CN118072101A - Noisy label image learning method and system based on balanced selection and contrastive learning - Google Patents

Abstract

The present invention provides a method and system for learning noisy labeled images based on balanced selection and contrastive learning, which relates to the fields of artificial intelligence and computer vision. The learning method includes the following steps: using a joint loss function, based on the original data set samples, pre-training a DNN model numbered m = {1,2} for several rounds; in the historical sequence S _m , in chronological order, marking the index of the sample containing the clean label as True, marking the other indexes as False, and putting the samples marked as True into the subset Remove labels from the remaining samples and put them into the subset Subset-based Subset Robust training _is performed on the DNN model numbered m = {1,2}; after all rounds of pre-training are completed, the last consecutive Among the results, select the sequence with the largest number of samples marked as True, and put the samples marked as True in the sequence into the benchmark set D _c ; reinitialize the DNN model numbered m = {1, 2}, and repeat the above steps until the preset total number of training rounds is reached.

Description

Noisy label image learning method and system based on balanced selection and contrastive learning

Technical Field

The present invention relates to the fields of artificial intelligence and computer vision, and in particular to a noisy label image learning method and system based on balanced selection and contrast learning.

Background technique

In the past few decades, deep neural networks (DNNs) have achieved great success in the field of image processing. Its development has been influenced by large-scale well-annotated datasets. However, the collection of such datasets requires huge manpower and material resources, which hinders the further application of DNNs. Weakly supervised and semi-supervised learning have received great attention due to their low requirements on the quality of model annotation. The datasets they use usually contain a large number of inaccurate labels (noise labels), and traditional training methods make the model prone to overfitting the noisy labels, resulting in a decrease in test performance and generalization performance.

There are existing technologies that use two models to alternately select image samples with clean labels for each other for training based on the small loss criterion, but their performance is poor.

In addition, there is an existing technology that introduces a mixed Gaussian model to model based on sample training loss to identify noise label samples, and then uses a semi-supervised method to train and achieve certain results, but this method performs poorly on data sets with high noise ratios.

There is also an existing technology that first collects a clean benchmark set with a smaller size but lower noise ratio through a sliding window method, and then optimizes the performance of the model based on the benchmark set in the second stage. It has made certain progress, but it ignores the category imbalance problem of the benchmark set, resulting in a decrease in the performance of DNN in asymmetric noise scenarios.

Summary of the invention

Purpose of the invention: The present invention proposes a noisy label image learning method and system based on balanced selection and contrastive learning, aiming to effectively solve the above-mentioned problems existing in the prior art.

First, a noisy label image learning method based on balanced selection and contrastive learning is proposed. The steps are as follows:

S1. Using the joint loss function, the DNN model numbered m = {1, 2} is pre-trained for several rounds based on the original dataset samples.

In the historical sequence _Sm , in chronological order, the index containing the clean label sample is marked as True, and the other indexes are marked as False. The samples marked as True are put into the subset Remove labels from the remaining samples and put them into the subset/>

S2. Based on the subset Subset/> Perform robust training on the DNN model numbered m={1,2};

S3, after all rounds of pre _- training are completed, the last consecutive Among the results, select the sequence with the largest number of samples marked as True, and put the samples marked as True in the sequence into the benchmark set D _c ;

S4. Reinitialize the DNN model numbered m={1,2} and repeat steps S1 to S3 until the preset total number of training rounds is reached.

In a further embodiment of the first aspect, the original data set sample in step S1 The expression is as follows:

In the formula, n is the number of samples in the data set; is the observed label of image _xi , and c represents the number of categories contained in the dataset.

In a further embodiment of the first aspect, the expression of the joint loss function L _joint in step S1 is as follows:

In the formula, _xi represents the input image; n represents the number of samples in the dataset; and p( _xi ) represents the predicted output value of the input feature _xi .

In a further embodiment of the first aspect, performing several rounds of pre-training on the DNN model numbered m={1,2} further includes:

Calculate the JSD loss d _{i of} all samples in the initial noise dataset under the DNN model:

Wherein, p ¹ ( _xi ) represents the output value of the DNN model with number m=1 based on the image _x ; p ² ( _xi ) represents the output value of the DNN model with number m=2 based on the image _x ;

KL(*) represents the Kullback-Leibler function, as follows:

Where m and n represent input parameters.

In a further embodiment of the first aspect, samples marked as True are placed into a subset The subset/> The expression is as follows:

Where, _xi represents the input image; represents the observed label of the input image x _i ; S _m represents the historical sequence; /> represents the weight value of the i-th sample at the t-th epoch; e=t-K+1 represents the t-K+1-th epoch;/> Indicates the weight of the i-th sample in the e-th epoch; /> Indicates the weight of the i-th sample in the j-th epoch; /> Represents the weight vector generated by all samples based on the m-th network at the tj-th epoch.

In a further embodiment of the first aspect, the historical sequence

In the formula, For the dataset/> The labeling results of all samples in the tth epoch based on the model numbered m.

In a further embodiment of the first aspect, in step S1, for the model numbered m=1:

Based on the JSD loss d _i , the samples contained in each category are sorted according to their corresponding divergence values, and the same number of samples are selected as samples with clean labels. In the historical sequence S ₁ , the indexes of the samples with clean labels are marked as True and the other indexes are marked as False in chronological order;

Initializing a Collection and/> Empty, select samples in sequence _S1 that are marked as True for K consecutive epochs and put them into the clean sample set/> Remove the labels from the remaining samples and place them in />

In a further embodiment of the first aspect, in step S1, for the model numbered m=2:

Based on the JSD loss d _i , the samples contained in each category are sorted according to their corresponding divergence values, and the same number of samples are selected as samples with clean labels. In the historical sequence S ₂ , the indexes of the samples with clean labels are marked as True and the other indexes are marked as False in chronological order;

Initializing a Collection and/> is empty, select samples in sequence S ₂ that are marked as True for K consecutive epochs and put them into the clean sample set/> Remove the labels from the remaining samples and place them in />

A second aspect of the present invention provides a noisy label image learning system, which includes a pre-training module, a robust training module, a selection module, and a repeated execution module.

The pre-training module uses the joint loss function to perform several rounds of pre-training on the DNN model numbered m = {1,2} based on the original dataset samples; in the historical sequence S _m , the index of the sample containing the clean label is marked as True, and the other indexes are marked as False in chronological order, and the samples marked as True are put into the subset Remove labels from the remaining samples and put them into the subset/>

The robust training module is based on the subset Subset/> Robust training is performed on the DNN model numbered m={1,2}.

The selection module is used to select the last _consecutive Among the results, the sequence with the largest number of samples marked as True is selected, and the samples marked as True in the sequence are put into the reference set D _c .

The repeated execution module is used to reinitialize the DNN model numbered m={1,2}, and feed back to the pre-training module, the robust training module, and the selection module until the preset total number of training rounds is reached.

According to a third aspect of the present invention, a computer-readable storage medium is provided, wherein at least one executable instruction is stored in the storage medium. When the executable instruction is executed on an electronic device, the electronic device executes the noisy label image learning method based on balanced selection and contrast learning as described in the first aspect.

Beneficial effects: The present invention proposes a noisy label image learning method and system based on balanced selection and contrastive learning. The method proposes a new balanced selection strategy to collect a clean subset with balanced categories and extremely low noise ratio, and then uses contrastive learning technology to further improve the model feature extraction ability and test performance, so that the model is suitable for various complex noise scenes, such as symmetrical noise, asymmetrical noise, instance-related noise and mixed noise. Compared with the common existing technology, the noisy label image learning method proposed by the present invention has a significant improvement in accuracy.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 shows the overall process of the noisy label image learning method based on balanced selection and contrastive learning.

Figure 2 shows the specific process of equilibrium selection.

Detailed ways

In the following description, a large number of specific details are provided to provide a more thorough understanding of the present invention. However, it is apparent to those skilled in the art that the present invention can be implemented without one or more of these details. In other examples, in order to avoid confusion with the present invention, some technical features well known in the art are not described.

Embodiment 1:

In view of the shortcomings of the prior art, the present invention proposes a noisy label image learning method of balanced selection and contrastive learning. The method proposes a new balanced selection strategy to collect a clean subset with balanced categories and extremely low noise ratio, and then uses contrastive learning technology to further improve the model feature extraction ability and test performance, so that the model is suitable for various complex noise scenes, such as symmetrical noise, asymmetrical noise, instance-related noise and mixed noise. In order to achieve the above purpose, the technical methods adopted by the present invention are as follows:

Step 1.1, for two DNN models with the same structure but different initialization parameters (model number is m = {1, 2}), use the joint loss function to calculate the given noisy label data set Perform warm-up training for a total of T _w epochs. After each epoch, calculate the Jensen–Shannon divergence of all training samples on the two models, and use the small loss criterion to sort the samples contained in each category according to their corresponding divergence values, and select the same number of samples as samples with clean labels. In the historical sequences S ₁ and S ₂ , mark the indexes of samples with clean labels as True values and other indexes as False values in chronological order;

Step 1.2: For the model with number m=1, at the beginning of each epoch, first calculate the Jensen–Shannon divergence of all training samples on the two models, and use the small loss criterion to sort the samples contained in each category according to their corresponding divergence values, and select the same number of samples as samples with clean labels. In the historical sequence _S1 , in chronological order, mark the index of the sample with clean label as True value, and mark other indexes as False value;

Then, initialize the collection and/> Empty, select samples in sequence _S1 that are marked as True for K consecutive epochs and put them into the clean sample set/> Remove the labels from the remaining samples and place them in />

Step 1.3: Use SSL technology and contrastive learning technology to divide the two subsets based on Step 1.2. and Perform robust training on the model numbered m=1.

Step 1.4, for the model numbered m = 2, at the beginning of each epoch, first calculate the Jensen–Shannon divergence of all training samples on the two models, and use the small loss criterion to sort the samples contained in each category according to their corresponding divergence values, and select the same number of samples as samples with clean labels. In the historical sequence S ₂ , in chronological order, mark the index of the sample with clean labels as True value, and mark other indexes as False value; then, initialize the set and/> is empty, select samples in sequence S ₂ that are marked as True for K consecutive epochs and put them into the clean sample set/> Remove the labels from the remaining samples and place them in />

Step 1.5: Use SSL technology and contrastive learning technology to divide the two subsets based on Step 1.2 and Robust training is performed on the model numbered m=2.

Step 1.6: If the preset total number of training rounds T _tot is not reached, jump to step 1.2 to continue training.

Step 1.7 After all the training in step ₁ is completed, the last continuous Among the results, the sequence with the largest number of samples marked as True is selected, and the samples marked as True in the sequence are put into the reference set D _c .

Step 1.8, re-initialize the DNN model with model number m = {1,2}, and also use the joint loss function based on the given noisy label dataset Perform warm-up training for a total of T _w epochs.

Step 1.9, for the model numbered m=1, at the beginning of each epoch, first calculate the Jensen–Shannon divergence of all training samples on the two models, and use the small loss criterion to sort the samples contained in each category according to their corresponding divergence values, and select the same number of samples as samples with clean labels, and mark their indexes as True values. In addition, for the samples included in the benchmark set D _c , their indexes are also marked as True, and the indexes of all remaining samples are marked as False values; then, initialize the set and/> Empty, the samples marked as True are put into the clean sample set/> The remaining False samples are removed from the label and put into />

Step 1.10: Use SSL technology and contrastive learning technology to divide the two subsets based on Step 1.9. and Perform robust training on the model numbered m=1.

Step 1.11. For the model with number m=2, at the beginning of each epoch, first calculate the Jensen–Shannon divergence of all training samples on the two models, and use the small loss criterion to sort the samples contained in each category according to their corresponding divergence values, and select the same number of samples as samples with clean labels, and mark their indexes as True values. In addition, for the samples included in the benchmark set D _c , their indexes are also marked as True, and the indexes of all remaining samples are marked as False values; then, initialize the set and/> Empty, the samples marked as True are put into the clean sample set/> The remaining False samples are removed from the label and put into />

Step 1.12: Use SSL technology and contrastive learning technology to divide the two subsets based on Step 1.11. and Robust training is performed on the model numbered m=2.

Step 1.13 If the preset total number of training rounds T _tot is not reached, jump to step 1.9 to continue training.

Embodiment 2:

Based on Example 1, the overall learning process of the present invention is shown with reference to FIG1. The specific process of the balanced selection in the process of collecting the benchmark set D _c proposed by the present invention is shown with reference to FIG2.

Assume the original data set is Where n is the number of samples in the data set,/> is the observed label of image x _i , which has a certain probability of not being equal to the true label of the image, and c represents the number of categories contained in the dataset. The feature extractor and classifier are denoted as h(·; φ _m ) and f(·; θ _m ), respectively, where m∈{1,2} is the model number, and φ _m and θ _m represent the parameters of the model numbered m. For the input image x _i , the output prediction value of the model numbered m can be p(x _i )=f(h(x _i ；φ _m ）；θ _m ).

Step 1.1, using the joint loss function of formula (1) based on Pre-train two DNN models (model number is m={1,2}) with the same structure but different initialization parameters, and the total number of epochs is _Tw ;

Then, the JSD loss of all samples in the initial noise data set under the DNN model numbered m=1 is calculated according to formula (2);

Here, p ^m ( _xi ) = h(f( _xi ; _θm ); _φm ) is the output value of the DNN model numbered m based on the image _xi . KL(·) represents the Kullback-Leibler function, as shown in Formula 3.

Based on the calculated JSD value, we sort the samples of each category in descending order. For example, for the samples contained in the jth category (j∈{1,2,…,c}), the sorted sequence is expressed as Where sort represents the sorting function. Based on the small loss criterion, we select the first R samples and mark their indexes as True, where And d _ts can be calculated from formula (4):

Here represents the mean of the JSD loss of all samples, and /> represents the minimum value of JSD of all samples, τ and d _μ are two predefined hyperparameters. Assuming that the current epoch number is t, all samples that are not marked as True can be marked as False, and the marks of all samples are stored in the history sequence according to their index order/> Among them/> For the dataset/> The labeling results of all samples in the tth epoch based on the model numbered m.

Step 1.2: For the model with number m=1, at the beginning of the tth epoch, first calculate the JSD loss of all training samples on the two models according to formula (2). Also use the small loss criterion, as in step 1.1, sort the samples contained in each category according to their corresponding divergence values, and select R samples as samples with clean labels according to formula (5), and mark them as True, and mark the remaining samples as False. Then store the marking results in the history sequence _S1 . Finally, initialize the set and/> is empty, as shown in formula (6), select samples in sequence _S1 that are marked as True for K consecutive epochs and put them into the clean sample set/> Remove the labels from the remaining samples and place them in />

Where t represents the tth epoch.

Step 1.3: Use SSL technology and contrastive learning technology to divide the two subsets based on Step 1.2. and Perform robust training on the model numbered m=1. First, calculate the SSL loss _Lssl according to formula (7);

Among them, L _l is based on the clean label subset The classification loss of _LuL is based on the unlabeled subset/> The classification loss, L _reg is the regularization loss, which can make the prediction value of the classifier not biased towards a specific category. λ _u is a preset weight coefficient. In addition, the present invention introduces the SimCLR contrastive learning method, which includes a mapping head for each model m = {1, 2}/> Where/> It is the parameter of the mapping header. /> The high-dimensional features f( _xi ) extracted by the feature extractor f(·; _θm ) can be mapped to a low-dimensional space, so as to facilitate further distinction. Based on the mapped low-dimensional features, for/> For each sample in , the contrast loss needs to be calculated according to formula (8) to improve the feature extraction and clustering capabilities of the model.

in is the low-dimensional feature after mapping; and x _2i and x _2i-1 are the results of two strong data enhancement transformations of sample _xi . Finally, the overall loss of the model numbered m is shown in formula (9):

Step 1.4: For the model with number m=2, at the beginning of the tth epoch, first calculate the JSD loss of all training samples on the two models according to formula (2). Also use the small loss criterion, as in step 1.1, sort the samples contained in each category according to their corresponding divergence values, and select R samples as samples with clean labels according to formula (5), and mark them as True, and mark the remaining samples as False. Then store the marking results in the history sequence _S2 . Finally, initialize the set and/> is empty, and according to formula (6), the samples that have been marked as True for K consecutive epochs in the historical sequence S ₂ are put into the clean sample set/> Remove the labels from the remaining samples and place them in />

Step 1.5: Use SSL technology and contrastive learning technology to divide the two subsets based on Step 1.4. and Robust training is performed on the model numbered m=2, and the loss function is shown in formula (9).

Step 1.6: If the preset total number of training rounds T _tot is not reached, jump to step 1.2 to continue training.

Step 1.7 After all the training in step ₁ is completed, as shown in formula (10), the last continuous Among the results, the sequence with the largest number of samples marked as True is selected, and the samples marked as True in the sequence are put into the reference set D _c .

Step 1.8, re-initialize the DNN model with model number m = {1,2}, and use formula (1) based on the given noisy label dataset Perform warm-up training for a total of T _w epochs.

Step 1.9: For the model with number m=1, at the beginning of the tth epoch, first calculate the JSD loss of all training samples on the two models according to formula (2). Also using the small loss criterion, as in step 1.1, sort the samples contained in each category according to their corresponding divergence values, and select R samples as samples with clean labels according to formula (5) and mark them as True. In addition, for the samples included in the benchmark set D _c , their indexes are also marked as True, and the indexes of all remaining samples are marked as False values.

Finally, initialize the collection and/> is empty, and according to formula (11), the samples marked as True in the original data set are put into the clean sample set/> Remove the labels from the remaining samples and place them in />

Step 1.10: Use SSL technology and contrastive learning technology to divide the two subsets based on Step 1.9. and Robust training is performed on the model numbered m=1, and the loss function is shown in formula (9).

Step 1.11, for the model numbered m = 2, at the beginning of the tth epoch, first calculate the JSD loss of all training samples on the two models according to formula (2), and also use the small loss criterion, as in step 1.1, sort the samples contained in each category according to their corresponding divergence values, and select R samples as samples with clean labels according to formula (5) and mark them as True. In addition, for the samples included in the benchmark set D _c , their indexes are also marked as True, and the indexes of all remaining samples are marked as False values. Finally, initialize the set and/> is empty, and according to formula (11), the samples marked as True in the original data set are put into the clean sample set/> Remove the labels from the remaining samples and place them in />

Step 1.12: Use SSL technology and contrastive learning technology to divide the two subsets based on Step 1.11. and Robust training is performed on the model numbered m=2, and the loss function is shown in formula (9).

Step 1.13 If the preset total number of training rounds T _tot is not reached, jump to step 1.9 to continue training.

Embodiment 3:

This embodiment discloses a noisy label image learning system, which includes a pre-training module, a robust training module, a selection module, and a repeated execution module. The pre-training module uses a joint loss function to perform several rounds of pre-training on a DNN model numbered m={1,2} based on the original data set samples; in the historical sequence _Sm , the index of the sample containing the clean label is marked as True, and the other indexes are marked as False in chronological order, and the samples marked as True are put into the subset Remove labels from the remaining samples and put them into the subset/> The robust training module is based on the subset Subset/> Perform robust training on the DNN model numbered m = {1,2}. The selection module is used to select the last consecutive in the historical sequence S _m after all rounds of pre-training are completed. Among the results, the sequence with the largest number of samples marked as True is selected, and the samples marked as True in the sequence are put into the benchmark set D _c . The repeated execution module is used to reinitialize the DNN model numbered m = {1, 2}, and feed back to the pre-training module, the robust training module, and the selection module until the preset total number of training rounds is reached.

By using the noisy labeled image learning system, the noisy labeled image learning method based on balanced selection and contrast learning disclosed in the above-mentioned embodiment 1 can be implemented.

Embodiment 4:

This embodiment proposes a computer-readable storage medium, in which at least one executable instruction is stored. When the executable instruction is executed on an electronic device, the electronic device performs the operation of the noisy label image learning method based on balanced selection and contrast learning as described in the above embodiment. More specific examples of computer-readable storage media in the present disclosure may include, but are not limited to: an electrical connection with one or more wires, a portable computer 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. The computer-readable storage medium may include a data signal propagated in a baseband or as part of a carrier wave, in which a readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The readable signal 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.

Based on the technical solutions disclosed in the above embodiments, the noisy label image learning method based on balanced selection and contrastive learning proposed in the present invention uses a deep model with the same structure as the backbone network on the synthetic noisy label dataset CIFAR-10, and the gain effect is as follows:

1) In the 90% symmetric noise label scenario, the current most advanced methods DivideMix and LongReMix achieved 93.2% and 93.8% test accuracy respectively, while the present invention achieved 94.0% accuracy improvement. When the symmetric noise ratio increased to 92%, the test accuracy of DivideMix and LongReMix methods were 57.6% and 63.9% respectively, while the test accuracy of the present invention was 90.9%;

2) In the 49% asymmetric noise label scenario, the accuracy of the similar methods DivideMix and LongReMix is 83.7% and 87.7%, while the accuracy of the present invention is 92.0%, which is a significant improvement;

3) When in the residual (RN) semantic noise label scenario, the accuracies of DivideMix and LongReMix are 84.57% and 85.13% respectively, while the accuracy of the present invention is 92.04%, which is a significant improvement.

Finally, facing the real noise label dataset Animal-10N, the accuracy of the competing methods DivideMix and LongReMix are 84.5% and 87.2% respectively, and the accuracy of the present invention is 88.84%, an improvement of 1.2%.

As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes may be made to it in form and detail without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims (10)

1. A noise-containing label image learning method based on balanced selection and contrast learning is characterized by comprising the following steps:

s1, performing a plurality of rounds of pre-training on a DNN model with the number of m= {1,2} based on an original data set sample by utilizing a joint loss function;

In the historical sequence S _m, the index containing the clean label sample is marked as True, other indexes are marked as False according to the time sequence, and the samples marked as True are put into the subset Removing labels from the remaining samples and then placing the samples into subsets/>

S2, based on the subsetSubset/>Robust training is carried out on a DNN model with the number of m= {1,2 };

S3, after the pre-training of all rounds is finished, the training is finally continued in a history sequence S _m Selecting the group of sequences with the largest number of samples marked as True from the results, and placing the samples marked as True in the sequences into a benchmark set D _c;

s4, re-initializing a DNN model with the number of m= {1,2} and repeating the steps S1 to S3 until a preset training total round is reached.

2. The noisy label image learning method according to claim 1, characterized in that: the expression of the original dataset sample D in step S1 is as follows:

Wherein n is the number of data set samples; Is the observation tag of image x _i, c represents the number of categories contained in the dataset.

3. The method according to claim 1, wherein the expression of the joint loss function L _joint in step S1 is as follows:

Wherein x _i represents an input image; n represents the number of data set samples; p (x _i) represents the predicted output value of the input feature x _i.

4. The noisy label image learning method according to claim 1, wherein the DNN model numbered m= {1,2} is pre-trained for several rounds, further comprising:

Calculating JSD loss d _i of all samples in the initial noise dataset under the DNN model:

Where p ¹(x_i) represents the output value of the DNN model numbered m=1 based on the image x _i; p ²(x_i) represents the output value of the DNN model numbered m=2 based on the image x _i;

KL (x) represents the Kullback-Leibler function as follows:

Where m and n represent input parameters.

5. The noisy label image learning method of claim 1, wherein the sample marked True is placed in the subsetThe subset/>The expression of (2) is as follows:

wherein x _i represents an input image; An observation tag representing an input image x _i; s _m represents a history sequence; /(I) A weight value representing the ith sample at the t-th epoch; e=t-k+1 represents the t-k+1th epoch; /(I)Representing the weight of the ith sample at the ith epoch; /(I)Representing the weight of the ith sample at the jth epoch; /(I)Representing the weight vector generated by all samples based on the mth network at the t-j epoch.

6. The method of claim 5, wherein the history sequence S _m =

In the method, in the process of the invention,For dataset/>At the t-th epoch, based on the labeling results of the model numbered m.

7. The noisy label image learning method according to claim 6, wherein in step S1, for a model numbered m=1:

based on the JSD loss d _i, sorting samples contained in each category according to the corresponding divergence value, selecting the same number of samples as samples containing clean labels, and marking indexes containing the clean label samples as True and other indexes as False in a historical sequence S ₁ according to time sequence;

Initializing a collection And/>For null, samples with consecutive K epochs in sequence S ₁, each labeled True, were selected for placement into the clean sample set/>Residual sample put after removal of tag/>

8. The noisy label image learning method according to claim 6, wherein in step S1, for a model numbered m=2:

Based on the JSD loss d _i, sorting samples contained in each category according to the corresponding divergence value, selecting the same number of samples as samples containing clean labels, and marking indexes containing the clean label samples as True and other indexes as False in a historical sequence S ₂ according to time sequence;

Initializing a collection And/>For null, samples with consecutive K epochs in sequence S ₂, each labeled True, were selected for placement into the clean sample set/>Residual sample put after removal of tag/>

9. A noisy label image learning system, comprising:

A pre-training module; the pre-training module performs a plurality of rounds of pre-training on a DNN model with the number of m= {1,2} based on an original data set sample by utilizing a joint loss function; in the historical sequence S _m, the index containing the clean label sample is marked as True, other indexes are marked as False according to the time sequence, and the samples marked as True are put into the subset Removing labels from the remaining samples and then placing the samples into subsets/>

A robust training module; the robust training module is based on the subsetSubset/>Robust training is carried out on a DNN model with the number of m= {1,2 };

Selecting a module; the selecting module is used for finally and continuously selecting the historical sequence S _m after the pre-training of all rounds is finished Selecting the group of sequences with the largest number of samples marked as True from the results, and placing the samples marked as True in the sequences into a benchmark set D _c;

Repeating the executing module; the repeated execution module is used for reinitializing the DNN model with the number of m= {1,2} and feeding back to the pre-training module, the robust training module and the selection module until the preset training total round is reached.

10. A computer readable storage medium, wherein at least one executable instruction is stored in the storage medium, which when executed on an electronic device, causes the electronic device to perform the noise-containing label image learning method based on equalization selection and contrast learning as claimed in any one of claims 1 to 8.