A noisy label correction method for unsupervised domain adaptation person re-identification

Abstract

The present invention provides a noise label correction method for unsupervised domain adaptive person re-identification, which relates to the field of unsupervised domain adaptive person re-identification. The noise label correction method for unsupervised domain adaptive person re-identification uses a source domain dataset and its corresponding labels to train a source model; uses a clustering-based unsupervised domain adaptive person re-identification method to fine-tune the source model, and uses the noise label correction framework proposed in this article to correct the noise labels of the fine-tuned domain adaptation model. Compared with previous noise label correction methods, this method uses a noise label corrector to process noise labels as variables, and updates the probability label variables through back propagation, gradually correcting the noise labels while updating the variables. This method can significantly improve model performance and is a plug-and-play noise correction method.

Description

Technical Field

[0001] The present invention relates to the technical field of unsupervised domain adaptive person re-identification, and in particular to a noise label correction method for unsupervised domain adaptive person re-identification. Background Art

[0002] The AGW method proposes a joint learning framework for high-precision neighborhood pseudo-labels and relatively high-recall group pseudo-labels. This framework leverages the mutually beneficial nature of complementary pseudo-labels to address the lack of hard samples in clustering-based person re-identification methods. It also introduces a similarity aggregation loss to rationally utilize group pseudo-labels and mitigate the impact of noise during the clustering process. The TCPC method proposes a two-stage clustering pseudo-label correction method. This method partitions the dataset into sub-camera domains based on camera ID and subjects them to the same clustering process to more accurately mine intra-camera connections. This process yields a list of reliable sub-pseudo-labels that can be used to correct global pseudo-labels. While these methods mitigate the negative impact of noisy labels inevitably generated during the clustering process, their effectiveness depends heavily on the quality of the feature representations used to generate the clusters, particularly the discriminability of the generated features. Consequently, they often result in limited performance improvements. Summary of the Invention

[0003] (1) Technical problems solved

[0004] In response to the shortcomings of the existing technology, the present invention provides a noise label correction method for unsupervised domain adaptive pedestrian re-identification, which can significantly improve model performance and is a plug-and-play noise correction method.

[0005] (2) Technical solution

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a noise label correction method for unsupervised domain adaptive pedestrian re-identification, comprising:

[0007] Use the source domain dataset to train the source model for cross-domain person re-identification and obtain the corresponding network model and weight file;

[0008] Load the source model and weight file to initialize the adaptive model, use the adaptive model to extract the target domain sample features, and use the target domain sample features to cluster to obtain the corresponding sample pseudo labels; use the target domain dataset with pseudo labels to train the adaptive model to obtain the corresponding network model and weight file;

[0009] Load the adaptive model and weight file to initialize the model to be corrected, and use the noise labels of the target domain samples predicted by the model to initialize the probability label variables in the noise label corrector;

[0010] Use the target domain training samples to update and maintain the probability label variables and model parameters through model parameter learning and label learning in a small batch manner, and use the update of the probability label variables to correct the noise to achieve noise label correction in the target domain until the network model converges and obtains the noise-corrected network model and weight file.

[0011] Preferably, the source model for cross-domain person re-identification is trained using a data set of the source domain to obtain a corresponding network model and weight file, and is specifically trained in a supervised manner using cross entropy loss and multiple similarity loss.

[0012] Preferably, the specific form of the cross entropy loss is expressed as:

[0013]

[0014] Among them, N s is the number of samples in the source domain dataset, Represented as a sample The model predicts the probability of Represents a sample Real identity label.

[0015] Preferably, the multiple similarity loss is iteratively implemented through two steps of sampling and weighting. In the sampling stage, information-rich sample pairs are selected by calculating the relative similarity of the sample pairs. In the sampling stage, information-rich sample pairs are roughly selected and pairs with poor information are discarded. In the weighting stage, the samples are weighted by jointly calculating their self-similarity and relative similarity. The multiple self-similarity loss is obtained by integrating the sampling and weighting stages. The multiple similarity loss function is expressed as:

[0016]

[0017] Among them, s ij = <g s,i ,g s,j > represents the similarity between two features, <.,.> represents the dot product; α, β, λ are all hyperparameters, and P represents the sample The positive example pair set, N i Representation sample The set of counterexample pairs;

[0018] The source model is optimized by minimizing the cross entropy loss and the multiple similarity loss. The total loss function of the source model is expressed as:

[0019] L src =L sce +αL ms .

[0020] Preferably, the adaptive model and weight file are loaded to initialize the model to be corrected, and the noise labels of the target domain samples predicted by the model are used to initialize the probability label variables in the noise label corrector; the target domain training samples are used to update and maintain the probability label variables and model parameters through model parameter learning and label learning in a small batch manner, and the noise is corrected by using the update of the probability label variables to achieve noise label correction in the target domain until the network model converges, thereby obtaining the noise-corrected network model and weight file. Specifically, it includes:

[0021] Given a target domain sample x, the attention wave block network acts as a feature extractor G to extract discriminative features The classifier outputs a vector via linear activation This value can be obtained by p=W T f is calculated, where is the weight matrix of the fully connected layer; the probability label of the target domain sample predicted by the model and the probability label variable y in the noisy label corrector a Can be respectively and y a =σ(p a ), where σ(·) is the soft-max activation function, and

[0022] Preferably, the The specific form of is expressed by the following formula;

[0023]

[0024] For the probability label variable y in the noise label corrector a , the target domain noisy label y predicted by the adaptive model fine-tuned by the clustering-based domain adaptation method n Initialize; the specific initialization formula is as follows:

[0025] y a =Hy n

[0026] Among them, H is a constant (H=10 is set during the experiment). In the initial stage of noise label correction, there is

[0027] In the noise label correction stage, for each target domain sample x i , the noise label corrector has a probability label variable variable is the sample x i The noise-free probability label estimate of can be Calculated; swapping the two operators in the cross entropy loss gives an inverse cross entropy loss; the inverse cross entropy loss function can be expressed by the following formula:

[0028]

[0029] The noise label correction stage can be achieved by the probability label variable y a and model predictions The difference between them is used to guide the update of the model parameters θ; the anti-cross entropy loss is used to calculate the classification loss; the classification loss in the noise label correction stage can be expressed as:

[0030]

[0031] y a Updates can be made in the following ways:

[0032]

[0033] Entropy loss is introduced, and the specific form of the entropy loss function can be expressed by the following formula:

[0034]

[0035] Combining the classification loss and the regularization term as the total loss of this stage, the overall optimization objective of the noise label correction stage can be expressed as:

[0036]

[0037] Where β is the equilibrium parameter, L id is the classification loss, L e is the entropy loss, y a is the probabilistic label variable in the noisy label corrector, and θ is a model parameter. The parameter β is related to the entropy loss, which acts as a regularization term to prevent the network from stalling. β is set to < 1. The network is constrained by the overall optimization objective and continuously iterates to correct the noisy labels.

[0038] In a second aspect, a computer-readable storage medium storing one or more programs is provided, wherein the one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods described.

[0039] According to a third aspect, a computing device is provided, including:

[0040] One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for executing any of the methods described.

[0041] (3) Beneficial effects

[0042] This paper proposes a method for correcting noisy labels for unsupervised domain-adaptive person re-identification. This framework requires no auxiliary clean datasets or prior knowledge about noise. Instead, it uses a noisy label corrector to treat the target domain noisy labels predicted by the model as variables and updates the variables through backpropagation. This framework is therefore highly user-friendly and robust.

[0043] The present invention performs noise label correction based on the noise labels predicted by the model for the first time, uses model parameter learning and label learning to update and maintain probability label variables and model parameters, and gradually corrects noise by updating the probability label variables.

[0044] The proposed noise label correction method can be used to correct noise labels for different adaptive models, and significantly improves model performance. This demonstrates that the method is a plug-and-play method with strong compatibility and versatility.

[0045] This paper proposes a new source domain pre-training method that uses multiple similarity losses instead of the traditional triplet loss and uses an attention wave block network instead of a ResNet50 network for source domain pre-training. Experimental results show that the performance of the proposed pre-training method is significantly better than other methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a flow chart of the method of the present invention;

[0047] Figure 2 This is a model architecture diagram of the source domain pre-training stage of the present invention;

[0048] Figure 3 2 is a model architecture diagram of the noise label correction stage of the present invention. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] Example

[0051] Reference Figure 1 The present invention proposes a noise label correction method for unsupervised domain adaptation person re-identification, comprising:

[0052] Use the source domain dataset to train the source model for cross-domain person re-identification and obtain the corresponding network model and weight file;

[0053] Load the source model and weight file to initialize the adaptive model, use the adaptive model to extract the target domain sample features, and use the target domain sample features to cluster to obtain the corresponding sample pseudo labels; use the target domain dataset with pseudo labels to train the adaptive model to obtain the corresponding network model and weight file;

[0054] Load the adaptive model and weight file to initialize the model to be corrected, and use the noise labels of the target domain samples predicted by the model to initialize the probability label variables in the noise label corrector;

[0055] Use the target domain training samples to update and maintain the probability label variables and model parameters through model parameter learning and label learning in a small batch manner, and use the update of the probability label variables to correct the noise to achieve noise label correction in the target domain until the network model converges and obtains the noise-corrected network model and weight file.

[0056] Specifically, the method consists of three stages: (1) source domain pre-training (2) clustering-based domain adaptation (3) noise label correction. The first two stages are to pre-train the model with a labeled source domain dataset, and then fine-tune the source domain pre-trained model using target domain samples with pseudo labels generated based on clustering, thereby obtaining a model that has basic discriminability for target domain samples. The third stage uses the adaptive model of the second stage to initialize the model to be corrected, and uses the noise label corrector to correct the noise labels of the target domain predicted by the model, thereby improving the accuracy and reliability of the model prediction. The following article will describe the contents of each stage in detail.

[0057] Source domain pre-training:

[0058] Reference Figure 2 , for the cross-domain person re-identification task, there is a labeled source domain and an unlabeled target domain Since the identity information of the source domain dataset Ds is available, the training process of the source model can be regarded as a traditional classification problem. Traditional source models are usually trained in a supervised manner using cross entropy loss and triple loss. The specific form of cross entropy loss can be expressed as:

[0059]

[0060] where N s is the number of samples in the source domain dataset, Represented as a sample The model predicts the probability of Representation sample Real identity label.

[0061] The traditional triplet loss function is expressed as:

[0062]

[0063] Among them, ||.|| represents the Euclidean distance, and Respectively positive and negative samples.

[0064] Because the traditional triplet loss focuses primarily on relative similarity and considers all selected pairs equally, this limits its ability to identify pairs containing more information among the selected pairs, making it insufficient for pulling similar samples closer and rejecting dissimilar samples in the pedestrian re-identification task. To fully utilize the multiple similarities in the representation features of the selected pairs, this paper adopts the multiple similarity loss instead of the traditional triplet loss to optimize the model.

[0065] The multiple similarity loss is implemented iteratively through two steps: sampling and weighting. The sampling stage selects information-rich sample pairs by calculating the relative similarity of the sample pairs. That is, the negative example pairs are compared with the most difficult positive example pairs (with the smallest similarity), and the positive example pairs are compared with the easiest negative example pairs (with the largest similarity). Since the sampling stage roughly selects information-rich sample pairs and discards information-poor pairs, the weighting stage weights the samples by jointly calculating their self-similarity and relative similarity. By integrating the sampling and weighting stages, the multiple self-similarity loss is obtained. The multiple similarity loss function is expressed as:

[0066]

[0067] where s ij = <g s,i ,g s,j > represents the similarity between two features, <.,.> represents the dot product. α, β, λ are all hyperparameters, and P represents the sample The positive example pair set, N i Representation sample The set of counterexample pairs.

[0068] In order to extract more discriminative features, this paper uses the attention wave block network instead of the traditional RestNet50 network for feature extraction. The source model can be optimized by minimizing the cross entropy loss and the multiple similarity loss. Therefore, the total loss function of the source model can be expressed as:

[0069] L src =L sce +αL ms (4)

[0070] Clustering-based Domain Adaptation:

[0071] Due to the domain gap between the source and target domains, directly applying the source model obtained in the previous stage to the target domain can significantly degrade model performance. To address this issue, many clustering-based unsupervised domain adaptation person re-identification methods have been proposed. In our experiments, we used the AWB method and the Dual-Refinement method to fine-tune the source model, respectively, to enable it to better generalize to the target domain sample set.

[0072] The AWB method designs two combination strategies of attention mechanism and wave block (pre-attention and post-attention). This paper adopts the AWB method with post-attention strategy to fine-tune the source model, and adopts post-attention strategy to implement the attention wave block network architecture. Figure 2 As shown in Figure 2, the post-attention strategy places the attention mechanism after the Wake Block module. First, the Wake Block module performs different fluctuations on different regions, generating different discriminative features. Then, the attention mechanism learns the discriminative features of the fluctuating regions, effectively amplifying the different discriminative features.

[0073] Noisy label correction:

[0074] Reference Figure 3 The noisy label correction stage mainly uses the noisy label corrector to correct the target domain noisy pseudo-labels predicted by the model, thereby improving the model's discriminability for target domain samples and the accuracy of model predictions. The noisy label corrector processes the target domain noisy label predicted by the model as a variable and updates the variable through backpropagation to obtain the corrected noise-free label, which is then used to supervise model training. For this probabilistic label variable, we initialize it with the target domain noisy label predicted by the model fine-tuned using the clustering-based domain adaptation method. Next, we will describe the specific correction process of the noisy label correction stage in detail.

[0075] (1) Construction of probability label variables:

[0076] Given a target domain sample x, the attention wave block network acts as a feature extractor G to extract discriminative features The classifier then outputs the vector via a linear activation This value can be obtained by p=W T f is calculated, where is the weight matrix of the fully connected layer. The probability label of the target domain sample predicted by the model and the probability label variable y in the noisy label corrector a Can be respectively and y a =σ(p a ), where σ(·) is the soft-max activation function, and therefore, The specific form can be expressed by the following formula:

[0077]

[0078] For the probability label variable y in the noise label corrector a , we use a clustering-based adaptive method to fine-tune the adaptive model to predict the target domain noisy label y n Initialize. The specific initialization formula is as follows:

[0079] y a =Hy n (6)

[0080] Where H is a constant (H=10 is set during the experiment). In the initial stage of noise label correction,

[0081] (2) Optimization of probability label variables:

[0082] In the noise label correction stage, for each target domain sample x i , the noise label corrector has a probability label variable This variable is the sample x i The noise-free probability label estimate of can be Calculated. We swap the two operators in the cross entropy loss to obtain an inverse cross entropy loss. The inverse cross entropy loss function can be expressed as follows:

[0083]

[0084] The noise label correction stage can be achieved by the probability label variable y a and model predictions The difference between them guides the update of the model parameters θ. In previous classification losses, the classic cross-entropy loss is often used to calculate the distance between these two components. However, through experiments, we found that the cross-entropy loss is not very effective in the noisy label correction stage, and compared with the classic cross-entropy loss, the inverse cross-entropy loss is more suitable for correcting noisy labels. Therefore, the inverse cross-entropy loss is used here to calculate the classification loss. The classification loss in the noisy label correction stage can be expressed as:

[0085]

[0086] (3) Update of probability label variables:

[0087] The noise label correction stage not only updates the model parameters θ, but also updates the probability label variable y in the noise label correctora .y a You can update it in the following ways:

[0088]

[0089] (4) Enhanced discriminability of probabilistic label variables:

[0090] In the classification loss, we convert the probability label variable y in the noise label corrector a As the supervised value of the model parameter θ may lead to the model prediction Quickly approach the probability label variable y a , which causes the network to stop updating. To solve this problem, we introduce an additional regularization term (entropy loss). Entropy loss is the most commonly used loss in unsupervised domain adaptation. Minimizing this loss can make the model peak only on one identity class and form a weak decision boundary. The intra-class samples are more compact and the inter-class boundaries are more obvious. The specific form of the entropy loss function can be expressed by the following formula:

[0091]

[0092] (5) Overall optimization goal:

[0093] We combine the classification loss and the regularization term as the total loss of this stage. Therefore, the overall optimization objective of the noise label correction stage can be expressed as:

[0094]

[0095] Where β is the equilibrium parameter, L id is the classification loss, L e is the entropy loss, y a is the probabilistic label variable in the noisy label corrector, and θ is a model parameter. The parameter β is related to entropy loss, which acts as a regularization term to prevent the network from stalling. Excessively large entropy loss can lead to ineffective noisy label correction. Therefore, in our experiments, we set β < 1. The network, constrained by the overall optimization objective function, continuously iterates to ultimately achieve the goal of correcting noisy labels.

[0096] The experimental effects of the present invention are described below in conjunction with Table 1, Table 2 and Table 3. Table 1 shows the experimental results of the noise label correction of the AWB method on two pedestrian re-identification benchmark datasets, where the bold represents the optimal experimental results. It can be seen from Table 1 that the noise label correction method (NCL) proposed in the present invention can significantly improve the performance of the AWB model, and the improvement effect on most domain adaptation tasks is about 2.0%. Table 2 shows the experimental results of the noise label correction of the Dual-Refinement method. It can be seen from Table 2 that the present invention can also significantly improve the performance of the Dual-Refinement method. This fully demonstrates that the noise label correction method proposed in the present invention is a plug-and-play method with strong compatibility and versatility. Table 3 compares the source domain pre-training method proposed in the present invention with the existing more popular source domain pre-training methods. It can be seen from Table 3 that the source domain pre-training method proposed in the present invention is significantly superior to these representative source domain pre-training methods.

[0097] Table 1 Evaluation results of noise label correction for AWB model using NCL method

[0098]

[0099] Table 2 Evaluation results of the Dual-Refinement model using the NCL method for noise label correction

[0100]

[0101] Table 3 Performance comparison of different source domain pre-training methods on the MSMT17 dataset

[0102]

[0103] The embodiments of the present application can be provided as methods or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and translated scripting language JavaScript, etc.

[0104] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0105] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0106] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0107] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

Claims

1. A noise label correction method for unsupervised domain adaptation person re-identification, characterized by: include: Use the source domain dataset to train the source model for cross-domain person re-identification and obtain the corresponding network model and weight file; Load the source model and weight file to initialize the adaptive model, use the adaptive model to extract the target domain sample features, and use the target domain sample features to cluster to obtain the corresponding sample pseudo labels; use the target domain dataset with pseudo labels to train the adaptive model to obtain the corresponding network model and weight file; Load the adaptive model and weight file to initialize the model to be corrected, and use the noise labels of the target domain samples predicted by the model to initialize the probability label variables in the noise label corrector; Use the target domain training samples to update and maintain the probability label variables and model parameters through model parameter learning and label learning in a small batch manner. Use the update of the probability label variables to correct the noise and achieve noise label correction in the target domain until the network model converges. The noise-corrected network model and weight file are obtained. The source model for cross-domain person re-identification is trained using the source domain dataset to obtain the corresponding network model and weight file, specifically using cross entropy loss and multiple similarity loss for training in a supervised manner; The specific form of the cross entropy loss is expressed as: Among them, N s is the number of samples in the source domain dataset, Represented as a sample The model predicts the probability of Representation sample The real identity label; The multiple similarity loss is implemented iteratively through two steps: sampling and weighting. In the sampling stage, the relative similarity of the sample pairs is calculated to select the sample pairs with rich information. In the sampling stage, the sample pairs with rich information are roughly selected and the pairs with poor information are discarded. In the weighting stage, the samples are weighted by jointly calculating their self-similarity and relative similarity. The multiple self-similarity loss is obtained by integrating the sampling and weighting stages. The multiple similarity loss function is expressed as: Among them, s ij = <g s,i ,g s,j > represents the similarity between two features, <.,.> represents the dot product; α, β, λ are all hyperparameters, and P represents the sample The positive example pair set, N i Representation sample The set of counterexample pairs; The source model is optimized by minimizing the cross entropy loss and the multiple similarity loss. The total loss function of the source model is expressed as: L src =L sce +αL ms 。 2. The noise label correction method for unsupervised domain-adaptive person re-identification according to claim 1, characterized in that: The adaptive model and weight file are loaded to initialize the model to be corrected, and the noise labels of the target domain samples predicted by the model are used to initialize the probability label variables in the noise label corrector; the target domain training samples are used to update and maintain the probability label variables and model parameters through model parameter learning and label learning in a small batch manner, and the noise is corrected by using the update of the probability label variables to achieve noise label correction in the target domain until the network model converges, and the noise-corrected network model and weight file are obtained. Specifically, it includes: Given a target domain sample x, the attention wave block network acts as a feature extractor G to extract discriminative features The classifier outputs a vector via linear activation This value is given by p=W T f is calculated, where is the weight matrix of the fully connected layer; the probability label of the target domain sample predicted by the model and the probability label variable y in the noisy label corrector a Respectively by and y a =σ(p a ), where σ(·) is the soft-max activation function, and 3. The noise label correction method for unsupervised domain adaptive person re-identification according to claim 2, characterized in that: The specific form is expressed by the following formula: For the probability label variable y in the noise label corrector a , the target domain noisy label y predicted by the adaptive model fine-tuned by the clustering-based domain adaptation method n Initialize; the specific initialization formula is as follows: the a =Hy n Among them, H is a constant. During the experiment, H=10 is set. In the initial stage of noise label correction, In the noise label correction stage, for each target domain sample x i , the noise label corrector has a probability label variable variable is the sample x i The noise-free probability label estimate of Depend on Calculated; swapping the two operators in the cross entropy loss gives an inverse cross entropy loss; the inverse cross entropy loss function is expressed by the following formula: The noise label correction stage uses the probability label variable y a and model predictions The difference between them is used to guide the update of the model parameters θ; the anti-cross entropy loss is used to calculate the classification loss; the classification loss in the noise label correction stage is expressed as: y a Update via: Entropy loss is introduced, and the specific form of the entropy loss function is expressed by the following formula: Combining the classification loss and the regularization term as the total loss of this stage, the overall optimization objective of the noise label correction stage is expressed as: Where β is the equilibrium parameter, L id is the classification loss, L e is the entropy loss, y a is the probability label variable in the noise label corrector, θ is the model parameter; the parameter β is related to the entropy loss. The entropy loss is used as a regularization term to prevent the network from stopping updating. β<1 is set; the network is constrained by the total optimization objective function and continuously iterates to achieve the purpose of correcting the noise label.

4. A computer-readable storage medium storing one or more programs, characterized in that: The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any one of the methods according to claims 1-3.

5. A computing device, characterized in that include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs comprising instructions for executing any one of the methods according to claims 1-3.

Images