Method and system for detecting target objects through discriminant models

Abstract

The present invention relates to the field of machine learning. The present invention provides a method for detecting a target object using a discriminant model, the method comprising the following steps: obtaining raw data samples of the target object using a data acquisition module; detecting the target object based on the raw data samples using a discriminant model; and enhancing the discriminant model based on a generative model, wherein the enhancement is performed in the following manner: preliminarily training the discriminant model using the raw data samples; performing defect analysis on the raw data samples based on the preliminary training results of the discriminant model to obtain a list of defect scenario data; training a generative model using the list of defect scenario data to directionally generate defect scenario data using the generative model; and retraining the discriminant model based on at least the defect scenario data. The present invention also provides a system for enhancing a discriminant model based on a generative model and a machine-readable program carrier.

Description

Technical Field

[0001] The present invention relates to a method for detecting a target object through a discriminant model, a system for enhancing a discriminant model based on a generative model, and a machine-readable program carrier. Background Art

[0002] With the development of artificial intelligence and big data, discriminative models based on deep neural networks have been widely used in natural language processing, image recognition and understanding, and goal planning and decision-making. They have also spurred the development of multidisciplinary industrial technologies such as autonomous driving. To meet the requirements for predictive accuracy, reliability, and generalization in different application scenarios, optimizing the performance of discriminative model algorithms has been a common challenge in both academia and industry. Generally speaking, there are two approaches to optimizing discriminative model performance. One is to improve model performance metrics by optimizing the model structure and various parameters (such as the neural network structure and hyperparameters) within the algorithm itself. This approach not only requires extensive data mining of the original data but also requires continuous debugging during the neural network training process, resulting in significant overhead. The other approach is to improve model performance metrics by enhancing the training data samples (for example, increasing the number of training data samples and enriching the data diversity).

[0003] There are two main data enhancement methods that already exist in the prior art. One is to collect and label the original data in a specific application scenario to enrich the training subset. However, this method is not only costly and inefficient, but also difficult to obtain low-probability events, which often leads to an uneven distribution of data samples and affects the robustness of the model in low-probability scenarios. Another method is to achieve the purpose of data enhancement by preprocessing existing data samples before model training. For example, in the field of image detection, image samples can be flipped, translated, and scaled. However, this method can only enhance the diversity of data samples in limited dimensions such as the size and angle of existing samples, and cannot generate new data samples. Therefore, it is impossible to fundamentally improve the scale and diversity of data samples. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for detecting a target object through a discriminant model, a system for enhancing a discriminant model based on a generative model, and a machine-readable program carrier, so as to at least solve some of the problems in the prior art.

[0005] According to a first aspect of the present invention, a method for detecting a target object using a discriminant model is provided, the method comprising the following steps:

[0006] S1: Obtain the original data sample of the target object through the data acquisition module;

[0007] S2: Detect the target object based on the original data sample through the discriminant model;

[0008] S3: Enhance the discriminant model based on the generative model, wherein the enhancement is performed in the following manner:

[0009] S31: Preliminary training of the discriminant model with the help of original data samples;

[0010] S32: performing defect analysis on the original data sample based on the preliminary training result of the discriminant model to obtain a defect scenario data list;

[0011] S33: training a generative model with the defect scenario data list to generate defect scenario data in a directionally manner through the generative model;

[0012] S34: Retrain the discriminant model based at least on the defect scenario data.

[0013] The present invention particularly includes the following technical advantages: It is different from the prior art in which data samples are directly enhanced and expanded. The present invention proposes to first screen out the deficiencies of data samples under the current discriminant model task through pre-training of the discriminant model, and then use the generative model to achieve targeted enhancement of the data samples, thereby advantageously solving the problems of low data collection efficiency, insufficient data diversity and uneven distribution in the existing model training process. Through this two-stage training process, the performance of the discriminant model can be optimized more efficiently. In addition, since the generated data samples are "new" data samples generated in a directionally manner with the help of the generative model, rather than being limited to a simple transformation of the existing data samples, it can essentially make up for the problem of uneven distribution of the original data samples and play a role in balancing the data samples.

[0014] Optionally, step S31 includes: defining corresponding evaluation indicators according to the discriminant task of the discriminant model, so as to evaluate the preliminary training results of the discriminant model.

[0015] This particularly achieves the following technical advantages: by defining corresponding evaluation indicators, the quality of the discrimination results of the discriminant model can be quantified, so that the statistical distribution of the model training results can be more intuitively derived.

[0016] Optionally, step S32 includes: analyzing the original data sample in terms of data distribution characteristics; and / or analyzing the original data sample in terms of performance of the discriminant model.

[0017] This method achieves the following technical advantages: it not only considers the uneven distribution of raw data samples, thereby directly combining the specific tasks or annotation information of the discriminant model to extract rare data categories, but also considers the performance differences of the discriminant model for raw data samples, thereby extracting special data categories based on the characteristics of the discriminant model itself. As a result, it can perform defect analysis of raw data samples more comprehensively from different aspects.

[0018] Optionally, the defect scenario data list includes: a low-probability scenario list and / or a failure scenario list, wherein the low-probability scenario represents a data category whose quantity share in the original data sample is lower than a determined threshold, and wherein the failure scenario represents a data category for which the discriminant model has performance defects.

[0019] Optionally, step S32 includes: using unsupervised learning algorithms such as cluster analysis and principal component analysis to reduce the dimension and cluster the original data samples, so as to find clusters with smaller distribution in the original data samples as low-probability scenarios; and / or, classifying the preliminary training results of the discriminant model, so as to find data categories for which the discriminant model has performance defects as failure scenarios.

[0020] Optionally, in terms of image-based person detection, low-probability scenarios and / or failure scenarios include: strange person postures, strange person clothing, large-area person occlusion, and / or blurred person images.

[0021] Optionally, the step S33 includes: obtaining a model control condition belonging to the defect scene; and generating defect scene data in a decoupled manner in terms of the model control condition with the help of a generative model. Preferably, the model control condition includes appearance information and posture information.

[0022] Optionally, the step S33 includes: directionally generating the defect scenario data by means of a generative adversarial network, a variational autoencoder and / or an autoregressive model.

[0023] Optionally, step S34 includes: merging the defect scene data generated in a targeted manner with the original data sample, and retraining the discriminant model with the combined data sample; and / or, continuing to train the discriminant model that has been preliminarily trained directly with the defect scene data generated in a targeted manner.

[0024] According to a second aspect of the present invention, a system for enhancing a discriminative model based on a generative model is proposed, the system comprising:

[0025] A data acquisition module configured to acquire raw data samples of a target object;

[0026] a discriminative model training module configured to perform preliminary training on the discriminative model using raw data samples;

[0027] a defect analysis module configured to perform defect analysis on raw data samples based on the preliminary training results of the discriminant model to obtain a defect scenario data list; and

[0028] A generative model training module is configured to train a generative model with the help of a defect scenario data list, so as to generate defect scenario data in a directionally manner through the generative model;

[0029] The discriminant model training module is further configured to be able to retrain the discriminant model based at least on the defect scenario data.

[0030] According to a third aspect of the present invention, a machine-readable program carrier is provided, on which a computer program is stored for executing a method according to an aspect of the present invention when the computer program is run on a computer. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The present invention will be described in more detail below with reference to the accompanying drawings, so that the principles, features and advantages of the present invention can be better understood. The accompanying drawings include:

[0032] Figure 1 A flow chart of a method for detecting a target object by using a discriminant model according to an exemplary embodiment of the present invention is shown;

[0033] Figure 2 A flowchart of a method for detecting a target object by using a discriminant model according to the present invention is shown in conjunction with a specific application example;

[0034] Figure 3 A structural block diagram of a system for enhancing a discriminant model based on a generative model according to an exemplary embodiment of the present invention is shown;

[0035] Figure 4a and Figure 4b The training process and generation process of the generative model used in the method according to the present invention are illustrated with reference to a specific application example;

[0036] Figure 5 An exemplary embodiment of defect scene data directionally generated by means of the method according to the present invention is shown. DETAILED DESCRIPTION

[0037] In order to make the technical problems, technical solutions and beneficial technical effects to be solved by the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and multiple exemplary embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the scope of protection of the present invention.

[0038] Figure 1 A flowchart of a method for detecting a target object by using a discriminant model according to an exemplary embodiment of the present invention is shown.

[0039] In step S1, the raw data samples of the target object are obtained with the help of the data acquisition module. Here, for example, the natural data is collected and labeled in combination with the specific functional tasks of the discriminant model (such as target detection, intention estimation, speech recognition, motion control, etc.), the raw data format required for training the discriminant model (such as image, speech, control signal, etc.) and the annotation information format (object posture, object category, object intention, etc.), thereby obtaining the raw data samples with annotation information. Then, for example, the raw data samples are divided into training subsets data train , validation subset data val And the test subset data test .

[0040] In step S2, the target object is detected using a discriminant model based on the original data sample.

[0041] In step S3, the discriminant model is enhanced based on the generative model, wherein the enhancement is performed in the following manner:

[0042] In step S31, the selected discriminant model is trained based on the original training data sample. Here, for example, the corresponding machine learning framework platform is first selected in combination with the system environment and software architecture and the discriminant model is developed. Then the training subset data is used to train the discriminant model. train Input into the discriminant model for training. During the training iteration, the validation subset data can be used val The intermediate model D1 obtained by training the discriminant mode model is verified, and its optimal result is used as the preliminary training result D of the discriminant model. raw .

[0043] According to one embodiment, in step S31, corresponding evaluation indicators can be defined according to the specific discriminant task of the selected discriminant model to evaluate the preliminary training results of the discriminant model, and a corresponding evaluation tool chain can be developed. For example, the discriminant model can be trained on the training subset data. train Preliminary training results on D raw On the test subset data testTest on the test, and input all test results into the evaluation tool chain to obtain D raw The statistical results on the test subset are used to estimate the generalization ability of the model in actual use by using the discriminant effect of the trained model on the test subset. For example, confusion matrix, ROC curve, AUC value, KS curve, PR curve, PSI index, etc. can be used as model performance evaluation indicators.

[0044] In step S32, based on the preliminary training results of the discriminant model, the original data samples are subjected to defect analysis to obtain a list of defect scenario data. Here, such defect analysis can be performed in two aspects: on the one hand, the original data samples, especially the training subset data, can be subjected to defect analysis. train This analysis mainly focuses on the labeling information of the training data samples and combines the specific algorithm principles and task types of the selected discriminant model; on the other hand, the original data samples, especially the training subset data, can also be analyzed based on the performance evaluation indicators of the discriminant model. train To obtain the training subset data used train A list of potential problems under this discriminant model task.

[0045] In this case, it is especially possible to combine the model D raw By jointly analyzing the statistical results on the test subsets, we can obtain a list of defective data scenarios for the original data samples under the target task and application scenario. Based on the data-driven hypothesis, this defect scenario data list can be further divided into a list of low-probability scenarios and a list of performance defects. Low-probability scenarios, for example, represent data categories whose proportion in the original data samples is below a certain threshold, while failure scenarios, for example, represent data categories for which the discriminant model has performance defects.

[0046] In step S33, a generative model is built according to the output / input relationship under the target task and application scenario. Currently commonly used generative models include generative adversarial networks (GAN), variational autoencoders (VAE), and autoregression models. In order to meet the needs of subsequent data-oriented enhancement, the input variables used in the construction of the generative model should at least contain the annotation information required by the discriminant model, and as much as possible contain controllable variables of other characteristics. In addition, the generative model should be controlled to output data samples (such as images, videos, audio, etc.) in the same data format as the current annotated data samples.

[0047] After completing the construction of the generative model, for example, the training subset datatrain Input into the generative model for training. During the training process of the generative model, the validation subset data is used val The intermediate result G1 of the generative model is tested online, and the best result is selected as the final training result M of the generative model. gen It is worth mentioning that the input data used to train the generative model can not only use the training subset data of the discriminative model train , you can also use other existing test subset data test Incremental learning and transfer learning of labeled data samples that are independent of each other and meet the needs of the current discriminant application scenario

[0048] Next, the low-probability scenario list and the failure scenario list obtained in step S32 can be input into the trained generative model together to generate defect scenario data samples in batches. gen .

[0049] Then, in step S34, the defect scene data sample data generated by the orientation can be gen Compared with the original data samples, especially the original training subset data train The combined defect scenario data can be used as input data for retraining the discriminant model. Here, the combined defect scenario data can be input into the discriminant model built and preliminarily trained in step S31 to retrain the discriminant model. In particular, the intermediate model D1' obtained by training the discriminant model can be verified according to the validation subset in iterative training, and the optimal result can be selected as the final result D of the retraining of the discriminant model. gen '.

[0050] Figure 2 A flowchart of the method for detecting a target object by using a discriminant model according to the present invention is shown in conjunction with a specific application example.

[0051] exist Figure 2 In the illustrated embodiment, the discriminant model to be enhanced is an image-based object detection model, specifically a person keypoint detection model. During the person keypoint detection process, an image containing a person is input into the corresponding detection model, thereby extracting the person's keypoints and forming a person skeleton model. However, it should be noted that the present invention does not impose any restrictions on the target objects to be detected, the output format, or the task. Therefore, the target objects to be detected may also include vehicles, traffic lights, and the like.

[0052] In step S201, the original character image and its key point annotation information are obtained. Specifically, for example, image acquisition is performed in the following manner: different or single characters are photographed by a camera in combination with the required application scenario. In the case of application in an autonomous driving scenario, it is necessary to use a vehicle-mounted camera to collect pedestrian images on the road. Since this embodiment enhances the character key point detection model, the character key point annotation information is completed through manual annotation. In other usage scenarios, of course, a character detection algorithm can also be used to extract character key points (such as OpenPose, Pifpaf, HRNet, etc.). Here, the original data sample is divided into a training subset data train , validation subset data val And the test subset data test .

[0053] In step S202, the character key point detection model is trained based on the original character annotated data sample in step S201. Specifically, first select and build a character key point detection model framework D(x), such as OpenPose, Pifpaf and HRNet, etc. This embodiment does not impose any restrictions on the specific model framework. Here, the training subset data in step S201 is trained. train Input into the selected character key point detection model framework D(x) for preliminary training, and use the validation subset data in the training iteration process val Test and evaluate to obtain the best training results as the preliminary human key point detection model D raw .

[0054] In step S203, based on the test subset data obtained in step S201 test The preliminary character key point detection model D obtained by training in step S202 raw In the character key point detection algorithm, commonly used performance evaluation indicators include but are not limited to OKS, PCK, Lamir, etc. Specifically, the preliminary character key point detection model D can be used. raw For the test subset data test The key points of the person Ii in each frame image are detected to obtain the detection result Pi, and then the selected performance indicator equation (such as OKS) is used to calculate each detection result Pi based on the true value of the person annotation of each frame image to obtain the accuracy indicator OKSi of each detection result. In addition, the test subset data can also be used to calculate the accuracy indicator OKSi of each detection result. test Perform statistical analysis to obtain a preliminary character key point detection model D raw On the test subset data test The accuracy statistics on .

[0055] In step S204, starting from the data-driven hypothesis, its main purpose is to conduct statistical analysis of relevant dimensions on the original data samples and the models trained therefrom, thereby mining defect scenarios of the original data samples. Specifically, in this embodiment, on the one hand, the training subset data in step S201 can be combined with the human key point detection model algorithm selected in step S202 (such as clustering, principal component analysis and other unsupervised learning algorithms) to perform statistical analysis on the human key point detection model algorithm selected in step S202 (such as clustering, principal component analysis and other unsupervised learning algorithms). train Perform dimensionality reduction and clustering to find the clusters with smaller distribution in the data sample; on the other hand, the preliminary training results D in step S203 can also be used to raw On the test subset data test The statistical results of the performance analysis on the above data are classified to find the data samples corresponding to the poor performance. When the data samples are analyzed and classified, it can be obtained in which data scenarios the model has a large performance defect. It should be pointed out that the specific performance defect analysis and mining methods include but are not limited to the above two methods, and in specific operations, the above two performance defect analysis and mining methods can be used separately. Finally, a list of low-probability scenarios with deficiencies in data distribution and a list of failure scenarios with model performance defects are obtained through the corresponding defect analysis. In this embodiment, defect scenes may include but are not limited to strange character postures, clothing, large-area character occlusion, blurred character images, etc.

[0056] In step S205, a generative model is built based on the required training samples (defect scene data list) and their annotated data format. The discrimination task in this embodiment is character key point detection, so the output result of the generative model built should be an image containing a character, and the input of the generative model should at least include but not be limited to the key point annotation information corresponding to the character image. Commonly used generative models include generative adversarial networks, variational autoencoders, autoregressive models, and extensions or variants of these models. This embodiment uses a variational autoencoder, and the basic expression of its image generation process is:

[0057] Y=G(x c , x a )

[0058] Among them, x c Indicates the key point of the character Ii in the image, x a Represents the appearance of the character, and Y represents the output of the generative model including the key points x of the character c (corresponding to the character's posture) and appearance x a Character images, usually x c and x a is called the model control condition. The basic expression of the generative model training process is:

[0059] G=argmin likelihood[Yi,G(x ci ,x ai )]

[0060] Here, based on the training subset data divided in step S201 train The generative model is iteratively trained, wherein the intermediate model obtained in each training cycle is Gi. During the training iteration process, the validation subset data can be used val Each intermediate model Gi is verified, and the optimal result is selected as the final generative model G.

[0061] In step S206, according to the defect scenario list obtained in step S204, a certain amount of model control conditions (x c , x a It is worth pointing out that (x c , x a ) can satisfy only one or all of the failure scenario conditions. Specifically, these model control conditions can be directly obtained from the existing training subset data train The model control conditions corresponding to the defect scene are input into the generative model G obtained in step S205, so that a large amount of defect scene data (here refers to human image data) can be generated in a targeted manner. gen , the key point information of the characters in these defect scene data and the model control conditions x input into the generative model c One-to-one correspondence, so the model control condition x c The key point annotation data of the person can be used as additional data samples for directional generation. In particular, each model control condition can be considered independently when generating defect scene data. For example, the key point x of each person image can be c With character appearance x a The split is performed, and then a new combination of character key points and character appearance is generated through a directed generation process.

[0062] In step S207, the generated defect scene data data gen And the original training subset data train Input into the character key point recognition model built in step S202 for training. It should be noted that the defect scene data generated in the direction can be firstly gen With the training subset data trainMerge and then input them together into the character key point detection model framework D(x) for retraining. However, it is also possible that after the training subset data train Based on the training model, input the newly generated defect scenario data gen Retraining (retrain) can also be considered as additional data samples data gen Any other meaningful addition time and method. During the training iteration, use the validation subset data val The retrained key point detection model is tested and evaluated to obtain the best training result as the final key point detection model M gen Since in the final character key point detection model M gen The training process introduces a large number of generated data samples under low probability scenarios / failure scenarios gen Therefore, its comprehensive performance index, especially the performance index in the face of defect scenes, is better than the preliminary human key point detection model D raw Will be significantly improved.

[0063] Figure 3 A structural block diagram of a system for enhancing a discriminant model based on a generative model according to an exemplary embodiment of the present invention is shown.

[0064] like Figure 3 As shown, the system 1 includes a data acquisition module 10, a discriminant model training module 20, a defect analysis module 30, and a generative model training module 40. In addition, the system may further include a data merging module 50.

[0065] The data acquisition module 1 is used to obtain raw data samples of the target object. Here, depending on the specific functional tasks and usage scenarios of the discriminant model, the data acquisition module can be constructed as an image sensor, video sensor, radar sensor, lidar sensor, ultrasonic sensor, infrared sensor, etc.

[0066] After the raw data is collected, the raw data samples are provided to the discriminant model training module 20 for preliminary training of the discriminant model. For example, the discriminant model training module 20 may include a discriminant model training framework 210, a discriminator 220, and a discriminant model performance evaluation unit 230. Based on the discriminative task of the selected discriminant model, corresponding evaluation indicators may be defined in the discriminant model performance evaluation unit 230 to evaluate the preliminary training results of the discriminant model.

[0067] The discriminant model training module 20 is also connected to the defect analysis module 30 in a data transmission manner, so that the preliminary training results of the discriminant model can be provided to the defect analysis module 30. In addition, the defect analysis module 30 is also connected to the data acquisition module 10. Exemplarily, the defect analysis module 30 may include a data distribution analysis unit 310, a performance defect analysis unit 320, and a joint analysis unit 330. The defect analysis module 30 is configured to be able to perform defect analysis on the original data samples based on the preliminary training results of the discriminant model. This defect analysis can be performed in particular with the help of the data distribution analysis unit 310 and the performance defect analysis unit 320. Then, in the joint analysis unit 330, a defect scenario data list can be obtained by comprehensively considering the defect analysis of different methods.

[0068] This defect scenario data list is then provided to the generative model training module 40. This generative model training module 40 is also connected to the data acquisition module 10 to build a generative model using the annotation information required by the discriminant model. This generative model can also be trained based on the original data samples (particularly the corresponding training subset). After training is complete, the defect scenario data list obtained from the defect analysis module 30 can be input into the generative model, enabling targeted generation of defect scenario data.

[0069] Next, the directionally generated defect scenario data can be directly provided by the generative model training module to the discriminative model training framework 210 so as to input the newly generated defect scenario data for retraining based on the discriminative model that has been trained with the original data samples.

[0070] Additionally or alternatively, the targeted defect scene data can be merged with the original data samples in the data merging module 50, thereby effectively adjusting the data distribution of the original data samples to obtain a data-balanced merged data sample. The merged data sample is then input into the person discriminant model training framework to retrain the discriminant model.

[0071] Figure 4a and Figure 4b The training process and generation process of the generative model used in the method according to the present invention are illustrated with reference to a specific application example.

[0072] exist Figure 4a, which can be performed offline. For example, the original person image 402 and its annotated information 401 are input into a generative model 404 for training, thereby obtaining a reconstructed image 406 of the target person. The generative model includes, but is not limited to, generative adversarial neural networks, variational autoencoders, and their derivatives. Optionally, the generative model also includes a pre-processing module 403 and a post-processing module 405 for data processing.

[0073] exist Figure 4b The process of generating a generative model is shown in FIG. Here, the generative model has already been trained and learned using, for example, raw data samples. Therefore, the model control conditions corresponding to the defect scene data samples (i.e., the target person's pose information 401' and appearance information 402') are now provided to the generative model as input data, thereby generating defect scene data samples in a targeted manner using the generative model.

[0074] Figure 5 An exemplary embodiment of defect scene data directionally generated by means of the method according to the present invention is shown.

[0075] like Figure 5 As shown, defect scene data generated directionally using the method according to the present invention is shown, where the discriminant model's functional task is character posture detection. The first row shows the first model control condition (here, character key point information or posture information) considered during the generative model training process. Subsequent rows show additional model control conditions (e.g., character appearance information) considered during the generative model training process.

[0076] exist Figure 5 In each column, new defect scene data samples are generated in a decoupled manner using pose and appearance information belonging to low-probability scenarios and / or failure scenarios. This means, for example, combining different target person poses with different target person appearances to form new person images. In addition to pose and appearance information, other control conditions for the targeted generation of defect scene data include person occlusion, person blur (ghosting), and image brightness defects.

[0077] Although specific embodiments of the present invention are described in detail herein, they are provided for illustrative purposes only and should not be considered to limit the scope of the present invention. Various substitutions, changes, and modifications may be conceived without departing from the spirit and scope of the present invention.

Claims

1. A method for detecting a target object using a discriminant model, the method comprising the following steps: S1: Obtain the original data sample of the target object through the data acquisition module; S2: Detect the target object based on the original data sample through the discriminant model; S3: Enhance the discriminant model based on the generative model, wherein the enhancement is performed in the following manner: S31: Preliminary training of the discriminant model with the help of original data samples; S32: Based on the preliminary training results of the discriminant model, defect analysis is performed on the original data samples to obtain a defect scenario data list, wherein the original data samples are analyzed in terms of the performance of the discriminant model to find data samples corresponding to poor performance of the discriminant model in the original data samples. After analyzing and classifying the data samples, failure scenarios with model performance defects are obtained. The defect scenario data list includes a list of the failure scenarios. S33: training a generative model with the defect scenario data list to generate defect scenario data in a directionally manner through the generative model; S34: Retrain the discriminant model based at least on the defect scenario data.

2. The method according to claim 1, wherein The step S31 includes: defining corresponding evaluation indicators according to the discriminant task of the discriminant model, so as to evaluate the preliminary training results of the discriminant model.

3. The method according to claim 1 or 2, wherein The step S32 includes: The original data samples are analyzed with respect to data distribution characteristics.

4. The method according to claim 1 or 2, wherein: The defect scenario data list includes: a low-probability scenario list and / or a failure scenario list, wherein the low-probability scenario represents a data category whose quantity share in the original data sample is lower than a certain threshold.

5. The method according to claim 3, wherein The defect scenario data list includes: a low-probability scenario list and / or a failure scenario list, wherein the low-probability scenario represents a data category whose quantity share in the original data sample is lower than a certain threshold.

6. The method according to claim 4, wherein: The step S32 includes: Using unsupervised learning algorithms such as cluster analysis and principal component analysis to reduce the dimensionality and cluster the original data samples, so as to identify clusters with smaller distribution in the original data samples as low-probability scenarios; and / or The preliminary training results of the discriminative model are classified in order to find out the data categories for which the discriminative model has performance deficiencies as failure scenarios.

7. The method according to claim 5, wherein: The step S32 includes: Using unsupervised learning algorithms such as cluster analysis and principal component analysis to reduce the dimensionality and cluster the original data samples, so as to identify clusters with smaller distribution in the original data samples as low-probability scenarios; and / or The preliminary training results of the discriminative model are classified in order to find out the data categories for which the discriminative model has performance deficiencies as failure scenarios.

8. The method according to claim 4, wherein: In terms of image-based person detection, the low-probability scenarios and / or failure scenarios include: strange person postures, strange person clothing, large-area person occlusion and / or blurred person images.

9. The method according to any one of claims 5 to 8, wherein In terms of image-based person detection, the low-probability scenarios and / or failure scenarios include: strange person postures, strange person clothing, large-area person occlusion and / or blurred person images.

10. The method according to any one of claims 1, 2, 5, 6, 7 and 8, wherein: The step S33 includes: Obtain the model control conditions (x c , x a );as well as With the help of the generative model, the control conditions (x c , x a ) are decoupled to generate the defect scenario data.

11. The method according to claim 3, wherein: The step S33 includes: Obtain the model control conditions (x c , x a );as well as With the help of the generative model, the control conditions (x c , x a ) are decoupled to generate the defect scenario data.

12. The method according to claim 4, wherein: The step S33 includes: Obtain the model control conditions (x c , x a );as well as With the help of the generative model, the control conditions (x c , x a ) are decoupled to generate the defect scenario data.

13. The method according to any one of claims 1, 2, 5, 6, 7, 8, 11, and 12, wherein: The step S34 includes: Merging the directed generated defect scenario data with the original data samples, and retraining the discriminant model with the merged data samples; and / or The discriminant model that has been initially trained is further trained directly with the help of the targeted defect scenario data.

14. The method according to claim 3, wherein: The step S34 includes: Merging the directed generated defect scenario data with the original data samples, and retraining the discriminant model with the merged data samples; and / or The discriminant model that has been initially trained is further trained directly with the help of the targeted defect scenario data.

15. The method according to claim 4, wherein The step S34 includes: Merging the directed generated defect scenario data with the original data samples, and retraining the discriminant model with the merged data samples; and / or The discriminant model that has been initially trained is further trained directly with the help of the targeted defect scenario data.

16. The method according to claim 10, wherein The step S34 includes: Merging the directed generated defect scenario data with the original data samples, and retraining the discriminant model with the merged data samples; and / or The discriminant model that has been initially trained is further trained directly with the help of the targeted defect scenario data.

17. A system (1) for enhancing a discriminative model based on a generative model, the system being configured to implement the method according to any one of claims 1 to 16, the system comprising: a data acquisition module (10), configured to acquire raw data samples of a target object; a discriminant model training module (20), configured to perform preliminary training on the discriminant model using raw data samples; A defect analysis module (30) configured to perform defect analysis on raw data samples based on the preliminary training results of the discriminant model to obtain a defect scenario data list; as well as A generative model training module (40) is configured to train a generative model with the aid of a list of defect scenario data, so as to directionally generate defect scenario data through the generative model; The discriminant model training module is further configured to be able to retrain the discriminant model based at least on the defect scenario data.

18. A machine-readable program carrier having stored thereon a computer program for executing the method according to any one of claims 1 to 16 when the computer program is run on a computer.

Images