Adversarial learning and training method, system and device for eliminating overfitting, and storage medium
By introducing an adversarial network into the adversarial training of the live recognition model, the problem of overfitting the live recognition model is solved, and the generalization ability and recognition effect of the model are improved.
Patent Information
- Application Number
- CN202411593893.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-08
- Publication Date
- 2025-02-28
AI Technical Summary
In the adversarial training of the live body recognition model, the overfitting phenomenon is serious, resulting in the model's performance on the test set and the generalization ability is weak.
In the process of introducing an adversarial network into the training of living models, the body identity is classified through the discriminator, and the generator generates identity-free samples. Through iterative training, the living models filter out the impact of identity and improve the recognition effect.
Through the introduction of adversarial networks, the training samples change targetedly according to the training situation, correct the deviations in the model training process, improve the training stability and success rate, and enhance the generalization ability of the model.
Smart Images
Figure CN119541059A_ABST
Abstract
Description
[0001] Technology Neighborhood
[0002] The present invention relates to the field of living body recognition technology, and in particular to an adversarial learning and training method, system, device and storage medium for eliminating overfitting. Background Art
[0003] In the adversarial training of liveness recognition models, overfitting is an issue that requires special attention. Overfitting refers to the phenomenon that the model performs well on the training set, but the performance on the test set (or in actual applications) drops significantly, which usually means that the model's generalization ability is weak.
[0004] In adversarial training of liveness detection models, overfitting may manifest as follows: as training progresses, the adversarial robustness of the model on the training set (i.e., the model's ability to identify adversarial samples) continues to increase, but the adversarial robustness on the test set begins to decrease. At the same time, the classification accuracy of both may also continue to increase, but the accuracy improvement on the test set may not be as significant as that on the training set.
[0005] The main reasons for overfitting are as follows:
[0006] Too high model complexity: When the complexity of the model is much higher than the complexity of the problem, the model may learn the noise and unnecessary features in the training data, resulting in overfitting.
[0007] Insufficient training data: If the training data is insufficient to support the complexity of the model, the model may overfit to the limited information in the training data and fail to generalize to new data.
[0008] Impure data: Noise, incorrect labels, or outliers contained in the training data may cause the model to learn incorrect features.
[0009] The disclosure of the above background technology content is only used to assist in understanding the inventive concept and technical solution of the present invention. It does not necessarily belong to the prior art of this patent application. In the absence of clear evidence that the above content has been disclosed on the filing date of this patent application, the above background technology should not be used to evaluate the novelty and creativity of the present application. Summary of the invention
[0010] To this end, the present invention introduces an adversarial network into the living model training process, classifies body markers according to the discriminator, and generates identity-free samples through a generator. Through iterative training, the living model filters out the influence of identity and obtains a more robust recognition effect.
[0011] In a first aspect, the present invention provides an adversarial learning and training method for eliminating overfitting, characterized by comprising:
[0012] Step S1: performing initial training on the living body model according to the first training set and obtaining a first classification result;
[0013] Step S2: using a discriminator to identify the first training set, and using a generator to generate a first non-identity sample according to the identity identification and the first classification result, and adding the first non-identity sample to the first training set to obtain a second training set;
[0014] Step S3: training the living model according to the second training set to obtain a second classification result;
[0015] Step S4: using the discriminator to enable the generator to generate a second identity-free sample according to the identity identifier and the second classification result, and adding the second identity-free sample to the second training set, and repeatedly training the living model until the proportion of the identity identifier in the second classification result meets a preset range.
[0016] Optionally, the adversarial learning and training method for eliminating overfitting is characterized in that the first training set contains both living subjects and prostheses under the same identity.
[0017] Optionally, the adversarial learning and training method for eliminating overfitting is characterized in that, when generating the first identity-free sample or the second identity-free sample, the identity information of the sample misclassified by the living model is removed, and the label is kept unchanged to remove the influence of the identity information on the living model.
[0018] Optionally, the adversarial learning and training method for eliminating overfitting is characterized in that step S2 comprises:
[0019] Step S21: using a discriminator to identify the first image in the first training set, marking the same person as the same identity identifier, and marking different people as different identity identifiers;
[0020] Step S22: for the erroneous samples in the first classification result, a generator is used to generate N first non-identity samples; when the discriminator identifies the first non-identity samples, the identity information cannot be identified; wherein N≥1;
[0021] Step S23: adding the generated first identity-free samples to the first training set to obtain a second training set.
[0022] Optionally, the adversarial learning and training method for eliminating overfitting is applicable to a living face model, and is characterized in that step S22 comprises:
[0023] Step S221: For the wrong samples in the first classification result, obtain the key areas of the face; wherein the key areas of the face include at least one of the glasses, the nose, and the mouth;
[0024] Step S222: using a generator to remove one of the plurality of key face regions, and respectively generate a plurality of samples;
[0025] Step S223: using the generator to modify or replace the key face area so that the discriminator cannot recognize the identity information, thereby obtaining multiple samples.
[0026] Optionally, the adversarial learning and training method for eliminating overfitting is applicable to a palm live model, and is characterized in that step S22 comprises:
[0027] Step S224: obtaining a palm or vein pattern for an erroneous sample in the first classification result;
[0028] Step S225: For the living body sample, the generator replaces at least part of the palm or vein pattern with other living body samples to generate multiple samples;
[0029] Step S226: For the prosthetic sample, the generator replaces at least part of the palm or vein pattern with other prosthetic samples to generate multiple samples.
[0030] Optionally, the adversarial learning and training method for eliminating overfitting is characterized by further comprising:
[0031] Step S227: For the prosthetic sample, the generator replaces at least part of the palm or vein pattern with the living sample to generate multiple samples.
[0032] In a second aspect, the present invention provides an adversarial learning and training system for eliminating overfitting, which is used to implement any of the adversarial learning and training methods for eliminating overfitting described above, and is characterized by comprising:
[0033] A first training module, used to perform initial training on the living body model according to a first training set and obtain a first classification result;
[0034] a first generating module, configured to use a discriminator to identify the first training set, and enable a generator to generate a first non-identity sample according to the identity identification and the first classification result, and add the first non-identity sample to the first training set to obtain a second training set;
[0035] A second training module, used for training the living body model according to the second training set to obtain a second classification result;
[0036] The second generating module is used to use the discriminator to enable the generator to generate a second identity-free sample according to the identity identifier and the second classification result, and add the second identity-free sample to the second training set, and repeatedly train the living model until the proportion of the identity identifier in the second classification result meets a preset range.
[0037] In a third aspect, the present invention provides an adversarial learning and training device for eliminating overfitting, characterized in that it includes:
[0038] processor;
[0039] a memory storing executable instructions of the processor;
[0040] Wherein, the processor is configured to execute the steps of any of the aforementioned adversarial learning and training methods for eliminating overfitting by executing the executable instructions.
[0041] In a fourth aspect, the present invention provides a computer-readable storage medium for storing a program, characterized in that when the program is executed, the steps of any of the aforementioned adversarial learning and training methods for eliminating overfitting are implemented.
[0042] Compared with the prior art, the present invention has the following beneficial effects:
[0043] The present invention introduces the adversarial network into the training of the living model, so that the training samples can be changed in a targeted manner according to the training situation. The deviations occurring during the training of the living model can be corrected in a timely manner, the stability of the training process can be improved, the success rate of the living model training can be improved, and the generalization ability of the training model can be improved.
[0044] The present invention utilizes an adversarial network to generate identity-free samples, so that identity information is removed in a targeted manner, thereby enabling a living model with good generalization ability to be trained with a small sample set. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without creative work. Other features, purposes and advantages of the present invention will become more obvious by reading the detailed description of the non-limiting embodiments with reference to the following drawings:
[0046] Figure 1 A flowchart of the steps of an adversarial learning and training method for eliminating overfitting in an embodiment of the present invention;
[0047] Figure 2 is a flow chart of steps for obtaining a second training set in an embodiment of the present invention;
[0048] Figure 3 This is a flow chart of steps for generating a first non-identity sample in an embodiment of the present invention;
[0049] Figure 4 This is another flowchart of steps for generating a first non-identity sample in an embodiment of the present invention;
[0050] Figure 5 This is another flowchart of steps for generating a first non-identity sample in an embodiment of the present invention;
[0051] Figure 6 A schematic diagram of the structure of an adversarial learning and training system for eliminating overfitting in an embodiment of the present invention;
[0052] Figure 7 is a schematic diagram of the structure of an adversarial learning and training device for eliminating overfitting in an embodiment of the present invention; and
[0053] Figure 8 Schematic diagram of the structure of a computer-readable storage medium in an embodiment of the present invention. DETAILED DESCRIPTION
[0054] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several variations and improvements may be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.
[0055] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein, for example, can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0056] An adversarial learning and training method for eliminating overfitting provided in an embodiment of the present invention is intended to solve the problems existing in the prior art.
[0057] The following specific embodiments are used to describe in detail the technical solutions of the present invention and how the technical solutions of the present application solve the above-mentioned technical problems. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present invention will be described below in conjunction with the accompanying drawings.
[0058] The present invention introduces an adversarial network into the living body model training process, classifies body markers according to the discriminator, and generates identity-free samples through a generator. Through iterative training, the living body model filters out the influence of identity and obtains a more robust recognition effect.
[0059] Figure 1 The present invention is a flowchart of the steps of an adversarial learning and training method for eliminating overfitting in an embodiment of the present invention.
[0060] like Figure 1 As shown, the steps of an adversarial learning and training method for eliminating overfitting in an embodiment of the present invention include:
[0061] Step S1: Perform initial training on the living body model according to the first training set and obtain a first classification result.
[0062] In this step, there are both live bodies and prostheses under the same identity in the first training set. This step is to allow the live model (usually a deep learning model, such as a neural network) to preliminarily learn and identify the data features in the first training set. The first training set should contain enough representative samples to ensure that the model can learn the basic features of the data. Use these samples to train the live model, and the training process includes steps such as forward propagation, loss calculation, back propagation, and parameter update. Use the model after initial training to predict the first training set to obtain the first classification result. These results will be used for evaluation and generation of adversarial samples in subsequent steps. It should be noted that in this step, it is not necessary to train the live model until the parameters converge, that is, the live model in this step is not in the state after the training is completed. This embodiment is carried out during the training process of the live model, and the sample set is changed in a targeted manner according to the learning effect of each round of the live model during the training process to improve the training efficiency, rather than the evaluation after the model training.
[0063] Step S2: Use a discriminator to identify the first training set, and use a generator to generate a first non-identity sample based on the identity identification and the first classification result, and add the first non-identity sample to the first training set to obtain a second training set.
[0064] In this step, when generating the first identity-free samples, the identity information of the samples misclassified by the living model is removed, and the labels are kept unchanged to remove the influence of the identity information on the living model. The samples in the first training set are identified by a discriminator. The discriminator is used to classify the samples according to the identity information to distinguish the identity information of the samples. Based on the identity identification and the first classification result, a generator is used to generate the first identity-free samples. These samples remove specific identity identification information while maintaining the basic characteristics of the data. The generated first identity-free samples are added to the first training set to obtain a second training set. In this way, the second training set contains the original samples and the identity-free samples generated based on the misclassified samples, thereby specifically removing the influence of the identity information, which can greatly reduce the influence of the identity information on the judgment of the living model.
[0065] Step S3: training the living body model according to the second training set to obtain a second classification result.
[0066] In this step, the second training set is used to train the in vivo model. The training process is similar to step S1, including steps such as forward propagation, loss calculation, back propagation, and parameter update. Compared with step S1, the training set in this step is different, and the parameters of the in vivo model obtained by training are also different.
[0067] Step S4: using the discriminator to enable the generator to generate a second identity-free sample according to the identity identifier and the second classification result, and adding the second identity-free sample to the second training set, and repeatedly training the living model until the proportion of the identity identifier in the second classification result meets a preset range.
[0068] In this step, when generating the second identity-free sample, the identity information of the sample misclassified by the living model is removed, and the label is kept unchanged to remove the influence of the identity information on the living model. This step is to continuously optimize the performance of the living model by iteratively generating and adding identity-free samples until the preset generalization ability standard is reached. The discriminator is used to generate the second identity-free sample using the generator according to the identity identifier and the second classification result. These samples further remove the identity identifier information while maintaining the basic characteristics of the data. The generated second identity-free sample is added to the second training set to obtain an updated training set. Repeat steps S3 and S4, continuously iteratively train the living model, and generate new identity-free samples. In each iteration, the updated training set is used to train the living model and evaluate its performance on the test set. When the proportion of the identity identifier in the second classification result meets the preset range (that is, the performance of the model on the test set meets the preset standard), the iterative training is stopped. At this point, the living model already has a good generalization ability and can effectively reduce the overfitting phenomenon.
[0069] Figure 2 FIG. 1 is a flow chart of steps for obtaining a second training set in an embodiment of the present invention. Figure 2 As shown, in an embodiment of the present invention, a step of obtaining a second training set includes:
[0070] Step S21: using a discriminator to identify the first image in the first training set, marking the same person with the same identity identifier and marking different people with different identity identifiers.
[0071] In this step, the purpose is to use the discriminator to identify the images in the first training set and assign a unique identity to each image. The discriminator is a trained deep learning classifier that can distinguish between images of different individuals. For each image in the first training set, the discriminator is used to identify the image. Based on the identification results, an identity is assigned to each image. For multiple images of the same person, the same identity is assigned; for images of different people, different identities are assigned.
[0072] Step S22: for the erroneous samples in the first classification result, a generator is used to generate N first non-identity samples; when the discriminator identifies the first non-identity samples, the identity information cannot be identified.
[0073] In this step, N≥1. This step is to generate adversarial samples without identity information for samples that are wrong in the first classification results of the living model. These samples will be used to increase the diversity of the training set and help the living model better learn the essential characteristics of the data instead of focusing too much on identity information. Filter out the wrong samples from the first classification results, that is, those samples that are misclassified by the living model. Use the generator model to generate N first identity-free samples based on these wrong samples. The generator is usually a deep learning model that can generate adversarial samples that are similar to real samples. When generating identity-free samples, it is necessary to ensure that these samples remove or blur the identity information while retaining the basic characteristics of the data. This can be achieved by adding specific regularization terms to the loss function of the generator. The generated first identity-free samples are verified to ensure that the discriminator cannot recognize the identity information in them. This can be achieved by inputting the sample into the discriminator and checking whether its output meets certain threshold conditions.
[0074] Step S23: adding the generated first identity-free samples to the first training set to obtain a second training set.
[0075] In this step, the generated first identity-free samples are merged with the original samples in the first training set to form a new training set, namely the second training set. During the merging process, it is necessary to ensure that each sample retains its corresponding label so that it can be used in the subsequent training process. After the second training set is prepared, it can be used to train the living model in the subsequent steps.
[0076] Through this embodiment, a second training set containing more samples and being more challenging is obtained, which will help the living model to improve its generalization ability and reduce overfitting in the subsequent training process.
[0077] Figure 3 FIG. 1 is a flow chart of steps for generating a first non-identity sample in an embodiment of the present invention. Figure 3 As shown, this embodiment is applicable to the training of a live face model. In the embodiment of the present invention, a step of generating a first non-identity sample includes:
[0078] Step S221: For the erroneous samples in the first classification result, obtain the key face area.
[0079] In this step, the key areas of the face include at least one of the glasses, nose, and mouth. Filter out the wrong samples from the first classification results. Use face detection and feature point positioning algorithms (such as Dlib, MTCNN, etc.) to locate the key areas of the face. These algorithms can automatically detect and mark the key parts of the face such as eyes, nose, mouth, etc. These key areas are extracted and prepared for processing in subsequent steps.
[0080] Step S222: using a generator to remove one of the plurality of key face regions, and respectively generate a plurality of samples.
[0081] In this step, samples without identity information are generated by removing one or more of the key areas of the face, so that the living model can reduce its reliance on identity information during training, thereby improving its generalization ability. For each misclassified sample, a generator model (such as GAN) is used to remove one or more extracted key areas. When removing the key areas, it is necessary to ensure that the generated samples still visually maintain the basic structure and features of the face, but the identity information has been effectively removed or blurred. For each misclassified sample, the above operation is performed to generate multiple samples without identity information.
[0082] Step S223: using the generator to modify or replace the key face area so that the discriminator cannot recognize the identity information, thereby obtaining multiple samples.
[0083] In this step, samples without identity information are further generated by modifying or replacing the key areas of the face. Unlike step S222, here the key areas are not simply removed, but modified or replaced to eliminate the influence of the identity information. For each misclassified sample, the generator model is used to modify or replace its key areas. This can be achieved by generating a new key area image and fusing it into the original face image. When modifying or replacing the key area, it is necessary to ensure that the generated sample still visually maintains the coherence and authenticity of the face, while the identity information has been completely removed or replaced with meaningless information. For each misclassified sample, the above operation is performed to generate more samples without identity information. After the samples are generated, they are verified using a discriminator to ensure that the identity information has been effectively removed, that is, the discriminator cannot recognize the identity information in the sample. This can be achieved by inputting the sample into the discriminator and checking whether its output meets a certain threshold condition.
[0084] This embodiment helps to improve the generalization ability of the model and reduce the occurrence of overfitting.
[0085] Figure 4 FIG. 1 is another flow chart of steps for generating a first non-identity sample in an embodiment of the present invention. Figure 4 As shown, this embodiment is applicable to the training of a palm live model. In the embodiment of the present invention, another step of generating a first non-identity sample includes:
[0086] Step S224: For the erroneous samples in the first classification result, a palm or vein pattern is obtained.
[0087] In this step, palm or vein patterns are extracted from the misclassified samples. These patterns are key features in palm liveness recognition. By replacing or modifying in subsequent steps, samples for adversarial training can be generated. First, the wrong samples are screened out from the first classification results. These samples may be live objects that are mistakenly identified as prostheses, or prostheses that are mistakenly identified as live objects. Use image processing techniques (such as image segmentation, feature extraction, etc.) to extract palm or vein patterns from the misclassified samples. This can be achieved using deep learning models or traditional image processing algorithms.
[0088] Step S225: For the living body sample, the generator replaces at least part of the palm or vein pattern with other living body samples to generate multiple samples.
[0089] In this step, multiple new adversarial samples are generated by replacing the palm or vein patterns in the live samples. These samples will be used to train the palm live model to improve its recognition and generalization capabilities of live features. For each misclassified live sample, a generator model (such as GAN) is used to select part of the palm or vein pattern from other live samples for replacement. When replacing, it is necessary to ensure that the generated sample still visually maintains the basic structure and features of the palm while introducing the diversity of other live samples. Repeat this process to generate multiple adversarial samples for each misclassified live sample.
[0090] Step S226: For the prosthetic sample, the generator replaces at least part of the palm or vein pattern with other prosthetic samples to generate multiple samples.
[0091] In this step, multiple new adversarial samples are generated by replacing the palm or vein patterns in the prosthetic samples. These samples will be used to improve the recognition and generalization capabilities of the palm live model for prosthetic features. For each misclassified prosthetic sample, the generator model is used to select part of the palm or vein pattern from other prosthetic samples for replacement. When replacing, it is also necessary to ensure that the generated samples remain visually consistent and introduce the diversity of other prosthetic samples. Repeat this process to generate multiple adversarial samples for each misclassified prosthetic sample.
[0092] Figure 5 FIG. 1 is another flow chart of steps for generating a first non-identity sample in an embodiment of the present invention. Figure 5 As shown, compared with the above-mentioned embodiment, another step of generating a first non-identity sample in the embodiment of the present invention further includes:
[0093] Step S227: For the prosthetic sample, the generator replaces at least part of the palm or vein pattern with the living sample to generate multiple samples.
[0094] In this step, parts of the palm or vein pattern of the live sample are used to replace the corresponding parts in the prosthetic sample to generate multiple new adversarial samples. These samples will be used to train the palm live model to better identify those prosthetic samples that may be disguised as live. For each misclassified prosthetic sample, the generator model is used to select part of the palm or vein pattern from the live sample library for replacement. During the replacement process, it is necessary to ensure that the generated sample has a certain visual authenticity, that is, the replaced part maintains a certain coordination with the other parts of the prosthetic sample in terms of texture, color, etc. At the same time, the replaced part should contain enough information to simulate the live features so that the model can learn the differences between these features. Repeat this process to generate multiple adversarial samples for each misclassified prosthetic sample.
[0095] The adversarial samples generated by steps S224 to S227 can greatly enrich the training set of the palm live model, thereby improving the model's ability to identify live and prosthetic samples. In particular, the introduction of step S227 enables the model to better learn the subtle differences between live and prosthetic samples, reducing misjudgments caused by camouflage.
[0096] Figure 6 Schematic diagram of the structure of an adversarial learning and training system for eliminating overfitting in an embodiment of the present invention.
[0097] like Figure 6 As shown, an adversarial learning and training system for eliminating overfitting in an embodiment of the present invention includes:
[0098] A first training module, used to perform initial training on the living body model according to a first training set and obtain a first classification result;
[0099] a first generating module, configured to use a discriminator to identify the first training set, and enable a generator to generate a first non-identity sample according to the identity identification and the first classification result, and add the first non-identity sample to the first training set to obtain a second training set;
[0100] A second training module, used for training the living model according to the second training set to obtain a second classification result;
[0101] The second generating module is used to use the discriminator to enable the generator to generate a second identity-free sample according to the identity identifier and the second classification result, and add the second identity-free sample to the second training set, and repeatedly train the living model until the proportion of the identity identifier in the second classification result meets a preset range.
[0102] This embodiment introduces an adversarial network into the living body model training process, classifies body markers according to the discriminator, and generates identity-free samples through the generator. Through iterative training, the living body model filters out the influence of identity and obtains a more robust recognition effect.
[0103] An embodiment of the present invention also provides an adversarial learning and training device for eliminating overfitting, comprising a processor, a memory, in which executable instructions of the processor are stored, wherein the processor is configured to execute the steps of an adversarial learning and training method for eliminating overfitting by executing the executable instructions.
[0104] As mentioned above, this embodiment introduces an adversarial network into the living body model training process, classifies body markers according to the discriminator, and generates identity-free samples through the generator. Through iterative training, the living body model filters out the influence of identity and obtains a more robust recognition effect.
[0105] It will be appreciated by those skilled in the art that various aspects of the present invention may be implemented as systems, methods or program products. Therefore, various aspects of the present invention may be specifically implemented in the following forms, namely: a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or a combination of hardware and software implementations, which may be collectively referred to herein as "circuits", "modules" or "platforms".
[0106] Figure 7 Schematic diagram of the structure of an adversarial learning and training device for eliminating overfitting in an embodiment of the present invention. Figure 7 The electronic device 600 according to this embodiment of the present invention is described. Figure 7 The electronic device 600 shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.
[0107] like Figure 7 As shown, the electronic device 600 is in the form of a general computing device. The components of the electronic device 600 may include, but are not limited to: at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different platform components (including the storage unit 620 and the processing unit 610), a display unit 640, etc.
[0108] The storage unit stores program codes, which can be executed by the processing unit 610, so that the processing unit 610 executes the steps of various exemplary embodiments of the present invention described in the above-mentioned adversarial learning and training method for eliminating overfitting. For example, the processing unit 610 can execute the following steps: Figure 1 Follow the steps shown in .
[0109] The storage unit 620 may include a readable medium in the form of a volatile storage unit, such as a random access memory unit (RAM) 6201 and / or a cache memory unit 6202 , and may further include a read-only memory unit (ROM) 6203 .
[0110] The storage unit 620 may also include a program / utility 6204 having a set (at least one) of program modules 6205, such program modules 6205 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a grid environment.
[0111] Bus 630 may represent one or more of several types of bus structures, including a memory unit bus or memory unit controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.
[0112] The electronic device 600 may also communicate with one or more external devices 700 (e.g., keyboards, pointing devices, Bluetooth devices, etc.), one or more devices that enable a user to interact with the electronic device 600, and / or any device that enables the electronic device 600 to communicate with one or more other computing devices (e.g., routers, modems, etc.). Such communication may be performed through an input / output (I / O) interface 650. Furthermore, the electronic device 600 may also communicate with one or more grids (e.g., a local area network (LAN), a wide area network (WAN), and / or a public grid, such as the Internet) through a grid adapter 660. The grid adapter 660 may communicate with other modules of the electronic device 600 through the bus 630. It should be understood that although Figure 7 Not shown, other hardware and / or software modules may be used in conjunction with electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage platforms.
[0113] In an embodiment of the present invention, a computer-readable storage medium is also provided for storing a program, and when the program is executed, the steps of an adversarial learning and training method for eliminating overfitting are implemented. In some possible implementations, various aspects of the present invention can also be implemented in the form of a program product, which includes a program code. When the program product is run on a terminal device, the program code is used to enable the terminal device to execute the steps of various exemplary embodiments of the present invention described in the above-mentioned adversarial learning and training method for eliminating overfitting in this specification.
[0114] As shown above, this embodiment introduces an adversarial network into the living body model training process, classifies body markers according to the discriminator, and generates identity-free samples through the generator. Through iterative training, the living body model filters out the influence of identity and obtains a more robust recognition effect.
[0115] Figure 8 Schematic diagram of the structure of a computer-readable storage medium in an embodiment of the present invention. Figure 8 As shown, a program product 800 for implementing the above method according to an embodiment of the present invention is described, which can adopt a portable compact disk read-only memory (CD-ROM) and include program code, and can be run on a terminal device, such as a personal computer. However, the program product of the present invention is not limited thereto. In this document, a readable storage medium can be any tangible medium containing or storing a program, which can be used by or in combination with an instruction execution system, an apparatus or a device.
[0116] The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.
[0117] Computer readable storage media may include data signals propagated in baseband or as part of a carrier wave, wherein readable program codes are carried. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The readable storage medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, device, or device. The program codes contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination thereof.
[0118] Program code for performing the operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages such as Java, C++, etc., and conventional procedural programming languages such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving a remote computing device, the remote computing device may be connected to the user computing device through any type of grid, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., via the Internet using an Internet service provider).
[0119] This embodiment introduces an adversarial network into the living body model training process, classifies body markers according to the discriminator, and generates identity-free samples through the generator. Through iterative training, the living body model filters out the influence of identity and obtains a more robust recognition effect.
[0120] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the embodiments can be referred to each other. The above description of the disclosed embodiments enables professionals and technicians in this field to implement or use the present invention. The various modifications to these embodiments will be obvious to professionals and technicians in this field, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown in this article, but will comply with the widest range consistent with the principles and novel features disclosed herein.
[0121] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.
Claims
1. A method for eliminating overfitting through adversarial learning and training, characterized in that: include: Step S1: performing initial training on the living body model according to the first training set and obtaining a first classification result; Step S2: using a discriminator to identify the first training set, and using a generator to generate a first non-identity sample according to the identity identification and the first classification result, and adding the first non-identity sample to the first training set to obtain a second training set; Step S3: training the living model according to the second training set to obtain a second classification result; Step S4: using the discriminator to enable the generator to generate a second identity-free sample according to the identity identifier and the second classification result, and adding the second identity-free sample to the second training set, and repeatedly training the living model until the proportion of the identity identifier in the second classification result meets a preset range.
2. The adversarial learning and training method for eliminating overfitting according to claim 1, characterized in that: The first training set contains both living subjects and prosthetic subjects under the same identity.
3. The adversarial learning and training method for eliminating overfitting according to claim 1, characterized in that: When generating the first identity-free sample or the second identity-free sample, the identity information of the sample misclassified by the living model is removed, and the label is kept unchanged, so as to remove the influence of the identity information on the living model.
4. The adversarial learning and training method for eliminating overfitting according to claim 1, characterized in that: Step S2 includes: Step S21: using a discriminator to identify the first image in the first training set, marking the same person as the same identity identifier, and marking different people as different identity identifiers; Step S22: for the erroneous samples in the first classification result, a generator is used to generate N first non-identity samples; when the discriminator identifies the first non-identity samples, the identity information cannot be identified; wherein N≥1; Step S23: adding the generated first identity-free samples to the first training set to obtain a second training set.
5. The adversarial learning and training method for eliminating overfitting according to claim 4 is applicable to a living face model, characterized in that: Step S22 includes: Step S221: For the wrong samples in the first classification result, obtain the key areas of the face; wherein the key areas of the face include at least one of the glasses, the nose, and the mouth; Step S222: using a generator to remove one of the plurality of key face regions, and respectively generate a plurality of samples; Step S223: using the generator to modify or replace the key face area so that the discriminator cannot recognize the identity information, thereby obtaining multiple samples.
6. The adversarial learning and training method for eliminating overfitting according to claim 4 is applicable to a palm live model, characterized in that: Step S22 includes: Step S224: obtaining a palm or vein pattern for an erroneous sample in the first classification result; Step S225: For the living body sample, the generator replaces at least part of the palm or vein pattern with other living body samples to generate multiple samples; Step S226: For the prosthetic sample, the generator replaces at least part of the palm or vein pattern with other prosthetic samples to generate multiple samples.
7. The adversarial learning and training method for eliminating overfitting according to claim 6, characterized in that: Also includes: Step S227: For the prosthetic sample, the generator replaces at least part of the palm or vein pattern with the living sample to generate multiple samples.
8. An adversarial learning and training system for eliminating overfitting, used to implement the adversarial learning and training method for eliminating overfitting as claimed in any one of claims 1 to 7, characterized in that: include: A first training module, used to perform initial training on the living body model according to a first training set and obtain a first classification result; a first generating module, configured to use a discriminator to identify the first training set, and enable a generator to generate a first non-identity sample according to the identity identification and the first classification result, and add the first non-identity sample to the first training set to obtain a second training set; A second training module, used for training the living body model according to the second training set to obtain a second classification result; The second generating module is used to use the discriminator to enable the generator to generate a second identity-free sample according to the identity identifier and the second classification result, and add the second identity-free sample to the second training set, and repeatedly train the living model until the proportion of the identity identifier in the second classification result meets a preset range.
9. An adversarial learning and training device for eliminating overfitting, characterized in that: include: processor; a memory storing executable instructions of the processor; Wherein, the processor is configured to execute the steps of the adversarial learning and training method for eliminating overfitting as described in any one of claims 1 to 7 by executing the executable instructions.
10. A computer-readable storage medium for storing a program, characterized in that: When the program is executed, the steps of the adversarial learning and training method for eliminating overfitting described in any one of claims 1 to 7 are implemented.