Image processing method and device, electronic equipment and machine readable storage medium

Abstract

The application provides an image processing method and device, electronic equipment and machine readable storage medium, the method comprises the following steps: according to initial training set, using preset training system, model training is carried out, and initial model is obtained;The output data obtained by using the initial model for image processing on the business application end is filtered, and the target output data meeting the preset condition in confidence and accuracy is obtained;According to the initial training set is expanded according to the target output data, and the expanded training set is obtained;According to the initial model is optimized training according to the expanded training set, and the evolution model is obtained, so that the business application end carries out image processing according to the evolution model.This method can realize model automatic evolution, and the processing effect is optimized by carrying out image processing according to the evolution model.

Description

Technical Field

[0001] This application relates to the field of computer vision technology, and in particular to an image processing method, apparatus, electronic device, and machine-readable storage medium. Background Technology

[0002] With the rapid rise of applications such as intelligent video surveillance, autonomous driving, and face detection, the field of computer vision has received widespread attention due to huge market demand. Machine learning / deep learning methods have ensured the efficiency and accuracy of object detection. Furthermore, in recent years, the rapid improvement in the computing power of hardware devices, the massive amount of data accumulated rapidly in the internet era, and the improvement of training methods have greatly promoted the development of machine learning and deep learning, leading to groundbreaking progress in the field of computer vision.

[0003] However, in real-world applications, the amount of data that typically covers all types of target objects is enormous, making data cleaning and labeling very costly. Summary of the Invention

[0004] In view of the above, this application provides an image processing method, apparatus, electronic device, and machine-readable storage medium.

[0005] According to a first aspect of the embodiments of this application, an image processing method is provided, comprising:

[0006] Based on the initial training set, the model is trained using a pre-set training system to obtain the initial model;

[0007] The output data obtained by image processing using the initial model in the business application is filtered to obtain target output data whose confidence and accuracy meet preset conditions.

[0008] The initial training set is expanded based on the target output data to obtain an expanded training set;

[0009] Based on the expanded training set, the initial model is optimized and trained to obtain an evolutionary model, so that the business application can perform image processing based on the evolutionary model.

[0010] According to a second aspect of the embodiments of this application, an image processing apparatus is provided, comprising:

[0011] The pre-training unit is used to train the model based on the initial training set using a preset training system to obtain the initial model;

[0012] The filtering unit is used to filter the output data obtained by the business application terminal using the initial model for image processing, and to obtain target output data whose confidence and accuracy meet preset conditions.

[0013] An expansion unit is used to expand the initial training set based on the target output data to obtain an expanded training set.

[0014] An optimization training unit is used to optimize and train the initial model based on the expanded training set to obtain an evolutionary model, so that the business application terminal can perform image processing based on the evolutionary model.

[0015] According to a third aspect of the present application, an electronic device is provided, including a processor and a machine-readable storage medium, the machine-readable storage medium storing machine-executable instructions executable by the processor, the processor being configured to execute the machine-executable instructions to implement the method provided in the first aspect.

[0016] According to a fourth aspect of the embodiments of this application, a machine-readable storage medium is provided, wherein a computer program is stored therein, and when the computer program is executed by a processor, it implements the method provided in the first aspect.

[0017] According to a fifth aspect of the embodiments of this application, a computer program is provided, which is stored in a machine-readable storage medium, and when a processor executes the computer program, causes the processor to perform the method provided in the first aspect.

[0018] The image processing method of this application embodiment trains a model using a preset training system based on an initial training set to obtain an initial model. Then, based on the confidence and accuracy of the output data, it filters the output data obtained by the business application using the initial model to obtain target output data whose confidence and accuracy meet preset conditions. The initial training set is then expanded based on the target output data, achieving automatic expansion of the training set and reducing the requirement for the number of samples in the initial training set. Furthermore, the initial model is optimized and trained using the expanded training set to obtain an evolutionary model, achieving automatic model evolution. Finally, image processing is performed based on the evolutionary model to optimize the processing effect. Attached Figure Description

[0019] Figure 1 This is a schematic flowchart of an image processing method provided in an embodiment of this application;

[0020] Figure 2 This is a schematic diagram illustrating a specific implementation process of model optimization training provided in an embodiment of this application;

[0021] Figure 3 This is a schematic diagram of the structure of an image processing device provided in an embodiment of this application;

[0022] Figure 4 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0023] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0024] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0025] To enable those skilled in the art to better understand the technical solutions provided in the embodiments of this application, and to make the above-mentioned objectives, features and advantages of the embodiments of this application more apparent and understandable, the technical solutions in the embodiments of this application will be further described in detail below with reference to the accompanying drawings.

[0026] Please see Figure 1 This is a flowchart illustrating an image processing method provided in an embodiment of this application, as shown below. Figure 1 As shown, the image processing method may include:

[0027] Step S100: Based on the initial training set, use a preset training system to train the model and obtain the initial model.

[0028] For example, the number of training samples in the initial training set can be relatively small, such as hundreds or even dozens.

[0029] For example, training samples in the initial training set can be labeled manually or by using an auxiliary labeling system.

[0030] For example, the preset training system can be any machine learning training system, which may include, but is not limited to, training frameworks such as Tensorflow, PyTorch, and Caffe.

[0031] Step S110: Filter the output data obtained by image processing using the initial model on the business application side to obtain target output data whose confidence and accuracy meet the preset conditions.

[0032] For example, the business application may include an application system that loads the initial model trained in step S100. This can be any machine learning inference system, including but not limited to inference servers, mobile phones, embedded devices, etc.

[0033] In this embodiment of the application, the business application can load the initial model trained in step S100 and use the initial model for image processing.

[0034] For example, the output data obtained by image processing using the initial model in the business application can be filtered using certain rules to select data that is beneficial to improving the model's generalization ability.

[0035] Generalization ability refers to the ability of a machine learning algorithm to adapt to new samples, that is, to output a reasonable result after training on an existing dataset with new data. The goal of machine learning is to learn the patterns hidden behind the data. The trained network can also give appropriate outputs on data outside the training set that have the same patterns. This ability is called generalization ability.

[0036] For example, the output data can be filtered based on its confidence level and accuracy to obtain output data (referred to as target output data in this paper) whose confidence level and accuracy meet preset conditions.

[0037] Step S120: Expand the initial training set according to the target output data to obtain the expanded training set.

[0038] In this embodiment of the application, when the target output data is obtained through the method described in the above steps, the initial training set can be expanded based on the target output data to obtain an expanded training set. Thus, the expansion of the training set can be automatically achieved when the number of training samples in the initial training set is small. Then, the model can be optimized and trained based on the expanded training set to achieve model evolution.

[0039] Step S130: Based on the expanded training set, optimize and train the initial model to obtain an evolutionary model, so that the business application can perform image processing based on the evolutionary model.

[0040] In this embodiment, the initial model can be optimized and trained based on the expanded training set to obtain an evolutionary model. The evolutionary model is then loaded into the business application, which performs image processing based on the evolutionary model to optimize the image processing effect.

[0041] It can be seen that, in Figure 1In the illustrated method, an initial model is obtained by training the model using a preset training system based on the initial training set. Then, based on the confidence and accuracy of the output data, the output data obtained by the business application using the initial model for image processing is filtered to obtain target output data whose confidence and accuracy meet preset conditions. The initial training set is then expanded based on the target output data, realizing automatic expansion of the training set and reducing the requirement for the number of samples in the initial training set. Subsequently, the initial model is optimized and trained based on the expanded training set to obtain an evolutionary model, realizing automatic model evolution. Finally, image processing is performed based on the evolutionary model to optimize the processing effect.

[0042] In some embodiments, step S110, filtering the output data obtained by the business application terminal using the initial model for image processing, may include:

[0043] Based on the confidence level of the output data for the target object, the output data is divided into at least two groups; the at least two groups include at least a first group and a second group with corresponding confidence levels decreasing sequentially.

[0044] The output data is filtered according to the grouping to obtain the target output data; the data filtering strategies for different groups are not exactly the same.

[0045] For example, considering that the proportion of erroneous data in output data with low confidence may be relatively high, the algorithm may not be able to automatically remove erroneous data, that is, it cannot automatically filter the data; while output data with high confidence has a poor effect on model improvement and will reduce training efficiency; while output data with neither too high nor too low confidence can be automatically filtered by the algorithm.

[0046] In other words, output data with different confidence levels can be filtered using different data filtering strategies.

[0047] Accordingly, for the output data obtained by the business application using the initial model image processing, the output data can be divided into at least two groups according to the confidence level of the output data for the target object (detection target), and the output data can be filtered according to the group to obtain the target output data.

[0048] For example, the at least two groups may include at least a first group and a second group with decreasing confidence levels.

[0049] For example, the data filtering strategies for different groups are not exactly the same.

[0050] For example, when the output data is divided into a first group and a second group based on the confidence level for the target, the data filtering strategies for the first group and the second group can be different.

[0051] In one example, filtering the output data based on grouping as described above may include:

[0052] Using a tracking algorithm, based on the correlation between the output data before and after the output data in the first group, erroneous data is eliminated from the output data in the first group to obtain the target output data in the first group.

[0053] For example, considering multiple frames of images acquired over a continuous time, such as consecutive video frames, the target object is typically continuous in those frames.

[0054] For example, in pedestrian detection for video data, the positions of the same pedestrian are usually continuous in consecutive video frames.

[0055] Accordingly, for output data groups with high confidence in the target object, such as the first group mentioned above, a tracking algorithm can be used to eliminate erroneous data in the output data of the first group based on the preceding and following related output data, so as to obtain the target output data in the first group.

[0056] For example, for any output data in the first group, the output data corresponding to the preceding N1 consecutive frames of the video frame to which the output data belongs, and the output data corresponding to the following N2 consecutive frames, are the associated output data of the output data in the first group.

[0057] For example, N1≥1, N2≥1.

[0058] For example, if a pedestrian is detected in a video frame, but the pedestrian is not detected in the previous or next frame, then the pedestrian detected in that video frame is a false detection.

[0059] In one example, filtering the output data based on grouping as described above may include:

[0060] The output data in the second group is displayed in the preset review interface, and the erroneous data in the output data in the second group is eliminated according to the received review instructions to obtain the target output data in the second group.

[0061] For example, considering the grouping of output data with low confidence levels for the target object, the proportion of erroneous data in the output data within this group will be relatively high, making it difficult to remove erroneous data using algorithms. However, the correct data in this portion of the output data can greatly improve the efficiency of model evolution.

[0062] Therefore, for output data groups with relatively low confidence in the target object, such as the second group mentioned above, the output data in the second group can be displayed in a preset review interface, and relevant personnel can manually filter it. Based on the received review instructions, the output data in the second group can be erroneous to obtain the target output data in the second group.

[0063] In one example, before dividing the output data into at least two groups based on the confidence level of the output data for the target, the following may also be included:

[0064] Based on the confidence level of the output data for the target object, the output data with a confidence level for the target object that is higher than a preset confidence threshold is filtered; wherein, the preset confidence threshold is greater than or equal to the upper limit of the confidence level corresponding to the first group.

[0065] Based on the confidence level of the output data for the target object, the output data is divided into at least two groups, which may include:

[0066] Based on the confidence level of the filtered output data for the target object, the filtered output data is divided into at least two groups.

[0067] For example, output data with excessively high confidence in the target object may not be used for training set expansion as it has a poor effect on model improvement and reduces training efficiency.

[0068] Accordingly, for the output data obtained by the business application using the initial model for image processing, the output data with a confidence level higher than the preset confidence threshold can be filtered based on the confidence level of the output data for the target object. Based on the confidence level of the filtered output data for the target object, the filtered output data can be divided into at least two groups.

[0069] It should be noted that for output data with excessively high confidence in the target object, it can also be divided into a group (such as the third group). For the output data in the third group, the data filtering strategy can be to delete all of them.

[0070] Based on the confidence level of the output data for the target object, the output data can be divided into three groups: data with a confidence level higher than the first confidence threshold (which can be set according to actual needs) is assigned to the third group; data with a confidence level lower than the second confidence threshold (which can be set according to actual needs) is assigned to the second group; and data with a confidence level between the second and first confidence thresholds is assigned to the first group. For the output data of the third group, the data filtering strategy can be to delete all of them. The data filtering strategies for the first and second groups can be found in the relevant descriptions in the above embodiments.

[0071] In some embodiments, after optimizing the initial model based on the expanded training set to obtain the evolutionary model in step S130, the process may further include:

[0072] When the detection rate of the evolutionary model on the test set is lower than the preset detection rate threshold, the training set is expanded and the model is optimized based on the output data obtained by the business application using the evolutionary model for image processing, until the detection rate of the resulting evolutionary model on the test set is not lower than the preset detection rate threshold.

[0073] For example, when an evolutionary model is obtained in accordance with the manner described in steps S100 to S130, the obtained evolutionary model can be tested based on a test set to determine the detection rate of the evolutionary model on the test set, and the detection rate can be compared with a preset detection rate threshold (which can be set according to actual needs).

[0074] When the detection rate of the evolutionary model on the test set is lower than the preset detection rate threshold, the training set can be expanded and the model optimized based on the output data obtained by the business application terminal using the evolutionary model for image processing, to obtain the latest evolutionary model. The detection rate of the evolutionary model on the test set is then compared with the preset detection rate threshold. If the detection rate of the evolutionary model on the test set is not lower than the preset detection rate threshold, the model evolution is completed. If the detection rate of the evolutionary model on the test set is still lower than the preset detection rate threshold, the training set expansion and model optimization are performed again in the above manner.

[0075] It should be noted that when the detection rate of the evolutionary model obtained in step S130 on the test set is not lower than the preset detection rate threshold, it can be determined that the evolutionary model meets the usage requirements, and the model evolution is stopped.

[0076] To enable those skilled in the art to better understand the technical solutions provided in the embodiments of this application, the technical solutions provided in the embodiments of this application are described below with reference to specific examples.

[0077] In this embodiment, a training set is constructed with a relatively small initial amount of labeled data. During application, data beneficial to improving the model's generalization ability is acquired to automatically expand the training set. The model is then optimized and trained based on the expanded training set to obtain an evolved model. Through a iterative process, the model's generalization ability is gradually expanded, i.e., its performance is improved, ultimately outputting a high-performance model. This achieves the effect of training with large amounts of data with minimal or no human intervention, through a system-wide iterative evolution process. For a detailed implementation process, please refer to [link to relevant documentation]. Figure 2 ,like Figure 2 As shown, the implementation process may include:

[0078] 1. Initial data collection: Initial data collection may include data collection and data annotation.

[0079] For example, data collection may include collecting images or videos. Data annotation can be done manually or using an auxiliary annotation system.

[0080] For example, the amount of training data required for the initial data collection can be very small, such as obtaining a few hundred or even a few dozen training samples.

[0081] 2. Initial model training to obtain the initial model.

[0082] For example, model training can be performed using any machine learning training system, including but not limited to training frameworks such as Tensorflow, PyTorch, and Caffe.

[0083] 3. Load the initial model into the application system (i.e., the business application).

[0084] For example, the application system can be any machine learning inference system, including but not limited to inference servers, mobile phones, embedded devices, etc.

[0085] 4. Data Acquisition: Acquire the output data (which can be called detection data) obtained by the application system through image processing using the initial model. For example, computer vision outputs typically include a confidence parameter, and the confidence parameter is positively correlated with the current model's target detection performance.

[0086] 5. Confidence Filtering: Based on the confidence level of the detected data, the obtained detected data is filtered to remove data with excessively high confidence.

[0087] For example, considering that detected data with excessively high confidence has a poor effect on improving the model and will reduce training efficiency, its value in expanding the training set is low and it can be filtered out.

[0088] For example, the detected data with a confidence level greater than 90% can be filtered out to obtain the detected data with a confidence level less than or equal to 90% for subsequent processing.

[0089] 6. Detection data screening: For the filtered detection data, different data screening strategies can be adopted according to the confidence level of the detection data to obtain detection data with accuracy that meets the requirements (i.e., the target output data mentioned above).

[0090] For example, considering that the proportion of erroneous data in the detected data with low confidence is relatively high, and erroneous data cannot be automatically removed by the algorithm, the detected data with low confidence can be screened manually.

[0091] For example, the filtered detected data (assuming the confidence level is less than or equal to 90%) can be divided into two groups: a group with confidence levels of [50%, 90%] (i.e., the first group mentioned above) and a group with confidence levels of [0%, 50%] (i.e., the second group mentioned above).

[0092] For the first group, a tracking algorithm can be used to exclude erroneous data from the detected data in the first group based on the correlation between the detected data and the detected data in the group.

[0093] For example, the tracking algorithms described above may include, but are not limited to, IOU (Intersection over Union) comparison and image content comparison.

[0094] In addition, algorithms can be used to filter and remove detected data with high similarity or blurred images.

[0095] For the second group, manual screening can be used to exclude erroneous data from the detected data in the second group.

[0096] Among them, for groups with low confidence, although manually filtering out erroneous data will consume some manpower, it can obtain more training data, and the training data obtained will have a better effect on improving the model's generalization ability, enabling the model to evolve rapidly and improving the efficiency of model optimization training.

[0097] It should be noted that, considering that during the model evolution process, as erroneous data is eliminated and model performance is optimized, the amount of low-confidence data will decrease. Therefore, in the initial model optimization training process, the data is manually screened in the first few rounds (which can be set according to the actual scenario). After that, the data can be automatically screened by the algorithm until the detection rate of the evolved model on the test set is not lower than the preset detection rate threshold.

[0098] 7. Filtered Data Output: Organize the detected data after filtering in step 6 into the data format required for training, and merge it into the training set to achieve automatic expansion of the training set.

[0099] 8. Model optimization training: Optimize and train the initial model based on the expanded training set.

[0100] For example, the training system used for model optimization training can be the same as or different from the training system used for the initial model training.

[0101] 9. Publish the evolution model: Load the evolution model into the application system, and the application system will use the evolution model for graphics processing.

[0102] 10. Model Testing: Test the latest model using the test set to determine whether the detection rate of the evolved model on the test set is not lower than the preset detection rate threshold. If so, the evolved model is determined to meet the usage requirements, and model evolution is stopped; otherwise, continue to optimize the model.

[0103] For example, when it is determined that the evolutionary model meets the usage requirements (i.e., whether the detection rate of the evolutionary model on the test set is not lower than the preset detection rate threshold), a prompt message can be output so that relevant personnel can determine whether to stop the model evolution. When an operation instruction indicating to stop the model evolution is detected, the model evolution is stopped; otherwise, the model continues to be optimized.

[0104] As can be seen, in the above process, the output data obtained by image processing of the trained model is filtered and added to the training set, continuously expanding the quantity and coverage of the training set and making it closer to the real scene. Based on the expanded training set, the training system is used to train and evolve the output model. At this time, the model's generalization ability will cover the newly added data range. In terms of application performance, some targets with high confidence (e.g., 50%–90%) will be added to the very high confidence interval (e.g., confidence greater than 90%), and some targets with low confidence (e.g., confidence less than 50%) will be added to the very high confidence interval. Through multiple rounds of evolution, the coverage of the "very high confidence interval" will eventually become larger and larger, that is, the edge region of the model's generalization ability is expanded, the detection performance of the model is effectively improved, and the model's automatic evolution is achieved.

[0105] The method provided in this application has been described above. The apparatus provided in this application is described below:

[0106] Please see Figure 3 This is a schematic diagram of the structure of an image processing device provided in an embodiment of this application, as shown below. Figure 3 As shown, the image processing apparatus may include:

[0107] The pre-training unit 310 is used to train the model based on the initial training set using a preset training system to obtain the initial model;

[0108] The filtering unit 320 is used to filter the output data obtained by the business application terminal using the initial model for image processing, and obtain target output data whose confidence and accuracy meet preset conditions.

[0109] The expansion unit 330 is used to expand the initial training set according to the target output data to obtain an expanded training set;

[0110] The optimization training unit 340 is used to optimize and train the initial model based on the expanded training set to obtain an evolutionary model, so that the business application terminal can perform image processing based on the evolutionary model.

[0111] In some embodiments, the filtering unit 320 filters the output data obtained by the business application terminal using the initial model for image processing, including:

[0112] Based on the confidence level of the output data for the target object, the output data is divided into at least two groups; the at least two groups include at least a first group and a second group with corresponding confidence levels decreasing sequentially.

[0113] The output data is filtered according to the grouping to obtain the target output data; the data filtering strategies for different groups are not exactly the same.

[0114] In some embodiments, the filtering unit 320 filters the output data according to groups, including:

[0115] Using a tracking algorithm, based on the preceding and following output data in the first group, erroneous data is eliminated from the output data in the first group to obtain the target output data in the first group.

[0116] In some embodiments, the filtering unit 320 filters the output data according to groups, including:

[0117] The output data in the second group is displayed in the preset audit interface, and the erroneous data in the output data in the second group is eliminated according to the received audit instructions to obtain the target output data in the second group.

[0118] In some embodiments, before the filtering unit 320 divides the output data into at least two groups based on the confidence level of the output data for the target, it further includes:

[0119] Based on the confidence level of the output data for the target object, output data with a confidence level for the target object higher than a preset confidence threshold are filtered; wherein, the preset confidence threshold is greater than or equal to the upper limit of the confidence level corresponding to the first group;

[0120] The filtering unit 320 divides the output data into at least two groups based on the confidence level of the output data for the target object, including:

[0121] Based on the confidence level of the filtered output data for the target object, the filtered output data is divided into at least two groups.

[0122] In some embodiments, after the optimization training unit 340 optimizes and trains the initial model based on the expanded training set to obtain an evolutionary model, it further includes:

[0123] When the detection rate of the evolutionary model on the test set is lower than the preset detection rate threshold, the training set is expanded and the model is optimized based on the output data obtained by the business application terminal using the evolutionary model for image processing, until the detection rate of the obtained evolutionary model on the test set is not lower than the preset detection rate threshold.

[0124] This application provides an electronic device including a processor and a memory, wherein the memory stores machine-executable instructions that can be executed by the processor, and the processor executes the machine-executable instructions to implement the image processing method described above.

[0125] Please see Figure 4 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. The electronic device may include a processor 401 and a memory 402 storing machine-executable instructions. The processor 401 and the memory 402 can communicate via a system bus 403. Furthermore, by reading and executing the machine-executable instructions corresponding to image processing logic in the memory 402, the processor 401 can execute the image processing method described above.

[0126] The memory 402 mentioned in this document can be any electronic, magnetic, optical, or other physical storage device that can contain or store information such as executable instructions, data, etc. For example, machine-readable storage media can be: RAM (Random Access Memory), volatile memory, non-volatile memory, flash memory, storage drives (such as hard disk drives), solid-state drives, any type of storage disk (such as optical discs, DVDs, etc.), or similar storage media, or combinations thereof.

[0127] In some embodiments, a machine-readable storage medium, such as Figure 4 The memory 402 in the memory, which is a machine-readable storage medium, stores machine-executable instructions that, when executed by a processor, implement the image processing method described above. For example, the storage medium may be ROM, RAM, CD-ROM, magnetic tape, floppy disk, or optical data storage device.

[0128] This application also provides a computer program stored in a read storage medium, such as... Figure 4 The memory 402 is used, and when the processor executes the computer program, it causes the processor 401 to execute the image processing method described above.

[0129] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0130] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. An image processing method, characterized by, The method comprises the following steps: According to the initial training set, the preset training system is used for model training to obtain an initial model; The output data obtained by the business application end using the initial model for image processing is screened to obtain target output data that meets the preset conditions in terms of confidence and accuracy; According to the target output data, the initial training set is expanded to obtain an expanded training set; According to the expanded training set, the initial model is optimized and trained to obtain an evolution model, so that the business application end performs image processing according to the evolution model; The screening of the output data obtained by the business application end using the initial model for image processing comprises the following steps: According to the confidence of the output data for the target object, the output data with a confidence higher than a preset confidence threshold for the target object is filtered; the preset confidence threshold is greater than or equal to the upper limit of the confidence corresponding to the first group; According to the confidence of the filtered output data for the target object, the filtered output data is divided into at least two groups; the at least two groups at least include a first group and a second group corresponding to a confidence that decreases in turn; According to the grouping, the output data is screened to obtain the target output data; the data screening strategies of different groups are not completely the same; The screening of the output data according to the grouping comprises the following steps: Using a tracking algorithm, the output data in the first group is excluded from error data according to the front and rear associated output data of the output data in the first group to obtain target output data in the first group; The output data in the second group is displayed in a preset audit interface, and the output data in the second group is excluded from error data according to the received audit instruction to obtain target output data in the second group; For the front and rear associated output data of any output data in the first group, it comprises: The output data corresponding to the N1 consecutive frames before the video frame to which the output data belongs, and the output data corresponding to the N2 consecutive frames after the video frame to which the output data belongs; N1≥1, N2≥1.

2. The method of claim 1, wherein, After the evolution model is obtained by optimizing and training the initial model according to the expanded training set, the following steps are further included: When the detection rate of the evolution model on the test set is lower than the preset detection rate threshold, the output data obtained by the business application end using the evolution model for image processing is used for training set expansion and model optimization until the detection rate of the obtained evolution model on the test set is not lower than the preset detection rate threshold.

3. An image processing apparatus characterized by comprising: The method comprises the following steps: A pre-training unit is configured to train a model according to an initial training set using a preset training system to obtain an initial model; A screening unit is configured to screen output data obtained by a business application end using the initial model for image processing to obtain target output data that meets preset conditions in terms of confidence and accuracy; An expansion unit is configured to expand the initial training set according to the target output data to obtain an expanded training set; An optimization unit is configured to optimize and train the initial model according to the expanded training set to obtain an evolution model, so that the business application end performs image processing according to the evolution model. An optimization training unit is configured to perform optimization training on the initial model based on the extended training set to obtain an evolved model, so that the service application end performs image processing based on the evolved model. The screening unit screens output data obtained by the service application end performing image processing based on the initial model, including: filtering output data for which the confidence in the target object is higher than a preset confidence threshold based on the confidence of the target object in the output data; the preset confidence threshold is greater than or equal to the upper limit of the confidence of the first group; dividing the filtered output data into at least two groups based on the confidence of the target object in the filtered output data; the at least two groups at least include a first group and a second group in which the confidence decreases successively; screening the output data based on the groups to obtain the target output data; the data screening strategies of different groups are not completely the same; The screening unit screens the output data based on the groups, including: performing error data exclusion on the output data in the first group based on the front and rear associated output data of the output data in the first group using a tracking algorithm to obtain target output data in the first group; displaying the output data in the second group in a preset review interface and performing error data exclusion on the output data in the second group based on a received review instruction to obtain target output data in the second group; For the front and rear associated output data of any output data in the first group, it includes output data corresponding to the previous N1 consecutive frames of the video frame to which the output data belongs, and output data corresponding to the subsequent N2 consecutive frames; N1≥1, N2≥1.

4. The apparatus of claim 3, wherein, After the optimization training unit performs optimization training on the initial model based on the extended training set to obtain the evolved model, the method further includes: When the detection rate of the evolved model on the test set is lower than a preset detection rate threshold, performing training set extension and model optimization based on output data obtained by the service application end performing image processing based on the evolved model until the detection rate of the evolved model on the test set is not lower than the preset detection rate threshold.

5. An electronic device, comprising: The processor is configured to execute the machine executable instructions to implement the method of claim 1 or 2.

6. A machine-readable storage medium, characterized in that, The machine readable storage medium stores a computer program, which is executed by the processor to implement the method of claim 1 or 2.

Images