A noisy label image classification method based on sample hierarchical learning and correction

Abstract

The application relates to a noisy label image classification method based on sample hierarchical learning and correction, which comprises the following steps: dividing an image dataset into a simple clean sample subset, a simple noise sample subset, a difficult clean sample subset and a difficult noise sample subset by using a sample hierarchical strategy; correcting the labels of part of image samples in the simple noise sample subset based on a prediction consistency strategy; correcting the labels of part of image samples in the difficult noise sample subset based on the prediction consistency strategy; constructing a supervised training loss function according to the simple clean sample subset and the difficult clean sample subset; constructing an unsupervised training loss function according to the simple noise sample subset and the difficult noise sample subset; linearly weighting the supervised training loss function and the unsupervised training loss function to obtain a total loss function, training an image classification model, and classifying a to-be-tested image by using the trained image classification model. The application can improve the classification performance of the model.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of image classification, and particularly relates to a noisy label image classification method based on sample hierarchical learning and correction. BACKGROUND

[0002] Image classification task is one of the core tasks in the field of computer vision, and in recent years, it has made significant progress due to the rapid development of deep learning technology. However, to achieve high-precision image classification, it depends on large-scale and accurately labeled data sets. However, in reality, it is difficult and expensive to obtain such data sets. Therefore, collecting large-scale data from Internet applications has become an alternative solution, but such data often contains a large number of noisy labels, i.e. incorrect labels.

[0003] Although deep neural networks have strong fitting ability and can fit any data, including completely random noise data, the noisy labels in the training data set will cause the model to learn incorrect feature representation, which will seriously damage the model performance. To solve this problem, existing methods mainly use sample selection and correction strategies, which distinguish between clean data and noisy data, and perform supervised learning on the former and unsupervised or semi-supervised learning on the latter. However, these methods often ignore the difficulty of samples and simply use small loss criteria to distinguish between clean and noisy samples, which may lead to simple noisy samples and difficult clean samples being misclassified, and thus causing more serious overfitting phenomenon in the label correction process. SUMMARY

[0004] In order to solve the problems in the background art, solve the problem of ignoring the difficulty of samples in the sample selection and correction process in the prior art, and improve the classification performance of the model, the application provides a noisy label image classification method based on sample hierarchical learning and correction, comprising:

[0005] S1: obtaining an image data set with noisy labels and constructing an image classification model;

[0006] S2: training the image classification model, and classifying the test image through the trained image classification model, wherein the training of the image classification model comprises:

[0007] S21: preheating training the image classification model to enable the image classification model to acquire feature extraction ability and feature classification ability;

[0008] S22: inputting the image data set into the image classification model, calculating the JS divergence between the prediction result of each image sample and the actual label, dividing the image data set into a small-loss clean subset and a large-loss noise subset, and recording the prediction result of each image sample;

[0009] S23: according to the distance of each image sample in the nearest consistency of the prediction results recorded in the training round divides the image dataset into a consistency clean subset and an inconsistency noise subset;

[0010] S24: according to the small-loss clean subset, the large-loss noise subset, the consistency clean subset and the inconsistency noise subset, the image dataset is divided into a simple clean sample subset, a simple noise sample subset, a difficult clean sample subset and a difficult noise sample subset by using a sample grading strategy;

[0011] S25: if the current training round is greater than or equal to , the labels of part of the image samples in the simple noise sample subset are corrected based on a prediction consistency strategy, and are transferred to the simple clean sample subset;

[0012] S26: if the current training round is greater than or equal to , the labels of part of the image samples in the difficult noise sample subset are corrected based on a prediction consistency strategy, and are transferred to the difficult clean sample subset;

[0013] S27: a supervised training loss function is constructed according to the updated simple clean sample subset and the difficult clean sample subset; an unsupervised training loss function is constructed according to the simple noise sample subset and the difficult noise sample subset; the supervised training loss function and the unsupervised training loss function are linearly weighted to obtain a total loss function, and the image classification model is trained, and steps S21-S27 are repeatedly executed until a preset training round is reached, and a trained image classification model is obtained.

[0014] The present application has at least the following beneficial effects

[0015] The present application discloses a sample difficulty grading method, which accurately divides samples into simple clean samples, simple noise samples, difficult clean samples and difficult noise samples by comprehensively considering JS divergence and consistency of historical prediction results. This method effectively avoids the confusion of simple noise samples and difficult clean samples in the sample selection process. By constructing supervised and unsupervised training loss functions for different difficulty sample subsets, the model is provided with rich clean samples for learning, and difficult clean samples are also included, which helps the model to learn more accurate class boundaries and improve the ability to distinguish different classes. For simple noise samples and difficult noise samples, the present application sets different correction rounds and correctable confidence threshold values. Simple samples are corrected at the beginning of model learning, ensuring that the model is fully learned after the corrected simple samples, and then the difficult samples are corrected. This grading learning strategy makes the model learning process smoother, enhances the generalization ability of the model, and improves the accuracy of label correction. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 Figure 1 is a schematic diagram of a method according to the present application. DETAILED DESCRIPTION

[0017] Other advantages and novel features of the present application will become apparent from the following detailed description of the application when taken in conjunction with the announcements, as indicated above, and the working examples. Various modifications to the embodiments will be apparent to those skilled in the art, from this disclosure, which is intended in all respects to be illustrative rather than restrictive. Unless otherwise defined, scientific and technical terms used in connection with the present application shall have the meanings that are commonly understood by those of ordinary skill in the art. Further, unless otherwise required by context, singular terms shall include pluralities and vice versa. The description herein of any particular embodiment of the application is intended to apply to each and every embodiment of the application. The description of any particular embodiment of the application is intended to provide

[0018] Referring to Figure 1 , the present application provides a noisy label image classification method based on sample hierarchical learning and correction, comprising:

[0019] S1: obtaining an image dataset with noisy labels, and constructing an image classification model;

[0020] Preferably, the image classification model comprises a feature extractor , a feature projector and a classifier ; the feature extractor is used for feature extraction of samples, the classifier is used for classification of features extracted by the feature extractor , and the feature projector is used for dimension reduction of features extracted by the feature extractor to obtain low-dimensional features of samples.

[0021] In this embodiment, the image dataset with noisy labels contains images of multiple categories, but the labels may contain noise. An image classification model is constructed, which is composed of three main parts: a feature extractor , a feature projector and a classifier ; the feature extractor is used for feature extraction of input image samples to generate high-dimensional feature vectors; a convolutional neural network (CNN) can be used as the feature extractor. CNN can extract local features and global features in images through multiple layers of convolution and pooling operations, for example, a feature extractor of the ResNet series; the feature projector , which is used to reduce the dimensionality of the high-dimensional feature vectors generated by the feature extractor, resulting in low-dimensional feature vectors. This helps to reduce computational complexity while preserving key information, and can use a fully connected layer or dimensionality reduction techniques such as Principal Component Analysis (PCA), t-SNE, etc. as the feature projector. In this example, a fully connected layer is used for dimensionality reduction, and its output dimension can be adjusted according to actual needs. The classifier is used to classify the low-dimensional feature vectors generated by the feature projector, outputting the class label of the sample, and can use a fully connected layer plus a softmax function as the classifier. The softmax function can convert the output into a probability distribution, so as to obtain the predicted probability of each class.

[0022] S2: training the image classification model, and classifying the test image through the trained image classification model, wherein the training of the image classification model comprises:

[0023] S21: preheating training the image classification model to enable the image classification model to acquire feature extraction capability and feature classification capability;

[0024] Preferably, the preheating training of the image classification model comprises:

[0025] S211: freezing the classifier , and using the image dataset with noisy labels to obtain a label-free dataset by discarding the labels , wherein N represents the number of samples, represents the i-th sample, represents the label of the sample , and the feature extractor and the feature projector are trained in an unsupervised contrastive learning manner, the loss function is a contrastive loss, and the training is repeated for L rounds.

[0026] In this embodiment, strong enhancement and weak enhancement are used to construct positive and negative sample pairs. The strong enhancement sample and the original sample constitute a negative sample pair, while the weak enhancement sample and the strong enhancement sample (or the original sample, but the strong enhancement sample is usually selected to maintain consistency) constitute a positive sample pair. This helps the model to learn more discriminative feature representations. The feature extractor and the feature projector are trained using a contrastive loss function (such as InfoNCE loss).

[0027] S212: freezing the feature extractor and the feature projector , and training the classifier according to the image dataset with noisy labels ​​​Supervised training is performed, and a cross-entropy loss function is used for training rounds.

[0028] Preferably, the manner of using unsupervised contrastive learning is used to train the feature extractor and the feature projector Training includes:

[0029] The sample is respectively strongly enhanced and weakly enhanced to obtain a strongly enhanced sample and a weakly enhanced sample The strongly enhanced sample and the original sample constitute a negative example pair, and the weakly enhanced sample and the original sample constitute a positive example pair; a contrastive loss function is used to reduce the distance between the features of the positive example pair output by the feature projector and to increase the distance between the features of the negative example pair output by the feature projector.

[0030] In this embodiment, the contrastive loss function includes:

[0031] =

[0032]

[0033]

[0034]

[0035]

[0036]

[0037]

[0038]

[0039]

[0040] wherein, represents the contrastive loss function, represents a cosine similarity function; represents a temperature coefficient; represents an exponential function; represents the number of samples in the dataset D.

[0041] In this embodiment, the sample Strong enhancement includes: rotation, scaling, cropping, color jitter, etc. For example, the image is rotated by a certain angle, such as ±30 degrees, ±45 degrees, etc. to simulate a wider range of image changes. Weak enhancement includes slight rotation, scaling, cropping, or color adjustment, rotating the image slightly by a certain angle, such as ±5 or ±10 degrees. The purpose of weak enhancement is to increase data diversity without significantly changing the image content, while avoiding the introduction of excessive noise or interference information.

[0042] S22: Input the image dataset into the image classification model, calculate the JS divergence between the predicted result and the actual label of each image sample, divide the image dataset into a small loss clean subset and a large loss noisy subset, and record the prediction result of each image sample;

[0043] Preferably, calculating the JS divergence between the prediction result and the actual label of each image sample includes:

[0044]

[0045] in, Indicates the The JS divergence of samples, represents the divergence function, is the KL divergence function; The output of the classifier The prediction results of image samples, Indicates the The actual labels of the image samples.

[0046] Preferably, the method of dividing the image dataset into a small loss clean subset and a large loss noise subset comprises: traversing each category of the image dataset, sorting the image samples under each category from small to large according to their JS divergence, and selecting the top The proportion of samples is added to the small loss clean subset, and the rest of the samples are added to the large loss noise subset. After traversing all categories, the small loss clean subset and the large loss noise subset are obtained. Among them, The calculation process is as follows;

[0047] S221: Calculate the divergence threshold based on the JS divergence of all image samples in the image dataset:

[0048]

[0049] in, Indicates the Divergence threshold at the round iteration, Represents the average value of JS divergence of all image samples in the image dataset; represents the minimum value of the JS divergence of all image samples JS in the image data set; represents that the filter coefficient takes a value of 0.3-0.6; represents that the adjustment threshold takes a value of 0.5.

[0050] S222: The proportion of image samples in the image data set with a JS divergence value less than the divergence threshold in the image data set is taken as the consistency of the image data set. .

[0051] In this embodiment, by dividing the data set into a small-loss clean subset and a large-loss noise subset, the model can focus more on learning clean and high-quality data, thereby improving its generalization ability and accuracy. The small-loss clean subset usually contains samples that are easier to classify, which helps the model learn more meaningful feature representations. The large-loss noise subset contains samples with more noise or difficult to classify, which may have a negative impact on the training of the model. By separating these samples, the interference of these samples on the training of the model can be reduced, thereby improving the stability and robustness of the model.

[0052] S23: Dividing the image data set into a consistency clean subset and an inconsistency noise subset according to the consistency of the prediction results recorded in the last round of training of each image sample;

[0053] Preferably, the step S23 comprises:

[0054] S231: Adding image samples with the same prediction label recorded in the last round of training to the consistency clean subset as consistency clean samples, and defining the consistency of the consistency clean samples as ;

[0055] S232: Adding image samples with different prediction labels recorded in the last round of training to the inconsistency noise subset as inconsistency noise samples, and defining the consistency of the inconsistency noise samples as ​In this embodiment, the consistent clean subset generally contains image samples that are easier to classify and have more obvious features. These samples have stable prediction results in multiple training, which helps the model learn more robust feature representation. By focusing on training these consistent samples, the model can reduce its dependence on noisy data, thereby improving its generalization ability on new data. The inconsistent noise subset contains image samples whose prediction results are unstable in multiple training. These samples may be caused by data annotation errors, poor image quality, or insufficient model capabilities, etc. By separating these samples, we can further analyze the source of noisy data and take appropriate measures to handle it, such as data cleaning, enhancing model capabilities, etc.

[0056] S24: Dividing the image dataset into simple clean sample subset, simple noise sample subset, difficult clean sample subset and difficult noise sample subset according to the sample grading strategy based on the small loss clean subset, the large loss noise subset, the consistent clean subset and the inconsistent noise subset;

[0057] Preferably, the step S24 comprises:

[0058] The intersection of the consistent clean subset and the small loss clean subset is taken as the simple clean sample subset; these samples are easy to classify and have consistent classification results.

[0059] The difference set of the small loss clean subset and the consistent clean subset is taken as the simple noise sample subset; these samples are easy to classify, but the classification results are inconsistent, which may be noise samples.

[0060] The intersection of the consistent clean subset and the large loss noise subset is taken as the difficult clean sample subset; these samples are difficult to classify, but the classification results are consistent, which may be boundary samples or complex samples.

[0061] The samples remaining in the image dataset after the simple clean sample subset, the simple noise sample subset, and the difficult clean sample subset are taken as the difficult noise sample subset; these samples are difficult to classify and have inconsistent classification results, which are more likely to be noise samples.

[0062] In this embodiment, by dividing the dataset into four subsets, the model can focus more on learning the features of simple clean samples and difficult clean samples, thereby improving the overall performance. For simple noise samples and difficult noise samples, further processing can be performed to reduce the impact of noise on model training.

[0063] S25: If the current training round is greater than or equal to , then correct the labels of part of the image samples in the simple noise sample subset based on the prediction consistency strategy and move them to the simple clean sample subset;

[0064] Preferably, the prediction consistency-based strategy corrects the label of part of the image samples in the simple noisy sample subset and transfers them to the simple clean sample subset comprises:

[0065] S251: Correct the labels of the image samples in the simple noisy sample subset and transfer them to the simple clean sample subset ; ;

[0066] S252: Input the strongly augmented samples and the weakly augmented samples into the image classification model respectively to obtain the corresponding prediction results and ;

[0067] S253: Determine whether the sample is a correctable sample according to the consistency of the prediction results of the strongly augmented sample and the weakly augmented sample . The sample is a correctable sample only when the following conditions are met. The label of the sample is corrected with the label , and the sample is transferred to the simple clean sample subset. The conditions are as follows:

[0068] Condition 1: and , where is the confidence of obtaining the maximum probability class in the prediction result, is the confidence threshold of simple sample correction, and is 0.95;

[0069] Condition 2: , is the function;

[0070] Condition 3: , represents the consistency of the predicted label recorded in the last round of training of the image sample . When consistent , when inconsistent .

[0071] ​In this example, by correcting incorrect labels in a subset of simple noise samples, the model is provided with more correct training samples, helping it learn more accurate feature representations and classification decisions. By performing prediction consistency checks on the enhanced images, reliable samples can be further screened and incorrect labels corrected, enhancing the model's robustness to noise and variation.

[0072] S26: If the current number of training rounds is greater than or equal to , based on the prediction consistency strategy, the labels of some image samples in the difficult noise sample subset are corrected and transferred to the difficult clean sample subset;

[0073] Preferably, the step of correcting labels of some image samples in the difficult noise sample subset based on the prediction consistency strategy and transferring the labels to the difficult clean sample subset comprises:

[0074] S261: Subset difficult noise samples Image samples in Perform strong enhancement and weak enhancement respectively to obtain strong enhancement samples and weakly enhanced samples ;

[0075] S262: Strongly enhance the sample and weakly enhanced samples Input the image classification model separately to obtain the corresponding prediction results and ;

[0076] S263: Based on strong enhancement samples and weakly enhanced samples Consistency judgment sample of the prediction results Is it a correctable sample? If and only if the following conditions are met, the sample For correctable samples, use labels Correction sample The label and the sample Transfer to the difficult clean sample subset, the conditions are as follows:

[0077] Condition 1: and ,in, The function is to obtain the confidence of the maximum probability category in the prediction result. The confidence threshold for difficult correction samples is 0.9;

[0078] Condition 2: , express function;

[0079] Condition 3: , Represents an image sample In recent The consistency of the predicted labels recorded in the round of training, when consistent , when inconsistent .

[0080] In this example, we correct incorrect labels due to labeling errors or misjudgments in the model's initial stages. By focusing on samples that are difficult to classify (i.e., hard samples), we make the model more sensitive to these samples, thereby improving its overall robustness and accuracy. By transferring these corrected samples to the hard and clean sample subset, subsequent training can be more efficient and targeted.

[0081] S27: Construct a supervised training loss function based on the updated simple clean sample subset and the difficult clean sample subset; construct an unsupervised training loss function based on the simple noise sample subset and the difficult noise sample subset; linearly weight the supervised training loss function and the unsupervised training loss function to obtain a total loss function to train the image classification model, and repeat steps S21 to S27 until the preset number of training rounds is reached to obtain a trained image classification model.

[0082] Preferably, the total loss function includes:

[0083]

[0084]

[0085]

[0086] in, represents the total loss function; and represents the weight parameter; represents the supervised training loss function; represents the unsupervised training loss function; Represents a simple and clean subset The number of samples in represents the cross entropy loss function; and represents the weight parameter; Represents a difficult clean subset The number of samples in represents a simple noise subset The number of samples in Representing difficult noise subsets The number of samples in Represents a simple and clean subset The a simple clean sample; a label representing a sample ; a first difficult clean sample in a difficult clean subset ; a label representing a sample ; a first simple noise sample in a simple noise subset ; a contrastive loss function; a strongly augmented sample of a sample ; a weakly augmented sample of a sample ; a strongly augmented sample of a sample ; a weakly augmented sample of a sample ; a strongly augmented sample of a sample ;

[0087] In this embodiment, by focusing on supervised learning of clean samples, the model can learn the correct classification boundary. At the same time, by processing noise samples through unsupervised learning, the model can learn a more robust feature representation, thereby reducing sensitivity to incorrect labels. Since the model processes both simple and difficult samples, as well as correct and incorrect labels during training, it is more likely to remain robust to various input changes. Through unsupervised methods such as contrastive learning, the model can learn the similarity between samples, which helps it identify and correct samples with incorrect labels. Through the divide-and-conquer strategy, the model can more effectively utilize training data, thereby possibly reducing the training time and resources required to achieve the same performance level.

[0088] In this embodiment, the weight parameters and are represented as:

[0089]

[0090]

[0091] wherein, is the base weight coefficient with a value of 0.5, is the adjustment weight coefficient with a value of 0.2, is the total number of training rounds, is the current training round. In this embodiment, by introducing dynamic weight parameters and , the model pays more attention to the difficult clean sample subset at the beginning of training.This helps the model learn more challenging feature representations. As training progresses, the model gradually shifts its focus to the subset of easy and clean samples to ensure good performance on easy-to-classify samples as well. This strategy helps balance the model's learning on difficult and easy samples, improving its generalization ability. Since the model can dynamically adjust its focus on different sample subsets during training, it is more likely to learn the correct classification boundary and reduce the error rate on the test set. The model pays more attention to difficult samples at the beginning of training, which allows it to learn important feature representations more quickly. As training progresses, the model gradually shifts its focus to easy samples, which helps consolidate the knowledge learned and further improve the model's accuracy, helping to speed up the training process.

[0092] The contrastive loss function in this embodiment includes:

[0093] =

[0094]

[0095]

[0096]

[0097]

[0098]

[0099]

[0100]

[0101]

[0102] Similarly:

[0103] =

[0104]

[0105]

[0106]

[0107]

[0108]

[0109]

[0110]

[0111]

[0112] wherein, represents a contrast loss function, represents a cosine similarity function; represents a temperature coefficient; represents an exponential function.

[0113] In summary, the present application provides a sample difficulty grading method, which accurately divides samples into simple clean samples, simple noise samples, difficult clean samples and difficult noise samples by considering the consistency of JS divergence and historical prediction results. This method effectively avoids the confusion of simple noise samples and difficult clean samples in the sample selection process. By constructing supervised and unsupervised training loss functions for different difficulty sample subsets, the model is provided with a rich set of clean samples for learning, and difficult clean samples are also included, which helps the model to learn more accurate class boundaries and improve the ability to distinguish different classes. For simple noise samples and difficult noise samples, different correction rounds and correctable confidence thresholds are set. Simple samples are corrected at the beginning of model learning to ensure that the model is fully learned after the corrected simple samples, and then the difficult samples are corrected. This grading learning strategy makes the model's learning process smoother, enhances the model's generalization ability, and improves the accuracy of label correction.

[0114] Finally, it should be pointed out that the above examples are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the purpose and scope of the present technical solutions, which should be covered in the scope of the claims of the present application.

Claims

1. A noisy label image classification method based on sample hierarchical learning and correction, characterized in that, The method comprises the following steps: S1: obtaining an image dataset with noise labels, and constructing an image classification model; S2: training the image classification model, and classifying test images through the trained image classification model, wherein the training of the image classification model comprises: S21: preheating the image classification model to obtain feature extraction and feature classification capabilities of the image classification model; S22: inputting the image dataset into the image classification model, calculating JS divergence between a prediction result and an actual label of each image sample, dividing the image dataset into a small-loss clean subset and a large-loss noise subset, and recording the prediction result of each image sample; S23: dividing the image dataset into a consistency clean subset and an inconsistency noise subset according to the consistency of the prediction results recorded in the training of each image sample in the wheel training. S24: dividing the image dataset into a simple clean sample subset, a simple noise sample subset, a difficult clean sample subset, and a difficult noise sample subset according to the small-loss clean subset, the large-loss noise subset, a consistent clean subset, and an inconsistent noise subset by using a sample grading strategy; The step S24 comprises: taking an intersection of the consistent clean subset and the small-loss clean subset as the simple clean sample subset; taking a difference set of the small-loss clean subset and the consistent clean subset as the simple noise sample subset; taking an intersection of the consistent clean subset and the large-loss noise subset as the difficult clean sample subset; taking the remaining samples in the image dataset after the simple clean sample subset, the simple noise sample subset, and the difficult clean sample subset as the difficult noise sample subset; S25: If the current number of training rounds is greater than or equal to , based on the prediction consistency strategy, the labels of some image samples in the simple noise sample subset are corrected and transferred to the simple clean sample subset; The step of correcting and transferring the labels of part of the image samples in the simple noise sample subset to the simple clean sample subset based on the prediction consistency strategy comprises: S251: image samples in the simple noise sample subset are respectively subjected to strong enhancement and weak enhancement to obtain strong enhancement samples and weak enhancement samples ;​ S252: inputting the strong enhancement sample and the weak enhancement sample into the image classification model respectively to obtain corresponding prediction results and the weak enhancement sample and ;​ S253: Based on strong enhancement samples and weakly enhanced samples Consistency judgment sample of the prediction results Is it a correctable sample? If and only if the following conditions are met, the sample For correctable samples, use labels Correction sample The label and the sample Move to a simple clean sample subset with the following conditions: Condition 1: and wherein, the function is a confidence of obtaining the maximum probability class in the prediction result, represents a confidence threshold of simple sample correction; Condition 2: , represents function; Condition 3: , representing image samples In recent consistency of the predicted labels recorded in the training rounds, consistent , inconsistent ; S26: If the current training round is greater than or equal to then correct the labels of part of the image samples in the difficult noisy sample subset based on the prediction consistency strategy, and transfer them to the difficult clean sample subset. The step of correcting and transferring the labels of part of the image samples in the difficult noise sample subset to the difficult clean sample subset based on the prediction consistency strategy comprises: S261: Subset difficult noise samples Image samples in Perform strong enhancement and weak enhancement respectively to obtain strong enhancement samples and weakly enhanced samples ; S262: inputting the strong enhancement sample and the weak enhancement sample into the image classification model respectively to obtain corresponding prediction results and the weak enhancement sample and ;​ S263: Based on strong enhancement samples and weakly enhanced samples Consistency judgment sample of the prediction results Is it a correctable sample? If and only if the following conditions are met, the sample For correctable samples, use labels Correction sample The label and the sample Transfer to the difficult clean sample subset, the conditions are as follows: Condition 1: and wherein, the function is a confidence of obtaining the maximum probability class in the prediction result, represents a confidence threshold of a difficult correction sample; Condition 2: , represents a function; Condition 3: , Indicating image samples In recent Consistency of predicted labels recorded in the training of the round, consistent , inconsistent ; S27: constructing a supervised training loss function according to the updated simple clean sample subset and the difficult clean sample subset, constructing an unsupervised training loss function according to the simple noise sample subset and the difficult noise sample subset, linearly weighting the supervised training loss function and the unsupervised training loss function to obtain a total loss function, training the image classification model, and repeating the steps S21-S27 until a preset training number is reached to obtain the trained image classification model. 2.The method of claim 1, wherein, The image classification model comprises a feature extractor , a feature projector and a classifier ; the feature extractor is configured to perform feature extraction on a sample, the classifier is configured to classify the features extracted by the feature extractor , and the feature projector is configured to reduce the dimension of the features extracted by the feature extractor to obtain low-dimensional features of the sample. 3.The method of claim 2, wherein, The step of calculating the JS divergence between the prediction result and the actual label of each image sample comprises: wherein, represents the JS divergence of the th sample, represents a divergence function, is a KL divergence function; represents the prediction result of the th image sample output by the classifier, represents the actual label of the th image sample.

4. The noisy label classification method based on sample hierarchical learning and correction according to claim 2, characterized in that, The dividing the image dataset into a small-loss clean subset and a large-loss noise subset includes: traversing each category of the image dataset, sorting image samples under each category according to their JS divergence from small to large, and selecting the top The proportion of samples is added to the small-loss clean subset, and the remaining samples are added to the large-loss noise subset. After traversing all categories, the divided small-loss clean subset and the large-loss noise subset are obtained. The calculation process is as follows. S221: calculating a divergence threshold according to the JS divergence of all image samples in the image dataset: wherein, denotes the number of iterations, denotes the divergence threshold value at the i-th iteration, denotes the average value of the divergence of all image samples JS in the image data set; denotes the minimum value of the divergence of all image samples JS in the image data set; denotes the filter coefficient; denotes the adjustment threshold value; S222: the proportion of image samples with a JS divergence value less than a divergence threshold value of the image dataset is taken as . .

5. The noisy label classification method based on sample hierarchical learning and correction according to claim 1, characterized in that, The step S23 comprises: S231: add the image sample whose predicted label recorded in the wheel training is same as the consistency clean sample to the consistency clean subset, and define the consistency of the consistency clean sample as ;​ S232: Will be in the near future Image samples with different predicted labels recorded in rounds of training are added as inconsistent noise samples to the inconsistent noise subset, and the consistency of inconsistent noise samples is defined as .

6. The noisy label image classification method based on sample hierarchical learning and correction according to claim 2, characterized in that, The total loss function comprises: wherein, represents the total loss function; and represents the weight parameter; represents the supervised training loss function; represents the unsupervised training loss function; represents the number of samples in the simple clean subset ; represents the cross-entropy loss function; and represents the weight parameter; represents the number of samples in the difficult clean subset ; represents the number of samples in the simple noisy subset ; represents the number of samples in the difficult noisy subset ; represents the th simple clean sample in the simple clean subset ; represents the label of sample ; represents the th difficult clean sample in the difficult clean subset ; represents the label of sample ; represents the th simple noisy sample in the simple noisy subset ; represents the contrastive loss function; represents the strong augmented sample of sample ; represents the weak augmented sample of sample ; represents the strong augmented sample of sample ; represents the weak augmented sample of sample .

Images