Method and system for classifying images with noise labels based on recalibration strategy and Mixup loss

By introducing Mixup loss and recalibration strategies in deep neural networks, the problems of deep neural networks being overfitted and noise-sensitive in noise label learning are solved, and high-performance image classification in high-noise scenarios are realized, which enhances the model's Robustness and generalization ability.

CN119027723APending Publication Date: 2024-11-26JIANGSU OPEN UNIVERSITY (THE CITY VOCATIONAL COLLEGE OF JIANGSU)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411030397.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-30
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The existing noisy label learning method has poor performance in high-noise scenarios and is sensitive to prior noise information, which limits its practical application. In particular, deep neural networks are prone to overfitting noise labels, resulting in a decrease in classification performance and generalization capabilities.

Method used

Using the noise-containing label image classification method based on Mixup loss and recalibration strategies, two deep neural networks are pre-trained through Mixup loss, and GMM segments the noise-label dataset, performs pure self-supervised learning training, and every few The wheel uses CE loss for recalibration to reduce dependence on prior noise information.

Benefits of technology

This method improves the robustness and performance of the network, is suitable for various noise scenarios, avoids overfitting, enhances the generalization ability of the model, significantly improves the test accuracy, and is suitable for high noise scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119027723A_ABST
    Figure CN119027723A_ABST
Patent Text Reader

Abstract

The invention discloses a noise label-containing image classification method and system based on a recalibration strategy and Mixup loss. The method comprises the following steps: (a) a pre-training process using the Mixup loss; a plurality of rounds of pre-training are carried out on a network A and a network B with the same structure by using Mixup loss. And (b) a pure SSL robust training process. In each SSL training process, a whole data set is input into the network A and the network B for data segmentation. In this stage, a marked subset and an unmarked subset segmented using a GMM model based on the network A are used for robust SSL training of the network B, and vice versa. And (c) an SSL training process with a re-calibration strategy. After every K rounds of pure SSL training process, the optimization directions of the network A and the network B are corrected by using a re-calibration strategy proposed based on CE loss. The method is insensitive to noise types, is suitable for various noise scenes, and can effectively enhance the robustness and performance of the network.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of computer vision, and in particular relates to a noisy label image classification method and system based on a recalibration strategy and Mixup loss. Background Art

[0002] The emergence of large manually annotated datasets has enabled deep neural networks (DNNs) to achieve great success in computer vision tasks. However, collecting high-quality annotated data is very expensive and time-consuming. Although we can collect large datasets with noisy labels through search engines and shopping websites, existing studies have shown that DNNs are prone to overfitting these noisy label samples, resulting in a significant decrease in classification performance and generalization ability. Therefore, research on learning with noisy labels has received increasing attention in the field of computer vision.

[0003] Existing research on noisy label learning (LNL) can be mainly divided into three categories: noise transition matrix estimation, robust loss function and sample selection method. The noise transition matrix estimation method estimates the probability of a label being converted from clean to noisy through a specific structure or sample, and uses this probability to assist in loss calculation. Since the CE loss commonly used in traditional classification tasks easily makes the model fit the noisy labels, some researchers have designed more robust and less fitting loss functions to assist in model training. Finally, the sample selection method identifies noisy samples through specific model structures or standards to reduce their impact on DNN training. These methods have achieved some success, but they still have disadvantages, such as poor performance in high-noise scenarios and sensitivity to prior noise information, which limits their practical application. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a method and system for classifying noisy labeled images based on a recalibration strategy and Mixup loss.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for classifying noisy labeled images based on a recalibration strategy and Mixup loss, comprising the following steps:

[0007] Based on Mixup loss, two deep neural networks A and B with the same structure are pre-trained for several rounds using a dataset with noisy labels to obtain two deep neural network models with preliminary discrimination capabilities; the dataset includes images and noisy labels for marking image categories;

[0008] Use pure SSL of the dataset with noisy labels to train two deep neural network models with preliminary discrimination ability; in each round of pure SSL training, first input the entire dataset with noisy labels into deep neural network A and deep neural network B for data segmentation, deep neural network A uses GMM to segment the dataset with noisy labels to obtain the labeled subset and unlabeled subset for pure SSL training of deep neural network B, and deep neural network B uses GMM to segment the dataset with noisy labels to obtain the labeled subset and unlabeled subset for pure SSL training of deep neural network A; after every K rounds of pure SSL training, use the recalibration strategy based on CE loss to correct the optimization direction of deep neural network A and deep neural network B; until the maximum training round is reached, the trained deep neural network A and deep neural network B are obtained;

[0009] Use the trained deep neural network A or deep neural network B for image classification.

[0010] To optimize the above technical solutions, the specific measures taken also include:

[0011] Furthermore, the pre-training is specifically:

[0012] Using datasets with noisy labels Pre-train two deep neural networks A and B with the same structure, x i represents the i-th picture, represents the noisy label for marking the i-th image, n is the size of the dataset with noisy labels; the serial numbers of deep neural network A and deep neural network B are represented as m=0 and m=1 respectively;

[0013] The Mixup loss is specifically:

[0014]

[0015] Where, L mix represents Mixup loss, l represents interpolation weight, L CE (·) represents CE loss, represents the predicted probability of the mth deep neural network output, Indicates that in the i-th image x i and any other images x j The i-th mixed input image obtained by linear interpolation between y i and j are the i-th picture x i and the jth image x j One-hot encoding label of ;

[0016] Interpolation weights λ and mixed input images The calculation formula is as follows:

[0017]

[0018]

[0019]

[0020] Among them, b represents the input parameter of the beta distribution function, Beta(b,b) represents the beta distribution function, Represents a random scalar in the interval [0,1] sampled from a beta distribution.

[0021] Furthermore, the GMM segmentation of the data set with noisy labels is specifically as follows:

[0022] Use a two-component GMM to model the loss values ​​of all samples and estimate the probability of each sample being a noise label or a clean label. The formula is as follows:

[0023]

[0024] Where g = {1,2} 1 represents the g-th Gaussian term of the GMM model, is the i-th picture x i Standardized CE loss of ; Represents the i-th picture x i The probability that the standardized CE loss belongs to the first Gaussian term, Represents the i-th picture x i The probability that the standardized CE loss belongs to the second Gaussian term;

[0025] The output probability of the Gaussian term with the smallest mean is selected as the probability that each image contains a clean label. The input pair (x i ,y i ) is the probability of clean data as:

[0026]

[0027] In the formula, represents the probability that the label of the i-th image is a clean label, st represents the constraint condition, Select the mean value m (g) The smallest Gaussian term, j represents the number of the selected Gaussian term; m (g) is the estimated mean of the g-th Gaussian term;

[0028] Based on estimated probability With a predefined probability threshold, the mth deep neural network transforms the dataset with noisy labels using the following equation Divide into labeled subsets X(m) and the unlabeled subset U (m) :

[0029]

[0030]

[0031] Where n is the size of the data set, t is the predefined probability threshold, represents the probability that the label of the i-th image is a clean label, is a dataset with noisy labels. It's a picture x i and noise labels The sample pairs composed of Considering that the noise label is the true label, the sample pair Put it into the labeled subset, otherwise, it is considered a noise label It is not the real label. After removing the noise label, the image x i into the unlabeled subset.

[0032] Furthermore, during the pure SSL training of deep neural network B and the pure SSL training of deep neural network A, the following formula is used to obtain the pseudo labels of the labeled subset to further eliminate the noise labels:

[0033]

[0034] In the formula, represents the pseudo-labels for the labeled subset, represents the probability that the label of the ith image is a clean label determined by the 1-mth deep neural network, represents the noise label of the i-th image, is the i-th image x input by the m-th deep neural network i The predicted probability of the h-th weak enhancement output;

[0035] Integrate the probabilities predicted by the two deep neural networks to jointly guess the true label of each unlabeled sample and obtain the pseudo labels of the unlabeled subset:

[0036]

[0037] In the formula, represents the pseudo-labels of the unlabeled subset, is the i-th image x input by the m-th deep neural network i The predicted probability of the output after the h-th weak enhancement; is the i-th image x input by the 1-m deep neural network i The predicted probability of the output after the h-th weak enhancement;

[0038] After obtaining the pseudo labels of the labeled subset and the pseudo labels of the unlabeled subset, a cooling operation is performed on the pseudo labels:

[0039]

[0040] In the formula, represents the pseudo label after cooling; is the pseudo label vector The jth component of ; c represents the number of image categories;

[0041] Use pseudo labels after cooling Replace the pseudo labels of the labeled subset and the pseudo labels of the unlabeled subset, and then and the image x corresponding to the label i Composition input pair Perform Mixup operation on the input pair to obtain a mixed input pair is a mixed input image, are the mixed input pseudo labels used to compute the loss for pure SSL training.

[0042] Furthermore, the loss of the pure SSL training is specifically:

[0043]

[0044] Where, L SSL represents the loss of pure SSL training, L x represents the supervision loss, L u represents the unsupervised loss, L reg represents the regularization term, a is the coefficient that controls the contribution of unlabeled samples to the update of the deep neural network model; X (1-m) and U (1-m) are the labeled subsets and unlabeled subsets divided by the 1st to mth deep neural networks, x i is the i-th picture, represents the predicted probability of the mth deep neural network output, Represents the predicted probability that the image category output by the mth deep neural network belongs to category c.

[0045] Furthermore, the recalibration strategy is implemented by performing a single supervised training using CE loss on all samples in the dataset with noisy labels;

[0046] The recalibration strategy corrects the recalibration loss of the optimization direction process of deep neural network A and deep neural network B as follows:

[0047]

[0048] Where, LRC represents the recalibration loss, n is the size of the dataset with noisy labels, represents the noise label for marking the i-th image, p (m) (x i ) represents the predicted probability of the output of the m-th deep neural network.

[0049] The present invention also proposes a noisy label image classification system based on a recalibration strategy and Mixup loss, comprising:

[0050] A pre-training module, used to perform several rounds of pre-training on two deep neural networks A and B with the same structure using a data set with noisy labels based on Mixup loss to obtain two deep neural network models with preliminary discrimination capabilities; the data set includes pictures and noisy labels for marking picture categories;

[0051] The formal training module is used to train two deep neural network models with preliminary discrimination ability using pure SSL of a dataset with noisy labels. In each round of pure SSL training, the entire dataset with noisy labels is first input into deep neural network A and deep neural network B for data segmentation. The labeled subset and unlabeled subset obtained by deep neural network A using GMM to segment the dataset with noisy labels are used for pure SSL training of deep neural network B. The labeled subset and unlabeled subset obtained by deep neural network B using GMM to segment the dataset with noisy labels are used for pure SSL training of deep neural network A. After every K rounds of pure SSL training, the optimization direction of deep neural network A and deep neural network B is corrected using a recalibration strategy based on CE loss. This is done until the maximum number of training rounds is reached, and the trained deep neural network A and deep neural network B are obtained.

[0052] The image classification module is used to perform image classification using the trained deep neural network A or deep neural network B.

[0053] The beneficial effects of the present invention are:

[0054] 1. We propose an improved sample selection framework based on Mixup loss and recalibration strategy, which is insensitive to the noise type, suitable for various noise scenarios, and can effectively enhance the robustness and performance of the network.

[0055] 2. Since existing sample selection methods require dynamic adjustment of the loss function based on the noise type during pre-training, which is impractical in practical applications, we replace the pre-training loss function with MixUp loss, avoiding excessive parameter adjustment and reducing dependence on prior noise information.

[0056] 3. We introduce a recalibration strategy based on CE loss for the SSL process. We have shown that periodic recalibration using only CE loss during training can prevent the model from falling into local optimality and improve test accuracy.

[0057] The effectiveness of our method is verified on a synthetic noise dataset. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a training framework diagram of the noisy label image classification method based on the recalibration strategy and Mixup loss proposed in the present invention during the model training stage. DETAILED DESCRIPTION

[0059] The present invention will now be described in further detail with reference to the accompanying drawings.

[0060] Embodiment 1

[0061] The present invention proposes a noisy label image classification method based on a recalibration strategy and Mixup loss, comprising the following steps:

[0062] Based on Mixup loss, two deep neural networks A and B with the same structure are pre-trained for several rounds using a dataset with noisy labels to obtain two deep neural network models with preliminary discrimination capabilities; the dataset includes images and noisy labels for marking image categories;

[0063] Use pure SSL of the dataset with noisy labels to train two deep neural network models with preliminary discrimination ability; in each round of pure SSL training, first input the entire dataset with noisy labels into deep neural network A and deep neural network B for data segmentation, deep neural network A uses GMM to segment the dataset with noisy labels to obtain the labeled subset and unlabeled subset for pure SSL training of deep neural network B, and deep neural network B uses GMM to segment the dataset with noisy labels to obtain the labeled subset and unlabeled subset for pure SSL training of deep neural network A; after every K rounds of pure SSL training, use the recalibration strategy based on CE loss to correct the optimization direction of deep neural network A and deep neural network B; until the maximum training round is reached, the trained deep neural network A and deep neural network B are obtained;

[0064] Use the trained deep neural network A or deep neural network B for image classification.

[0065] The training framework of the image classification method in the model training stage is shown in the figure Figure 1 As shown, the pseudo code of the training process is as follows:

[0066] Algorithm 1

[0067]

[0068] The specific implementation steps of the noisy label image classification method based on the recalibration strategy and Mixup loss in this embodiment are as follows:

[0069] 1 Pre-training process using Mixup loss

[0070] Using datasets with noisy labels Pre-train two deep neural networks A and B with the same structure, x i represents the i-th picture, represents the noisy label for marking the i-th image, n is the size of the dataset with noisy labels; the serial numbers of deep neural network A and deep neural network B are represented as m=0 and m=1 respectively.

[0071] Existing sample selection methods based on SSL technology, such as DivideMix, C2MT and DISC, mainly use a joint loss function with a penalty term and CE loss to train the model for several rounds in the warm-up phase to enable the model to obtain initial discrimination ability. The joint loss for the mth network can be expressed as formula (1).

[0072]

[0073] Where n∈{0,1} 1 Represents the control factor, which determines the penalty term L PL Whether to participate in pre-training. Since n needs to be dynamically adjusted according to the type of noise labels in the dataset, this greatly increases the difficulty of training and the adjustment time. Therefore, we introduce Mixup loss to replace the joint loss in formula (1). It is worth noting that we are the first to introduce Mixup loss for pre-training for the sample selection framework based on SSL technology. Mixup loss is expressed as:

[0074]

[0075] in, According to formula (3), i and any other input x j The mixed input is obtained by linear interpolation between yi and y j They are samples x i and x j One-hot encoded labels.

[0076]

[0077] Where b represents the input parameter of the beta distribution function. represents a random scalar in the interval [0,1] sampled from a beta distribution, l represents the interpolation weight, Represents the i-th image after mixing.

[0078] 2 Pure SSL Robust Training Process

[0079] The SSL training process follows a similar process to DivideMix. It mainly consists of two parts: GMM data partitioning and robust training of the two partitioned subsets using SSL technology. For the mth model, we use it to calculate The CE loss value of each sample in . Then, we use a two-component GMM to model the loss values ​​of all samples and estimate the probability that each sample is a noise label or a clean label. The output of GMM is

[0080]

[0081] g={1,2} 1 represents the g-th Gaussian term of the GMM model, Yes(x i ,y i ). For the mth network, It means (x i ,y i ) corresponding to the loss The probability of belonging to the Gaussian term g = 1, It means (x i ,y i ) corresponding to the loss The probability of belonging to the Gaussian term g = 2 follows the small loss criterion, that is, samples with smaller losses are more likely to contain clean labels. We choose the output probability of the Gaussian term with the smallest mean as the probability that each sample contains a clean label. Therefore, the input pair (x i ,y i ) is the probability of clean data can be expressed as

[0082]

[0083] In the formula, Indicates the selection of mean m (g) The smallest Gaussian term, j represents the number of the selected Gaussian term, m (g) is the estimated mean of the g-th Gaussian term. Represents the input pair (x i ,y i ) is the probability of clean data, based on the estimated probability We can use equation (6) to transform the dataset with noisy labels into Divide into labeled subsets X (m) and the unlabeled subset U(m) .

[0084]

[0085] where t is a predefined threshold that controls the size of the labeled subset. After data partitioning, the two subsets of the mth model are used to train the remaining 1-m models, and vice versa. When training the mth model, we can use the labeled subset X (1-m) Get pseudo labels (Algorithm 1, lines 12-13):

[0086]

[0087] To further eliminate noisy labels. is the mth network input x i The output probability of the h-th weak enhancement.

[0088] We integrate the predictions of the two networks to jointly guess the true label of each unlabeled sample, that is, (1-m) Get pseudo labels (Lines 13-14 of Algorithm 1):

[0089]

[0090] After obtaining the unlabeled and labeled sets X (1-m) UU (1-m) After predicting the pseudo-labels, we perform a cooling operation (Lines 15-16 of Algorithm 1):

[0091]

[0092] In the formula, is the pseudo label vector We use the pseudo label Replace X (1-m) and U (1-m) The noisy labels observed in Then, the Mixup operation is performed on the unlabeled and labeled data sets: for two random input pairs We perform the Mixup operation through equation (3) and then obtain the mixed input pair in (Algorithm 1, lines 17-18). Finally, the loss of the pure SSL robust training process is expressed as (line 19 of Algorithm 1):

[0093]

[0094] Among them, L x represents the supervision loss, L u Denotes the unsupervised loss L regrepresents the regularization term, and a is the coefficient that controls the contribution of unlabeled samples to the model update.

[0095] 3 SSL training process with recalibration strategy

[0096] Existing methods continue pure SSL training after the warm-up phase. Considering the limitations of sample partitioning and the loss of some label information due to missing labels in the SSL process, we propose a recalibration strategy. After every K rounds of pure SSL training, we correct the training direction of the deep neural network model. This recalibration process is achieved by performing a single supervised training on all samples using a simple CE loss, as shown in Equation (11).

[0097]

[0098] Where, L RC represents the recalibration loss.

[0099] By periodically allowing the model to learn the label information and feature information of all samples, this approach helps avoid the bias introduced by some samples that have always been classified as unlabeled, and enhances the generalization and classification performance of the model.

[0100] 4 Use the trained model for image classification tasks

[0101] Use the trained deep neural network A or deep neural network B for image classification tasks.

[0102] Embodiment 2

[0103] The present invention proposes a noisy label image classification system based on a recalibration strategy and Mixup loss corresponding to the method of embodiment 1, comprising:

[0104] A pre-training module, used to perform several rounds of pre-training on two deep neural networks A and B with the same structure using a data set with noisy labels based on Mixup loss to obtain two deep neural network models with preliminary discrimination capabilities; the data set includes pictures and noisy labels for marking picture categories;

[0105] The formal training module is used to train two deep neural network models with preliminary discrimination ability using pure SSL of a dataset with noisy labels. In each round of pure SSL training, the entire dataset with noisy labels is first input into deep neural network A and deep neural network B for data segmentation. The labeled subset and unlabeled subset obtained by deep neural network A using GMM to segment the dataset with noisy labels are used for pure SSL training of deep neural network B. The labeled subset and unlabeled subset obtained by deep neural network B using GMM to segment the dataset with noisy labels are used for pure SSL training of deep neural network A. After every K rounds of pure SSL training, the optimization direction of deep neural network A and deep neural network B is corrected using a recalibration strategy based on CE loss. This is done until the maximum number of training rounds is reached, and the trained deep neural network A and deep neural network B are obtained.

[0106] The image classification module is used to perform image classification using the trained deep neural network A or deep neural network B.

[0107] The implementation methods of each module and module function in the system are completely consistent with the steps of the method in Example 1, so they will not be repeated here.

[0108] The effectiveness of the present invention is demonstrated by experimental results below.

[0109] In Table 1, we compare the test accuracy of our proposed method with the SOTA methods on CIFAR-10 and CIFAR-100 datasets with symmetric noisy labels. Here, "Best" represents the highest accuracy achieved in all training rounds, while "Last" represents the average accuracy of the last 10 rounds. When the "Last" and "Best" values ​​are close, it indicates that the method remains robust to noisy labels throughout the training process. As shown in Table 1, except for slightly inferior to DivideMix in 80% of the symmetric noise scenes, our method achieves the best performance in almost all symmetric noise scenes. It is worth noting that although our method is based on the DivideMix framework, the introduced Mixup loss and recalibration strategy not only significantly improve its performance, but also surpass other recent advanced methods. For example, in the scenario with 20% symmetrical noise on both datasets, SMR outperforms DivideMix, MOIT+, DISC, and TCL by 0.28% / 2.0%, 2.2% / 3.3%, 0.28% / -, and 1.38% / 1.24%, respectively. As the noise ratio increases, the performance gap becomes more obvious. For example, in the scenario with 50% symmetrical noise, SMR outperforms these methods by 0.44 / 0.8%, 3.0% / 5.4%, 0.06% / -, and 1.14% / 2.1%, respectively.

[0110] Table 1 Test accuracy (%) on CIFAR-10 and CIFAR-100 with symmetric label noise

[0111]

[0112] The best accuracy in the table is indicated in bold.

[0113] In Table 2, we compare the test accuracy of our proposed method with the SOTA methods on CIFAR-10 with asymmetric noisy labels. Our method achieves the best performance on the 40% asymmetric noise scenario. DivideMix… reports the results of DivideMix on this scenario for us by running the public code. Although our method results are close to those reported in the original DivideMix paper, our results exceed the reproduction results. Moreover, our method still outperforms other SOTA methods. Similar to the comparison results in Table 1, introducing Mixup loss and recalibration strategy to DivideMix not only significantly improves its performance against its predecessor, but also surpasses other recent state-of-the-art methods. For example, on the 40% asymmetric noise scenario on CIFAR-10, SMR outperforms DivideMix, MOIT, ECMB, C2MT, and TCL by 0.5%, 0.1%, 0.4%, 0.6%, and 0.8%, respectively.

[0114] Table 2 Test accuracy (%) on CIFAR-10 and CIFAR-100 with asymmetric label noise

[0115]

[0116]

[0117] The best accuracy in the table is indicated in bold.

[0118] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A classification method for noisy labeled images based on recalibration strategy and Mixup loss, It is characterized in that The following steps are involved: Based on Mixup loss, two deep neural networks A and B with the same structure are pre-trained for several rounds using a dataset with noisy labels to obtain two deep neural network models with preliminary discrimination capabilities; the dataset includes images and noisy labels for marking image categories; Use pure SSL of the noisy labeled dataset to train two deep neural network models with preliminary discrimination capabilities; in each round of pure SSL training, first input the entire noisy labeled dataset into deep neural network A and deep neural network B for data segmentation, deep neural network A uses GMM to segment the noisy labeled dataset to obtain the labeled subset and unlabeled subset for pure SSL training of deep neural network B, and deep neural network B uses GMM to segment the noisy labeled dataset to obtain the labeled subset and unlabeled subset for pure SSL training of deep neural network A; After every K rounds of pure SSL training, the optimization direction of deep neural network A and deep neural network B is corrected using a recalibration strategy based on CE loss until the maximum number of training rounds is reached, and the trained deep neural network A and deep neural network B are obtained; Use the trained deep neural network A or deep neural network B for image classification.

2. The noisy label image classification method based on the recalibration strategy and Mixup loss as claimed in claim 1, It is characterized in that The pre-training is specifically as follows: Using datasets with noisy labels Pre-train two deep neural networks A and B with the same structure, x i represents the i-th picture, represents the noisy label for marking the i-th image, n is the size of the dataset with noisy labels; the serial numbers of deep neural network A and deep neural network B are represented as m=0 and m=1 respectively; The Mixup loss is specifically: Where, L mix represents Mixup loss, l represents interpolation weight, L CE (·) represents CE loss, represents the predicted probability of the mth deep neural network output, Indicates that in the i-th image x i and any other images x j The i-th mixed input image obtained by linear interpolation between y i and j are the i-th picture x i and the jth image x j One-hot encoding label of ; Interpolation weights λ and mixed input images The calculation formula is as follows: Among them, β represents the input parameter of the beta distribution function, Beta(β,β) represents the beta distribution function, Represents a random scalar in the interval [0,1] sampled from a beta distribution.

3. The noisy label image classification method based on the recalibration strategy and Mixup loss as claimed in claim 1, It is characterized in that The GMM segmentation of the data set with noisy labels is specifically: Use a two-component GMM to model the loss values ​​of all samples and estimate the probability of each sample being a noise label or a clean label. The formula is as follows: Where g = {1,2} 1 represents the g-th Gaussian term of the GMM model, is the i-th picture x i Standardized CE loss of ; Represents the i-th picture x i The probability that the standardized CE loss belongs to the first Gaussian term, Represents the i-th picture x i The probability that the standardized CE loss belongs to the second Gaussian term; The output probability of the Gaussian term with the smallest mean is selected as the probability that each image contains a clean label. The input pair (x i ,y i ) is the probability of clean data as: In the formula, represents the probability that the label of the i-th image is a clean label, st represents the constraint condition, Select the mean value m (g) The smallest Gaussian term, j represents the number of the selected Gaussian term; m (g) is the estimated mean of the g-th Gaussian term; Based on estimated probability With a predefined probability threshold, the mth deep neural network transforms the dataset with noisy labels using the following equation Divide into labeled subsets X (m) and the unlabeled subset U (m) : Where n is the size of the data set, t is the predefined probability threshold, represents the probability that the label of the i-th image is a clean label, is a dataset with noisy labels. It's a picture x i and noise labels The sample pairs composed of Considering that the noise label is the true label, the sample pair Put it into the labeled subset, otherwise, it is considered a noise label It is not the real label. After removing the noise label, the image x i into the unlabeled subset.

4. The noisy label image classification method based on the recalibration strategy and Mixup loss as claimed in claim 1, It is characterized in that During the pure SSL training of deep neural network B and the pure SSL training of deep neural network A, the following formula is used to obtain the pseudo labels of the labeled subset to further eliminate the noise labels: In the formula, represents the pseudo-labels for the labeled subset, represents the probability that the label of the ith image is a clean label determined by the 1-mth deep neural network, represents the noise label of the i-th image, is the i-th image x input by the m-th deep neural network i The predicted probability of the h-th weak enhancement output; Integrate the probabilities predicted by the two deep neural networks to jointly guess the true label of each unlabeled sample and obtain the pseudo labels of the unlabeled subset: In the formula, represents the pseudo-labels of the unlabeled subset, is the i-th image x input by the m-th deep neural network i The predicted probability of the output after the h-th weak enhancement; is the i-th image x input by the 1-m deep neural network i The predicted probability of the output after the h-th weak enhancement; After obtaining the pseudo labels of the labeled subset and the pseudo labels of the unlabeled subset, a cooling operation is performed on the pseudo labels: In the formula, represents the pseudo label after cooling; is the pseudo label vector The jth component of ; c represents the number of image categories; Use pseudo labels after cooling Replace the pseudo labels of the labeled subset and the pseudo labels of the unlabeled subset, and then and the image x corresponding to the label i Composition input pair Perform Mixup operation on the input pair to obtain a mixed input pair is a mixed input image, are the mixed input pseudo labels used to compute the loss for pure SSL training.

5. The noisy label image classification method based on the recalibration strategy and Mixup loss as claimed in claim 4, It is characterized in that The loss of pure SSL training is specifically: Where, L SSL represents the loss of pure SSL training, L x represents the supervision loss, L u represents the unsupervised loss, L reg represents the regularization term, a is the coefficient that controls the contribution of unlabeled samples to the update of the deep neural network model; X (1-m) and U (1-m) are the labeled subsets and unlabeled subsets divided by the 1st to mth deep neural networks, x i is the i-th picture, represents the predicted probability of the mth deep neural network output, Represents the predicted probability that the image category output by the mth deep neural network belongs to category c.

6. The noisy label image classification method based on the recalibration strategy and Mixup loss as claimed in claim 1, It is characterized in that The recalibration strategy is implemented by performing a single supervised training using CE loss on all samples in the dataset with noisy labels; The recalibration strategy corrects the recalibration loss of the optimization direction process of deep neural network A and deep neural network B as follows: Where, L RC represents the recalibration loss, n is the size of the dataset with noisy labels, represents the noise label for marking the i-th image, p (m) (x i ) represents the predicted probability of the output of the m-th deep neural network.

7. A noisy label image classification system based on recalibration strategy and Mixup loss, It is characterized in that include: A pre-training module, used to perform several rounds of pre-training on two deep neural networks A and B with the same structure using a data set with noisy labels based on Mixup loss to obtain two deep neural network models with preliminary discrimination capabilities; the data set includes pictures and noisy labels for marking picture categories; The formal training module is used to train two deep neural network models with preliminary discrimination capabilities using pure SSL of the noisy labeled dataset. In each round of pure SSL training, the entire noisy labeled dataset is first input into deep neural network A and deep neural network B for data segmentation. The labeled subset and unlabeled subset obtained by deep neural network A using GMM to segment the noisy labeled dataset are used for pure SSL training of deep neural network B. The labeled subset and unlabeled subset obtained by deep neural network B using GMM to segment the noisy labeled dataset are used for pure SSL training of deep neural network A. After every K rounds of pure SSL training, the optimization direction of deep neural network A and deep neural network B is corrected using a recalibration strategy based on CE loss until the maximum number of training rounds is reached, and the trained deep neural network A and deep neural network B are obtained; The image classification module is used to perform image classification using the trained deep neural network A or deep neural network B.