyperparameters corresponding to the network structure include at least one of the following: The width and depth of the backbone network.

5. The method according to any one of claims 1 to 2, characterized in that, The process of searching for at least two candidate combinations for generating an image processing model based on a preset search space includes: using a search strategy through a reinforcement learning agent to search for at least two candidate combinations for generating an image processing model based on a preset search space. The step of training based on the at least two candidate combinations to obtain at least two candidate image processing models corresponding to the at least two candidate combinations includes: training based on the at least two candidate combinations for a preset number of rounds to obtain at least two candidate image processing models corresponding to the at least two candidate combinations. The method further includes: obtaining reward values ​​corresponding to the at least two candidate image processing models based on the verification image set; and updating the search strategy according to the reward values.

6. The method according to claim 5, characterized in that, The reinforcement learning agent employs a search strategy based on a preset search space to obtain at least two candidate combinations for generating the image processing model, including: By employing a search strategy through a reinforcement learning agent, a preset search space is searched based on a preset search order to obtain at least two candidate combinations for generating an image processing model. The preset search order refers to the order in which each sub-search space is searched in advance.

7. The method according to claim 5, characterized in that, The step of obtaining the reward values ​​corresponding to the at least two candidate image processing models based on the verification image set includes: For any one of the at least two candidate image processing models, the accuracy and computational cost of the candidate image processing model are determined based on the validation image set. The reward value corresponding to the candidate image processing model is determined based on the accuracy and computational cost of the candidate image processing model.

8. The method according to any one of claims 1 to 2, characterized in that, The number of candidate image processing models must be at least three; The step of determining the target image processing model based on the at least two candidate image processing models includes: Based on the reward value, at least two alternative image processing models are determined from the at least three candidate image processing models, wherein the number of alternative image processing models is less than the number of candidate image processing models; The at least two candidate image processing models are trained separately until the at least two candidate image processing models meet the preset training conditions. Based on the performance information of at least two candidate image processing models that have been trained, a target image processing model is determined from the at least two candidate image processing models that have been trained.

9. An apparatus for generating an image processing model, characterized in that, include: The search module is used to search for at least two candidate combinations for generating an image processing model based on a preset search space; wherein the preset search space includes at least two items from a first sub-search space, a second sub-search space, and a third sub-search space; the first sub-search space is used to search for hyperparameters corresponding to data cleaning, wherein the hyperparameters corresponding to data cleaning are used to clean the training image set; the second sub-search space is used to search for hyperparameters corresponding to the loss function; and the third sub-search space is used to search for hyperparameters corresponding to the network structure. The training module is used to train on the at least two candidate combinations respectively to obtain at least two candidate image processing models that correspond one-to-one with the at least two candidate combinations. The determination module is used to determine the target image processing model based on the at least two candidate image processing models.

10. An electronic device, characterized in that, include: One or more processors; Memory used to store executable instructions; The one or more processors are configured to invoke executable instructions stored in the memory to perform the method according to any one of claims 1 to 8.

11. A computer-readable storage medium storing computer program instructions thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 8.

12. A computer program product, characterized in that, Includes computer-readable code, or a non-volatile computer-readable storage medium carrying computer-readable code, wherein when the computer-readable code is executed in an electronic device, the processor in the electronic device performs the method according to any one of claims 1 to 8.

Images