A self-label modification method for dealing with noisy labels

Abstract

The present invention discloses a self-label modification method for processing noisy labels. The method randomly selects small batches of data samples, performs data augmentation processing on the data samples to obtain different views, uses them as inputs to a pseudo-twin neural network, and outputs the predicted probability of the data sample category. The prediction calculations of different networks for different views and the JS divergence of the data sample label distribution are used to determine the possibility of them being clean data samples. According to a given judgment threshold, the batch of data samples are divided into clean data samples and noisy data samples. The labels of the clean data samples are only smoothed, and the noisy data samples are dynamically weighted according to the model's predictions and the sample's own labels to give them reliable labels. Finally, the classification loss function and the consistency loss function are used to update the model. The method of the present invention is used to solve image classification tasks under label noise and achieve good performance results.

Description

Technical Field

[0001] The invention belongs to the technical field of computer vision and mainly relates to a self-label modification method for processing noise labels. Background Art

[0002] Deep neural networks have achieved tremendous progress in various computer vision tasks, and these achievements are inseparable from large-scale, reliably annotated datasets, such as ImageNet. However, collecting well-annotated datasets is extremely labor-intensive, material-intensive, and time-consuming, especially when specialized knowledge is required (e.g., fine-grained classification). The high cost of obtaining large amounts of well-labeled data constitutes a bottleneck for the use of deep neural networks in real-world scenarios. To alleviate this problem, data annotation companies have resorted to crowdsourcing to raise data and annotations, scraping data from the web, or, due to limited resources, employing only one or a small number of annotators, or using alternative methods such as online queries to improve labeling efficiency. Unfortunately, while these approaches are cheaper and easier to obtain, they often result in unavoidable noisy labels due to the unreliability of non-expert annotations, error-prone automatic annotation systems, or limited annotators without the ability to conduct repeated verification.

[0003] The complexity of the network structure makes deep networks very susceptible to overfitting to noisy labels. Samples with noisy labels will inevitably be fitted by deep neural networks, thus affecting the performance of the model. Therefore, there is an urgent need to study robust learning methods that are resistant to noisy labels.

[0004] Methods for studying noisy labels mainly include: 1. Approaching the problem by estimating the underlying noisy transition matrix. The main difficulty of this type of problem lies in the need to accurately estimate the noisy transition matrix, which requires good prior knowledge. 2. Designing noise-resistant loss functions and correcting the loss based on the predictions of deep neural networks. However, these methods are prone to failure when the dataset is large. 3. Using selected or reweighted training samples to train deep neural networks. The main challenge of this type of problem lies in designing an appropriate criterion for identifying clean samples. At the same time, how to improve the reliability of clean data samples is also a question that needs to be considered. 4. Modifying the labels of data samples, mainly by combining the output of the prediction network to correct the labels of data samples that are considered noisy. However, the degree of trust in the prediction network is also a question that needs to be considered. 5. Studying the noisy label problem in the field of semi-supervised learning. However, the accuracy of many semi-supervised learning classifiers decreases significantly in the presence of label noise. Summary of the Invention

[0005] The present invention is precisely aimed at the problem of model performance degradation caused by noisy labels in the prior art, and provides a self-label modification method for processing noisy labels. A small batch of data samples are randomly selected, and the data samples are subjected to data enhancement processing to obtain different views. These are used as input to a pseudo-twin neural network, and the predicted probability of the data sample category is output. The JS divergence between the prediction calculation of different views by different networks and the data sample label is used to judge its possibility as a clean data sample; according to a given judgment threshold, the batch of data samples is divided into clean data samples and noisy data samples, and the labels of the clean data samples are only smoothed. The noisy data samples are dynamically weighted according to the model's prediction and the sample's own label to give them reliable labels, and the model is updated using the proposed classification loss function and consistency loss function. The method of the present invention can achieve good performance results in artificially synthesized noisy data sets and large-scale noisy data sets from real scenarios, and also has the characteristics of faster convergence during the training process.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is: a self-label modification method for processing noisy labels, characterized by comprising the following steps:

[0007] S1, in the process of using the dataset to train the model, randomly select small batch data samples For each data sample X, two data augmentation methods, scaling and cropping, are used to process it to obtain different views V and V';

[0008] S2, the different views V and V' obtained in step S1 are used as the input of two pseudo twin neural networks, and the outputs of the two pseudo twin networks are respectively passed through the soft-max layer to obtain the final prediction output result

[0009] P1, P1', P2, P2', where P1 and P1' are generated by the output of network 1 through the soft-max layer, and the input is

[0010] are V and V' respectively; P2 and P2' are generated by the output of network 2 through the soft-max layer, and the input is V and V' respectively;

[0011] S3, calculates the difference between the pseudo twin neural network output in step S2 and the label distribution given by the sample, specifically

[0012]

[0013] in, For data sample x i Jensen-Shannon (JS) divergence is used to measure the probability distribution of predictions; For data sample xi Given the true label distribution; D KL (·||·) represents the Kullback-Leibler (KL) divergence function;

[0014] The label distribution of the data sample is 0-1 distribution, and only the category to which it belongs is recorded as 1, and the rest are 0. In order to prevent the true number of the logarithm from being 0 during the calculation process, this distribution is converted into the following formula for calculation:

[0015]

[0016] Among them, the given label is l i ∈{1,2,3,...,C}, ε is a hyperparameter used to control the smoothness of the label distribution;

[0017] S4, using the distribution difference d obtained in step S3 i , calculate the data sample x i is the probability of a clean sample, which is expressed as follows:

[0018]

[0019] in, Indicates p i and y i consistency between

[0020] S5, calculate the threshold for selecting clean data samples according to the training rounds and determine the threshold τ clean Then, for the data sample x i If the following formula is satisfied, we can preliminarily determine x i Here is a clean data sample:

[0021]

[0022] S6, based on the output of the two pseudo-twin neural networks, clean data samples are selected. Only when both neural networks judge that the data sample is clean, the data sample will participate in the subsequent model update. The expression for selecting the sample set is as follows:

[0023]

[0024] in, and They are the results of using the output predictions of two neural networks to judge the data samples;

[0025] S7, by judging and selecting data samples, the training data is divided into two subsets: clean data sample set and a set of noisy data samples

[0026] S8, using smoothed label distribution to process clean data sample sets The sample labels in are expressed as follows:

[0027]

[0028] S9, relying on the pseudo twin neural network in step S2 to process the sample labels in the noisy data sample set, is expressed as follows:

[0029]

[0030] in, is the label after the smoothing process in step S3; i is the prediction result output by the pseudo-twin neural network in step S2. Here, you can select one of the pseudo-twin neural networks; ∈ is the weight given to the model output;

[0031] S10, the label distribution modified by steps S8 and S9 is subjected to cross entropy loss with the probability distribution predicted by the model, and the classification loss function is calculated. The classification loss expression is as follows:

[0032]

[0033] Among them, the data sample x i After different data enhancement processes, two different views v are obtained i and v i ', as the input of the two networks, the output prediction probability distribution is recorded as p i1 , p' i1 , p i2 , p' i2 ; is the modified label distribution. For the selected clean data samples, Obtained from step S8; for data samples that are considered to have noise, Obtained from step S9; N is the number of processed data samples;

[0034] S11, calculation of consistency loss function, specifically:

[0035]

[0036] Among them, D KL (·||·) represents the Kullback-Leibler (KL) divergence function, p i1 , p' i1 , p i2 , p' i2 , N represents the same classification loss function;

[0037] S12, integrating the classification loss function obtained in step S10 and the consistency loss function obtained in step S11 to calculate the overall loss function, which is expressed as follows:

[0038]

[0039] Among them, α is a hyperparameter used to adjust the weights of the two losses;

[0040] S13, using the overall loss function to calculate the gradient descent, to update the model parameters, to obtain the optimal model for solving the noisy labels:

[0041]

[0042] Where θ = {θ1, θ2}, θ1 and θ2 represent the parameters of the two networks respectively. The training process is repeated after the update. If the set number of iterations is not reached, step S1 is executed. Otherwise, the training round is exited and the next training round is executed until the training round is completed.

[0043] Compared with existing technologies, this invention offers the following advantages: it provides a self-label modification method for processing noisy labels, enabling image classification tasks in the presence of label noise and achieving higher performance within a given dataset. This method fully utilizes training data during training, while relying solely on its own model for training, independent of additional auxiliary models. This method achieves good performance on both synthetic noisy datasets and large-scale noisy datasets from real-world scenarios, while also exhibiting faster convergence during training. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a framework diagram of the method of the present invention;

[0045] Figure 2 This figure compares the classification performance of the present invention with the existing main methods, including Decoupling, Co-teaching, Co-teaching+, JoCOR, and Jo-SRC, on the Clothing1M dataset. The Standard method directly trains the network on a noisy dataset.

[0046] Figure 3 This paper compares the classification performance of the present invention with existing main methods, including CleanNet and DeepSelf, on the Food101N dataset. The Standard method directly trains the network on a noisy dataset.

[0047] Figure 4This figure compares the classification performance of the present invention on the artificially synthesized noisy dataset on the CIFAR100 dataset with the existing main methods, including Decoupling, Co-teaching, Co-teaching+, JoCoR, and Jo-SRC methods. Among them, Standard is a method for directly training the network on the noisy dataset. DETAILED DESCRIPTION

[0048] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0049] Example 1

[0050] A self-label modification method for processing noisy labels, the framework of this method is as follows Figure 1 As shown, the present invention uses a pseudo-twin neural network to strictly judge whether a data sample is a clean sample. For data samples that are considered to be noisy, the method of the present invention performs self-label modification, does not rely on additional auxiliary networks, and only relies on the pseudo-twin neural network in the method for modification. At the same time, dynamic weights are given to the output predictions of the pseudo-twin neural network, which makes its trustworthiness more reasonable as training progresses. Finally, the proposed consistency loss and classification loss are used to update the model. Specifically, the following steps are included:

[0051] Step S1: Randomly select small batch data samples while training the model using the dataset For each data sample X, we use data augmentation technology to process it to obtain different views V and V'. Specifically, for the same data sample, we use two data augmentation methods, scaling and cropping, to process the data sample to obtain two views.

[0052] Step S2: For each sample’s different views V and V’, they are used as inputs to Network 1 and Network 2. Network 1 and Network 2 are pseudo-twin neural networks. These two networks can predict labels independently. Their parameters are different, but they are updated simultaneously using the same loss function. The output of the network is obtained through soft-max.

[0053] The final prediction output P1, P1', P2, P2' is obtained by layer, where P1 and P1' are obtained by the output of network 1 through

[0054] The soft-max layer generates the outputs of network 2, with inputs V and V' respectively. P2 and P2' are generated by the soft-max layer from the output of network 2, with inputs V and V' respectively. Networks 1 and 2 have the same network architecture but do not share parameters. Both networks are updated simultaneously using stochastic gradient descent using the same loss function.

[0055] Step S3: Calculate the difference between the pseudo twin neural network output in step S2 and the label distribution given by the sample. For the data sample x i The present invention uses Jensen-Shannon (JS) divergence to measure the probability distribution of its prediction And the given true label distribution The difference between them is expressed as follows:

[0056]

[0057] Among them D KL (·||·) represents the Kullback-Leibler (KL) divergence function.

[0058] The label distribution of the data sample is 0-1 distribution, and only the category to which it belongs is recorded as 1, and the rest are 0. In order to prevent the true number of the logarithm from being 0 during the calculation process, the smoothed label distribution is used in the calculation formula (1) and converted into formula (2) for calculation.

[0059]

[0060] Among them, the given label is l i ∈{1,2,3,...,C}, ε is a hyperparameter used to control the smoothness of the label distribution. After debugging, the hyperparameter ε for smoothing the labels should be selected as 0.7.

[0061] Step S4: Calculate the data sample x i is the probability of a clean sample. JS divergence can be used to measure the difference between two probability distributions, and its value is between 0 and 1. Therefore, intuitively, we can use d i To measure the data sample x i is the probability of a clean sample, which is expressed as follows:

[0062]

[0063] In fact, Indicates p i and y i consistency between them.

[0064] Step S5: Calculate the threshold for selecting clean data samples based on the number of training rounds. The present invention dynamically processes the threshold for determining whether a data sample is clean by the following method:

[0065]

[0066] Where t represents the number of training rounds, △τ=τ m -τc , τ c is a hyperparameter, τ m is a custom constant, the threshold τ clean It is divided into two stages. The first stage is 1≤t≤t w This stage is to make the model have a certain predictive ability as soon as possible, so only clean data samples are selected and the model is updated using the selected data samples without modifying the labels. The second stage t w <t≤t max , when the model has a certain predictive ability, in order to make more effective use of the data, the labels of the data samples that are judged to be noisy are modified. clean It changes with training in a linear manner in two phases.

[0067] Given an appropriate threshold τ clean Then, for the data sample x i If the following formula is satisfied, we can preliminarily determine x i Here is a clean data sample:

[0068]

[0069] The ratio of training rounds between the two phases is not fixed. During training, the first phase can achieve saturation performance, while the second phase requires full training to achieve higher performance. The threshold for selecting clean data samples in the initial process should be low enough to prevent the number of samples selected from being too small.

[0070] Step S6: Select clean data samples based on the network's output predictions. Since the two networks in the framework have different learning capabilities, they can filter out errors caused by different types of noise labels. To improve the reliability of sample selection, this invention adopts a dual-model structure to strengthen the screening of clean labels. Only when both networks determine that a data sample is clean will the data sample be included in the subsequent model update. The expression for selecting the sample set is as follows:

[0071]

[0072] in and They are the results of judging the data samples using the output predictions of the two networks.

[0073] Although the two networks can filter different types of noise, at the same stage, the thresholds used by the two networks to determine whether a data sample is a clean data sample are the same.

[0074] Step S7: Divide the training data. By judging and selecting the data samples, we divide the training data into two subsets: the clean data sample set and the and a set of noisy data samples

[0075] Step S8: Processing clean data sample set The present invention keeps their labels unchanged, but in order to improve the generalization performance and prevent the logarithm from being 0 in the calculation of cross entropy, a smoothed label distribution is used, which is expressed as follows:

[0076]

[0077] For the labels of clean data samples, the process of calculating the distribution difference between the network output and the given label of the sample in the previous step still needs to be consistent.

[0078] Step S9: Processing the noisy data sample set The present invention adopts a self-label modification method and relies only on the pseudo-twin neural network in step S2 to process the sample labels in the noisy data sample set. When a data sample is predicted to be noisy, the given label and the model prediction are in conflict. Since it is not completely certain whether the corresponding label is wrong or correct, different weights should be given to the sample label itself and the model prediction output. The expression is as follows:

[0079]

[0080] In the formula is the label after smoothing by formula (2), p i is the prediction result output by the pseudo-twin neural network in step S2. Here, we can select one of the pseudo-twin neural networks. ∈ is the weight given to the model output. ∈ determines the extent to which we should trust the label distribution predicted by the model. Considering that the model should become more reliable as training progresses, ∈ should be dynamic and increase continuously as training progresses. At the same time, considering that the predicted output distribution should be reasonable, ∈ can be defined as:

[0081] ∈=g(t)×l(p)(9)

[0082] Among them, g(t) determines the extent to which the learner can be trusted. It is data-independent and is expressed as follows:

[0083]

[0084] Where Γ represents the total number of training iterations and t represents the current training round.

[0085] l(p) determines how much we can trust the predicted label distribution. It is data-dependent and can be expressed as follows:

[0086] l(p)=1-H(p) / H(u)(11)

[0087] Where H(p) represents the information entropy of the model output prediction, and H(u) is expressed as follows:

[0088]

[0089] The weight ∈ assigned to the model output is determined by weighting the predicted outputs of the two networks, and the weight is 0.5.

[0090] Step S10: Calculate the classification loss function. Perform cross entropy loss on the modified label distribution and the probability distribution predicted by the model. The classification loss expression is as follows:

[0091]

[0092] Among them, the data sample x i After different data enhancement processes, two different views v are obtained i and v i ', as the input of the two networks, the output prediction probability distribution is recorded as p i1 , p' i1 , p i2 , p' i2 . is the modified label distribution. For the selected clean data samples, From formula (7), we can get that for the data samples that are considered to have noise, Obtained by formula (8). N is the number of processed data samples. In the first stage of threshold dynamicization, in order to enable the model to have a certain predictive ability as soon as possible, only clean data samples are selected without labeling the noise data samples. At this time, the classification loss is as follows:

[0093]

[0094] in It is a set of data samples that satisfies both networks to judge that the data samples are clean, and is calculated according to formula (6).

[0095] In the second stage of threshold dynamicization, the model has a certain predictive ability and modifies the labels of data that are judged to be noisy. At this time, N is the number of data samples in a small batch.

[0096] The labels used to calculate the classification loss are processed differently in two stages. The first stage is to smooth the labels. The second stage is to process the clean data samples and the noisy data samples differently according to the above steps.

[0097] Step S11: Calculate the consistency loss function. The present invention designs the consistency loss to maximize the consistency between the two classifiers and the consistency between the output predictions obtained by the same network for different views of the input. Its expression is as follows:

[0098]

[0099] Among them D KL (·||·) represents the Kullback-Leibler (KL) divergence function, p i1 , p' i1 , p i2 , p' i2 , N represents the same classification loss function.

[0100] Step S12: Calculate the overall loss function. Integrate the classification loss function and the consistency loss function. The overall loss function expression is as follows:

[0101]

[0102] Among them, α is a hyperparameter used to adjust the weights of the two losses.

[0103] Step S13: Use the overall loss function to calculate the gradient descent to update the model parameters:

[0104]

[0105] Where θ = {θ1, θ2}, where θ1 and θ2 represent the parameters of the two networks. After the update, the training process is repeated. If the set number of iterations is not reached, step 1 is executed. Otherwise, the training round is exited and the next training round is executed until the training round is completed.

[0106] Test Case

[0107] The classification performance of the proposed method is compared with the existing advanced methods in the field of noisy label processing on the Clothing1M dataset, Food101N dataset and CIFAR100 dataset. For each dataset, the compared methods are different. The specific comparison methods are shown in the accompanying drawings. The comparison results are shown in Figure 2-Figure 4 .

[0108] Figure 2The comparison of the classification performance of the present invention with existing advanced methods in the field of noisy label processing on the Clothing1m dataset is shown, including Decoupling, Co-teaching, Co-teaching+, JoCOR, and Jo-SRC. Standard is a method of directly training the network on a noisy dataset. It can be seen that this method achieves the best results, with a performance approximately 0.2% higher than the previously achieved best performance method Jo-SRC. However, the training process of the Jo-SRC method requires the use of predictions from the teacher model, and therefore relies on an accurate auxiliary model to generate predictions. The present invention lightweights the network model during the training process and achieves higher performance on the Clothing1m dataset.

[0109] Figure 3 This paper demonstrates a comparison of our method with existing state-of-the-art methods for handling noisy labels on the Food101N dataset, including CleanNet, DeepSelf, and Standard, which train networks directly on noisy datasets. Our method achieves slightly higher performance than the previous best-performing method, Jo-SRC, demonstrating its effectiveness in handling real-world noise. However, Jo-SRC utilizes a teacher model during training, demonstrating the method's superior performance without relying on an additional auxiliary model.

[0110] Figure 4 This paper compares our method with existing state-of-the-art methods for handling noisy labels, including Decoupling, Co-teaching, Co-teaching+, JoCoR, and Jo-SRC, on a noisy dataset synthesized from the CIFAR100 dataset. Standard (a method that directly trains the network on the noisy dataset) is used. Noise types include "symmetric" and "asymmetric." For "symmetric," the noise rates are set to 0.2, 0.4, and 0.8, while for "asymmetric," the noise rate is set to 0.4. As shown in the figure, our method consistently outperforms existing state-of-the-art methods for handling noisy labels.

[0111] In summary, this case proposes a simple and effective method to solve the problem of performance degradation caused by noisy labels in image classification. In response to the problem that existing methods lack reliability in determining whether data samples are clean, this method adopts a dual-model structure to filter out errors caused by different types of noisy labels and maximize the consistency between the prediction outputs of the dual models. For clean data samples, this method smoothes their labels to improve the generalization performance of the model and prevents the occurrence of true numbers of 0 in the process of calculating cross entropy. For noisy data samples, this method determines their labels based on the model's predictions and its own labeled labels, and assigns dynamic weights between the two. This modification method does not rely on other models, but only on the model of the framework itself. In addition, this method also proposes a classification loss function and a consistency loss function to update the model. Experiments were conducted on synthetic noisy datasets and large-scale noisy datasets in real scenarios, and good performance results were achieved to prove the effectiveness of the proposed method.

[0112] It should be noted that the above content merely illustrates the technical idea of ​​the present invention and cannot be used to limit the scope of protection of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications all fall within the scope of protection of the claims of the present invention.

Claims

1. A self-label modification method for processing noisy labels, characterized by: Randomly select a small batch of data samples, perform data augmentation on the data samples to obtain different views, and use them as the input of the pseudo-twin neural network. Output the predicted probability of the data sample category. According to the prediction calculation of different networks for different views and the JS divergence of the data sample label distribution, it is used to judge its possibility as a clean data sample. According to the given judgment threshold, the batch of data samples is divided into clean data samples and noisy data samples. The labels of the clean data samples are only smoothed. The noisy data samples are dynamically weighted according to the model's prediction and the sample's own label to give them reliable labels. The classification loss function and consistency loss function are used to update the model. The judgment threshold for determining whether the data sample is clean is dynamically processed as follows: Where t represents the number of training rounds; △τ = τ m -τ c , τ c is a hyperparameter; τ m is a custom constant; the threshold τ clean It is divided into two stages. The first stage is 1≤t≤t w , only clean data samples are selected, and the model is updated using the selected data samples without modifying the labels; the second stage t w <t≤t max , modify the label of the data sample determined to be noisy; the threshold τ clean It changes with training in a linear manner in two phases.

2. A self-label modification method for processing noisy labels according to claim 1, characterized in that: The following steps are involved: S1, in the process of using the dataset to train the model, randomly select small batch data samples For each data sample X, two data augmentation methods, scaling and cropping, are used to process it to obtain different views V and V'; S2, the different views V and V' obtained in step S1 are used as the input of two pseudo-twin neural networks. The outputs of the two pseudo-twin networks are respectively passed through the soft-max layer to obtain the final predicted output results P1, P'1, P2, and P'2, where P1 and P'1 are generated by the output of network one through the soft-max layer, and the inputs are V and V' respectively; P2 and P'2 are generated by the output of network two through the soft-max layer, and the inputs are V and V' respectively; S3 calculates the difference between the pseudo twin neural network output in step S2 and the label distribution given by the sample, specifically: in, The data sample is x i Jensen-Shannon (JS) divergence is used to measure the probability distribution of predictions; For data sample x i Given the true label distribution; D KL (·||·) represents the Kullback-Leibler divergence function; The label distribution of the data sample is 0-1 distribution, and only the category to which it belongs is recorded as 1, and the rest are 0. In order to prevent the true number of the logarithm from being 0 during the calculation process, this distribution is converted into the following formula for calculation: Among them, the given label is l i ∈{1,2,3,...,C}, ε is a hyperparameter used to control the smoothness of the label distribution; S4, using the distribution difference d obtained in step S3 i , calculate the data sample x i is the probability of a clean sample, which is expressed as follows: in, Indicates p i and y i consistency between S5, calculate the threshold for selecting clean data samples according to the training rounds and determine the threshold τ clean Then, for the data sample x i , if the following formula is satisfied, then the preliminary judgment is x i Here is a clean data sample: S6, based on the output of the two pseudo-twin neural networks, clean data samples are selected. Only when both neural networks judge that the data sample is clean, the data sample will participate in the subsequent model update. The expression for selecting the sample set is as follows: in, and They are the results of using the output predictions of two neural networks to judge the data samples; S7, by judging and selecting data samples, the training data is divided into two subsets: clean data sample set and a set of noisy data samples S8, using smoothed label distribution to process clean data sample sets The sample labels in are expressed as follows: S9, relying on the pseudo twin neural network in step S2 to process the noisy data sample set The sample labels in are expressed as follows: in, is the label after the smoothing process in step S3; i is the prediction result output by the pseudo twin neural network in step S2; ∈ is the weight given to the model output; S10, the label distribution modified by steps S8 and S9 is subjected to cross entropy loss with the probability distribution predicted by the model, and the classification loss function is calculated. The classification loss expression is as follows: Among them, the data sample x i After different data enhancement processes, two different views v are obtained i and v i ', as the input of the two networks, the output probability distributions are recorded as p i1 , p' i1 , p i2 , p' i2 ; is the modified label distribution. For the selected clean data samples, Obtained from step S8; for data samples that are considered to have noise, Obtained from step S9; N is the number of processed data samples; S11, calculation of consistency loss function, specifically: Among them, D KL (·||·) represents the Kullback-Leibler divergence function, p i1 , p' i1 , p i2 , p' i2 , N represents the same classification loss function; S12, integrating the classification loss function obtained in step S10 and the consistency loss function obtained in step S11 to calculate the overall loss function, which is expressed as follows: Among them, α is a hyperparameter used to adjust the weights of the two losses; S13, using the overall loss function to calculate the gradient descent, to update the model parameters, to obtain the optimal model for solving the noisy labels: Where θ = {θ1, θ2}, θ1 and θ2 represent the parameters of the two networks respectively. The training process is repeated after the update. If the set number of iterations is not reached, step S1 is executed. Otherwise, the training round is exited and the next training round is executed until the training round is completed.

3. The self-label modification method for processing noisy labels according to claim 2, characterized in that: In step S2, the two pseudo-twin neural networks are identical in network structure but do not share parameters. The same loss function is used to update the two neural networks simultaneously using the stochastic gradient descent method.

4. The method for modifying self-labels for processing noisy labels according to claim 2, characterized in that: In step S6, the two neural networks use the same threshold value to determine whether a data sample is a clean data sample.

5. The method for modifying self-labels for processing noisy labels according to claim 4, characterized in that: In step S9, the weight ∈ of the model output is dynamic and increases as the training progresses, that is, ∈ can be defined as: ∈=g(t)×l(p) H(u)=-log(1 / C) Where Γ represents the total number of training iterations; t represents the current training round; and H(p) represents the information entropy of the model output prediction.

6. The method for modifying self-labels for processing noisy labels according to claim 5, characterized in that: In step S10, in the first stage of threshold dynamicization, only clean data samples are selected without labeling the noise data samples. The classification loss is as follows: in, It is a set of data samples that satisfies both networks to judge that the data samples are clean; In the second stage of threshold dynamicization, the model modifies the labels of data judged to be noisy, where N is the number of data samples in a mini-batch.

Images