Noisy label image learning method and system based on balanced selection and contrastive learning
Through balanced selection and comparison learning technology, the noise label data set is processed, and the overfitting problem of deep neural networks on the noise label data set is solved, the performance of the model in noise scenarios is improved, and the accuracy and generalization are achieved. ability.
Patent Information
- Application Number
- CN202410281796.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-13
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-03-13
AI Technical Summary
The prior art deep neural networks are prone to overfitting when processing data sets containing large amounts of noise labels, resulting in degradation of test performance and generalization performance. The existing methods perform poorly on data sets with high noise ratios, ignoring categories The imbalance problem leads to a decrease in the performance of the model in asymmetric noise scenarios.
Using a method based on balanced selection and contrast learning, model pre-training and robust training is performed through joint loss function and Jensen–Shannon divergence, clean subsets with a category equalization and low noise ratio are selected, and the feature extraction of the model is improved by using contrast learning technology. Capability and test performance.
It significantly improves the accuracy of the model in complex noise scenarios, is suitable for scenarios such as symmetric noise, asymmetric noise, instance-related noise and mixed noise, and improves the generalization ability of the model.
Smart Images

Figure CN118072101B_ABST
Abstract
Description
Technical Field
[0001] 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 Art
[0002] 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 it easy for models to overfit noisy labels, resulting in a decrease in test performance and generalization performance.
[0003] 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.
[0004] In addition, there is an existing technology that introduces a mixed Gaussian model to model based on the training loss of samples 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.
[0005] 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
[0006] 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.
[0007] First, a noisy label image learning method based on balanced selection and contrastive learning is proposed. The steps are as follows:
[0008] 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.
[0009] In the historical sequence S m In the chronological order, the index containing the clean label sample is marked as True, and the other indexes are marked as False, and the samples marked as True are put into the subset Remove labels from the remaining samples and put them into the subset
[0010] S2. Based on the subset Subset Perform robust training on the DNN model numbered m={1,2};
[0011] S3. After all rounds of pre-training are completed, in the historical sequence S m 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 this sequence into the benchmark set D. c ;
[0012] 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.
[0013] In a further embodiment of the first aspect, the original data set sample in step S1 The expression is as follows:
[0014]
[0015] In the formula, n is the number of samples in the data set; is the image x i The observation labels are , and c represents the number of categories contained in the dataset.
[0016] In a further embodiment of the first aspect, the joint loss function L in step S1 joint The expression is as follows:
[0017]
[0018] In the formula, x i represents the input image; n represents the number of samples in the dataset; p(x i ) represents the input feature x i The predicted output value of .
[0019] In a further embodiment of the first aspect, performing several rounds of pre-training on the DNN model numbered m={1,2} further includes:
[0020] Calculate the JSD loss d of all samples in the initial noise dataset under the DNN model i :
[0021]
[0022] In the formula, p 1 (xi ) indicates that the DNN model with number m=1 is based on image x i The output value of p 2 (x i ) indicates that the DNN model numbered m=2 is based on the image x i The output value of
[0023] KL(*) represents the Kullback-Leibler function, as follows:
[0024]
[0025] In the formula, m and n represent input parameters.
[0026] In a further embodiment of the first aspect, samples marked as True are placed into a subset The subset The expression is as follows:
[0027]
[0028] In the formula, x i represents the input image; Represents the input image x i Observation label of S m Represents a 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; Represents the weight of the i-th sample in the e-th epoch; Represents 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.
[0029] In a further embodiment of the first aspect, the historical sequence
[0030] In the formula, For the dataset The labeling results of all samples in the tth epoch based on the model numbered m.
[0031] In a further embodiment of the first aspect, in step S1, for the model numbered m=1:
[0032] Based on the JSD loss d i , 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 1In the chronological order, the index containing the clean label sample is marked as True, and the other indexes are marked as False;
[0033] Initializing a Collection and Empty, select sequence S 1 The samples that are marked as True for K consecutive epochs are put into the clean sample set The remaining samples are placed after removing the labels
[0034] In a further embodiment of the first aspect, in step S1, for the model numbered m=2:
[0035] Based on the JSD loss d i , 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 the chronological order, the index containing the clean label sample is marked as True, and the other indexes are marked as False;
[0036] Initializing a Collection and Empty, select sequence S 2 The samples that are marked as True for K consecutive epochs are put into the clean sample set The remaining samples are placed after removing the labels
[0037] 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.
[0038] 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; m In the chronological order, the index containing the clean label sample is marked as True, and the other indexes are marked as False, and the samples marked as True are put into the subset Remove labels from the remaining samples and put them into the subset
[0039] The robust training module is based on the subset Subset Robust training is performed on the DNN model numbered m={1,2}.
[0040] The selection module is used to select the historical sequence S after all rounds of pre-training are completed. m 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 this sequence into the benchmark set D. c .
[0041] 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.
[0042] 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 contrastive learning as described in the first aspect.
[0043] 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
[0044] Figure 1 This is the overall process of the noisy label image learning method based on balanced selection and contrastive learning.
[0045] Figure 2 The specific process for balanced selection. DETAILED DESCRIPTION
[0046] 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 known in the art are not described.
[0047] Embodiment 1:
[0048] 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:
[0049] 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 Conduct a joint w After each epoch, the Jensen–Shannon divergence of all training samples on the two models is calculated, and the samples in each category are sorted according to their corresponding divergence values using the small loss criterion, and the same number of samples are selected as samples with clean labels in the historical sequence S 1 and S 2 In the chronological order, the index containing the clean label sample is marked as True value, and the other indexes are marked as False value;
[0050] 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 S. 1 In the chronological order, the index containing the clean label sample is marked as True value, and the other indexes are marked as False value;
[0051] Then, initialize the collection and Empty, select sequence S 1 The samples that are marked as True for K consecutive epochs are put into the clean sample set The remaining samples are placed after removing the labels
[0052] Step 1.3: 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=1.
[0053] Step 1.4: 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 in the historical sequence S. 2 In the chronological order, the index containing the clean label sample is marked as True value, and the other indexes are marked as False value; then, initialize the set and Empty, select sequence S 2 The samples that are marked as True for K consecutive epochs are put into the clean sample set The remaining samples are placed after removing the labels
[0054] 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.
[0055] Step 1.6 If the preset total number of training rounds T is not reached tot , then jump to step 1.2 to continue training.
[0056] Step 1.7 After all the training in step 1 is completed, in the historical sequence S 1 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 this sequence into the benchmark set D. c .
[0057] 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 Conduct a joint w epochs of warm-up training.
[0058] Step 1.9: For the model with 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 contained in the benchmark set D c The samples in the set are also marked as True, and the indexes of all remaining samples are marked as False values; then, the set is initialized and Empty, samples marked as True are put into the clean sample set The remaining False samples are removed from the label and put into
[0059] 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.
[0060] 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 contained in the benchmark set D c The samples in the set are also marked as True, and the indexes of all remaining samples are marked as False values; then, the set is initialized and Empty, samples marked as True are put into the clean sample set The remaining False samples are removed from the label and put into
[0061] 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.
[0062] Step 1.13 If the preset total number of training rounds T is not reached tot , then jump to step 1.9 to continue training.
[0063] Embodiment 2:
[0064] Based on Example 1, refer to the attached Figure 1 The overall learning process of the present invention is shown. Figure 2 The collection benchmark set D proposed by the present invention is shown c The specific process of equilibrium selection in the process.
[0065] Assume the original data set is Where n is the number of samples in the data set, is the image x i The observed label of the image 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 represented as h(·; φ m ) and f(·;θ m ), where m∈{1,2} is the model number and φ m and θ m Represents 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 ).
[0066] 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 T w ;
[0067]
[0068] 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);
[0069]
[0070] Here p m (x i )=h(f(x i θ m );φ m ) is the DNN model numbered m based on image x i The output value of . KL(·) represents the Kullback-Leibler function, as shown in Formula 3.
[0071]
[0072] 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 According to formula (4), we can get:
[0073]
[0074] 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 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.
[0075]
[0076] 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 S 1 Finally, initialize the collection and is empty, as shown in formula (6), select sequence S 1 The samples that are marked as True for K consecutive epochs are put into the clean sample set The remaining samples are placed after removing the labels
[0077]
[0078] Where t represents the tth epoch.
[0079] 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 L according to formula (7): ssl ;
[0080]
[0081] Among them, L l Based on the clean label subset The classification loss, L ul Based on the unlabeled subset The classification loss, L reg is the regularization loss, which can encourage the classifier's prediction value not to be biased towards a specific category. u is a preset weight coefficient. In addition, the present invention introduces the SimCLR contrastive learning method, which includes a mapping header for each model m={1,2}. in Is the parameter of the mapping header. The feature extractor f(·;θ m ) extracts high-dimensional features f(x i ) is mapped to a low-dimensional space, which facilitates further distinction. Based on the mapped low-dimensional features, 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.
[0082]
[0083]
[0084] in is the low-dimensional feature after mapping; and x 2i and x 2i-1 is the sample x i The result after two strong data enhancement transformations. Finally, the overall loss of the model numbered m is shown in formula (9):
[0085]
[0086] Step 1.4: For the model with 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 S 2 Finally, initialize the collection and is empty, according to formula (6), the historical sequence S 2 The samples that are marked as True for K consecutive epochs are put into the clean sample set The remaining samples are placed after removing the labels
[0087] 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).
[0088] Step 1.6 If the preset total number of training rounds T is not reached tot , then jump to step 1.2 to continue training.
[0089] Step 1.7 After all the training in step 1 is completed, as shown in formula (10), in the historical sequence S 1 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 this sequence into the benchmark set D. c .
[0090]
[0091]
[0092] 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 Conduct a joint w epochs of warm-up training.
[0093] 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 contained in the benchmark set D c The indices of all remaining samples are marked as False values.
[0094]
[0095]
[0096] Finally, initialize the collection and is empty, and the samples marked as True in the original data set are put into the clean sample set according to formula (11) The remaining samples are placed after removing the labels
[0097] 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).
[0098] Step 1.11. 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 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 contained in the benchmark set D c The index of the sample in is also marked as True, and the index of all remaining samples is marked as False. Finally, initialize the set and is empty, and the samples marked as True in the original data set are put into the clean sample set according to formula (11) The remaining samples are placed after removing the labels
[0099] 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).
[0100] Step 1.13 If the preset total number of training rounds T is not reached tot , then jump to step 1.9 to continue training.
[0101] Embodiment 3:
[0102] 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 the DNN model numbered m={1,2} based on the original data set samples; mIn the chronological order, the index containing the clean label sample is marked as True, and the other indexes are marked as False, 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 The DNN model numbered m = {1,2} is robustly trained. The selection module is used to select the DNN model in the historical sequence S after all rounds of pre-training are completed. m 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 this sequence 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.
[0103] 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.
[0104] Embodiment 4:
[0105] 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.
[0106] 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:
[0107] 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%;
[0108] 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;
[0109] 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.
[0110] 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%.
[0111] 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
1. A noisy label image learning method based on balanced selection and contrastive learning, characterized in that: The steps include: S1. Using the joint loss function, based on the original dataset samples, The DNN model is pre-trained for several rounds; The joint loss function The expression is as follows: ; In the formula, represents the input image; n represents the number of samples in the data set; Represents input features The predicted output value of In the historical sequence In the chronological order, the index containing the clean label sample is marked as True, and the other indexes are marked as False, and the samples marked as True are put into the subset , remove the labels of the remaining samples and put them into the subset ; Historical Sequence ; In the formula, For the dataset The labeling results of all samples in the tth epoch based on the model numbered m; Subset The expression is as follows: ; In the formula, represents the input image; Represents the input image Observation labels of Represents a historical sequence; Represents the weight value of the i-th sample in the t-th epoch; Indicates the t-K+1th epoch; Represents the weight of the i-th sample in the e-th epoch; Represents 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; For the number The DNN model is pre-trained for several rounds, including: calculating the JSD loss of all samples in the initial noise dataset under the DNN model : ; In the formula, Indicates that the DNN model numbered m=1 is based on the image The output value of Indicates that the DNN model numbered m=2 is based on the image The output value of KL(*) represents the Kullback-Leibler function, as follows: ; Where m and n represent input parameters; S2. Based on the subset , subset , for pairs numbered Robust training of DNN models; 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 this sequence into the benchmark set. ; S4, reinitialize the number The DNN model is trained and steps S1 to S3 are repeated until the preset total number of training rounds is reached.
2. The noisy label image learning method according to claim 1, characterized in that: 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 an image The observation labels are , and c represents the number of categories contained in the dataset.
3. The noisy label image learning method according to claim 1, characterized in that: In step S1, for the model numbered m=1: Based on the JSD loss , 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 In the chronological order, the index containing the clean label sample is marked as True, and the other indexes are marked as False; Initializing a Collection and Empty, select sequence The samples that are marked as True for K consecutive epochs are put into the clean sample set , the remaining samples are removed from the labels and placed in .
4. The noisy label image learning method according to claim 1, characterized in that: In step S1, for the model numbered m=2: Based on the JSD loss , 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 In the chronological order, the index containing the clean label sample is marked as True, and the other indexes are marked as False; Initializing a Collection and Empty, select sequence The samples that are marked as True for K consecutive epochs are put into the clean sample set , the remaining samples are removed from the labels and placed in .
5. A noisy label image learning system, used to implement the noisy label image learning method according to any one of claims 1 to 4; characterized in that: The noisy label image learning system includes: Pre-training module; the pre-training module uses a joint loss function to train the samples numbered The DNN model is pre-trained for several rounds; In the chronological order, the index containing the clean label sample is marked as True, and the other indexes are marked as False, and the samples marked as True are put into the subset , remove the labels of the remaining samples and put them into the subset ; A robust training module; the robust training module is based on the subset , subset , for pairs numbered Robust training of DNN models; Selection module; the selection module is used to select the historical sequence 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 this sequence into the benchmark set. ; Repeat execution module; the repeated execution module is used to reinitialize the number The DNN model is fed back to the pre-training module, the robust training module, and the selection module until the preset total number of training rounds is reached.
6. A computer-readable storage medium, characterized in that: The storage medium stores at least one executable instruction. When the executable instruction is executed on the electronic device, the electronic device executes the noisy label image learning method based on balanced selection and contrast learning as described in any one of claims 1 to 4.
