A noisy image classification method based on sample selection and label distribution learning

Abstract

The present invention discloses a noisy image classification method based on sample selection and label distribution learning, which relates to the field of computer vision technology and solves the technical problem that the current classification method is difficult to distinguish between difficult samples and noise samples, resulting in low accuracy of image classification. The method comprises: performing weak data enhancement and strong data enhancement on training samples respectively; constructing an image classification network, inputting the training samples into the image classification network to start training; establishing a global memorization queue of the image classification network; if the loss value of the training sample in the current training round is greater than the loss value of the nth stage, the training sample is marked as a noise sample; otherwise, the training sample is marked as a clean sample. The present invention can better distinguish between easy samples, difficult samples and noise samples, avoid classifying difficult-to-distinguish samples whose losses have not yet decreased as noise label samples in the early stage of training, and improve the accuracy of the model.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a noisy image classification method based on sample selection and label distribution learning. Background Art

[0002] Image classification is a basic task in the field of computer vision. Its core task is to assign a label to an image from a given set of categories. It involves assigning images to multiple predefined categories. Image classification has a wide range of applications in many fields, such as quality inspection and machine vision in industry, disease diagnosis and surgical assistance in medicine, and target detection and reconnaissance in the military.

[0003] In image classification tasks, the quality of image data is crucial to the performance of the model. Traditional image classification methods usually rely on training a large number of accurately labeled image samples. However, in practical applications, obtaining high-quality labeled samples often faces challenges in time and cost, and there are often noise samples in the labeled sample set. These noise samples not only affect the learning effect of the model, but may also lead to a significant decrease in classification accuracy.

[0004] In recent years, with the rapid development of deep learning technology, how to effectively process data with noisy samples has received more and more attention. Existing noise sample processing methods can be roughly divided into two categories: one is to use the minimum loss criterion to filter out clean samples, and the other is to filter out clean samples through the fluctuation strategy. However, since there are difficult samples in clean samples, these methods often misidentify difficult samples as noise samples, which still affects the accuracy of image classification.

[0005] In the process of implementing the present invention, the inventors found that there are at least the following problems in the prior art:

[0006] Current classification methods have difficulty distinguishing between difficult samples and noisy samples, resulting in low accuracy in image classification. Summary of the invention

[0007] The purpose of the present invention is to provide a noisy image classification method based on sample selection and label distribution learning to solve the technical problem that the current classification method in the prior art is difficult to distinguish between difficult samples and noise samples, resulting in low accuracy of image classification. The preferred technical solutions among the many technical solutions provided by the present invention can produce many technical effects as described below.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] The present invention provides a noisy image classification method based on sample selection and label distribution learning, comprising the following steps: performing weak data enhancement and strong data enhancement on training sample images respectively to obtain weakly enhanced images and strongly enhanced images; constructing an image classification network, inputting the weakly enhanced images and the strongly enhanced images into the image classification network, and setting the warm-up training rounds to t and the maximum training rounds to T, so that the image classification network starts training; establishing a global memorization queue for the image classification network, dividing the first training round to the current training round into n stages, and recording the loss value of the average prediction result on the weakly enhanced images and the strongly enhanced images in each stage; if the If the current training round is greater than the warm-up training round t and less than the maximum training round T, it is determined whether the loss values ​​of the weakly enhanced image and the strongly enhanced image in the current training round are both greater than the loss value of the nth stage; if the loss values ​​of the weakly enhanced image and the strongly enhanced image in the current training round are both greater than the loss value of the nth stage, the weakly enhanced image and the strongly enhanced image are marked as noise samples; if the loss values ​​of the weakly enhanced image and the strongly enhanced image in the current training round are not both greater than the loss value of the nth stage, the weakly enhanced image and the strongly enhanced image are marked as clean samples; wherein the clean samples include easy samples and difficult samples.

[0010] Optionally, if the loss values ​​of the weakly enhanced image and the strongly enhanced image in the current training round are not both greater than the loss values ​​of the nth stage, the weakly enhanced image and the strongly enhanced image are marked as clean samples, including: if the loss values ​​of the weakly enhanced image and the strongly enhanced image in the current training round are both less than the loss value of the nth stage, the weakly enhanced image and the strongly enhanced image are marked as easy samples; if the loss value of one of the weakly enhanced image and the strongly enhanced image in the current training round is less than the loss value of the nth stage, the weakly enhanced image and the strongly enhanced image are marked as difficult samples.

[0011] Optionally, after marking the weakly enhanced image and the strongly enhanced image as clean samples, the method further includes: using the total loss value of the easy samples and the difficult samples to perform a next round of training on the image classification network.

[0012] Optionally, the loss value of the average prediction result on the weakly enhanced image and the strongly enhanced image at each stage and the loss function of the loss value of the weakly enhanced image and the strongly enhanced image in the current training round are calculated. for: ;

[0013] in, is the true label of the weakly enhanced image or the strongly enhanced image; is the predicted label of the weakly enhanced image or the strongly enhanced image.

[0014] Optionally, a loss function is used to calculate the total loss value of the easy sample and the difficult sample. for:

[0015] ;

[0016] ;

[0017] ;

[0018] in, is an easy sample parameter; is the difficult sample parameter; is the loss function of the easy sample; is the total number of easy samples; is the loss function of the difficult sample; is the total number of difficult samples; is the weakly enhanced image of the i-th round; is the prediction result of the weakly enhanced image in the i-th round; is the strongly enhanced image of the i-th round; is the prediction result of the strongly enhanced image in the i-th round; q is a hyperparameter; is the loss value of the weakly enhanced image in the i-th round; is the loss value of the strongly enhanced image in the i-th round; is the true label of the weakly enhanced image or the strongly enhanced image in the i-th round.

[0019] Optionally, the average prediction result of the nth stage for:

[0020] ;

[0021] in, is the current training round; Representative The prediction results of all the weakly enhanced images and strongly enhanced images in the rounds; the average prediction results of the nth stage As soft labels for the weakly enhanced image and the strongly enhanced image.

[0022] Optionally, the method further includes: in the establishment of the global memorization queue of the image classification network, dividing the first training round to the current training round into n stages, and recording the loss value of the average prediction result of each stage on the weakly enhanced image and the strongly enhanced image, the average prediction result from the 1st stage to the n-1th stage for:

[0023] = ;

[0024] Among them, m∈[1,n-1].

[0025] Optionally, the weak data enhancement and strong data enhancement are respectively performed on the training sample images to obtain the weakly enhanced images and the strongly enhanced images, including: performing weak data enhancement on the training sample images by means of rescaling, random horizontal flipping, center cropping and ImageNet image standardization to obtain the weakly enhanced images; performing strong data enhancement on the training sample images by means of rescaling, random horizontal flipping, random vertical flipping, center cropping, RandAugmentwogeo data enhancement, and ImageNet image standardization to obtain the strongly enhanced images.

[0026] Optionally, the method further includes: if the current training round is less than or equal to the warm-up training round t, using a cross entropy loss function to perform warm-up training on the image classification network.

[0027] Optionally, the method further includes: if the current training round is equal to the maximum training round T, stopping the training to obtain a trained image classification network.

[0028] Implementing one of the above technical solutions of the present invention has the following advantages or beneficial effects:

[0029] The present invention maintains a global memorized queue through multi-stage training, distinguishes clean samples from noise samples according to the network's staged prediction results and the loss value of the current round of prediction results, and uses them for the next round of training. Compared with the traditional method of using the minimum loss criterion, it can better distinguish clean samples from noise samples, avoid classifying difficult samples whose loss is not yet reduced as noise samples in the early stage of training, and greatly improves the accuracy of image classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. It is obvious that the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0031] Figure 1 The present invention is a flowchart of a noisy image classification method based on sample selection and label distribution learning. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the present invention clearer, the various exemplary embodiments to be described below will refer to the corresponding drawings, which constitute a part of the exemplary embodiments, wherein various exemplary embodiments that may be used to implement the present invention are described. Unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementation methods described in the following exemplary embodiments do not represent all implementation methods consistent with the present disclosure. It should be understood that they are only examples of processes, methods, devices, etc. that are consistent with some aspects of the present disclosure as detailed in the attached claims, and other embodiments may also be used, or the embodiments listed herein may be modified in structure and function without departing from the scope and essence of the present invention.

[0033] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", etc. indicate the orientation or positional relationship based on the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the elements referred to must have a specific orientation, be constructed and operated in a specific orientation. The terms "first", "second", etc. are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. The term "multiple" means two or more. The terms "connected" and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, an integral connection, a mechanical connection, an electrical connection, a communication connection, a direct connection, an indirect connection through an intermediate medium, and can be the internal connection of two elements or the interaction relationship between two elements. The term "and / or" includes any and all combinations of one or more related listed items. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0034] In order to illustrate the technical solution of the present invention, a specific embodiment is used below for description, and only the parts related to the embodiment of the present invention are shown.

[0035] Embodiment 1:

[0036] like Figure 1As shown, the present invention provides a noisy image classification method based on sample selection and label distribution learning, comprising the following steps: S1, performing weak data enhancement and strong data enhancement on the training sample images respectively to obtain weakly enhanced images and strongly enhanced images; S2, constructing an image classification network, inputting the weakly enhanced images and the strongly enhanced images into the image classification network, and setting the warm-up training rounds to t and the maximum training rounds to T, so that the image classification network starts training; S3, establishing a global memorization queue of the image classification network, dividing the first training round to the current training round into n stages, and recording the average prediction results of each stage on the weakly enhanced images and the strongly enhanced images. loss value; S4, if the current training round is greater than the warm-up training round t, and less than the maximum training round T, then determine whether the loss values ​​of the weakly enhanced image and the strongly enhanced image in the current training round are both greater than the loss value of the nth stage; S5, if the loss values ​​of the weakly enhanced image and the strongly enhanced image in the current training round are both greater than the loss value of the nth stage, then mark the weakly enhanced image and the strongly enhanced image as noise samples; S6, if the loss values ​​of the weakly enhanced image and the strongly enhanced image in the current training round are not both greater than the loss value of the nth stage, then mark the weakly enhanced image and the strongly enhanced image as clean samples; wherein, clean samples include easy samples and difficult samples.

[0037] This embodiment maintains a global memorized queue through multi-stage training, distinguishes clean samples from noise samples based on the network's staged prediction results and the loss value of the current round of prediction results, and uses them for the next round of training. Compared with the traditional method of using the minimum loss criterion, it can better distinguish clean samples from noise samples, avoid classifying difficult samples whose loss is not yet reduced as noise samples in the early stage of training, and greatly improves the accuracy of image classification.

[0038] Next, combine Figure 1 The specific implementation steps of the noisy image classification method based on sample selection and label distribution learning provided in this embodiment are introduced in detail:

[0039] First, execute step S1, perform weak data enhancement and strong data enhancement on the training sample images, respectively, to obtain weakly enhanced images and strongly enhanced images. Before step S1, it is necessary to first construct an image dataset and divide the image dataset into a training sample image set and a verification image set; specifically, Food101N can be used as an image dataset. Food101N contains about 101,000 images of food recipes, which are classified into 101 categories. It aims to learn how to solve the label noise problem with minimal manual supervision. It is an ideal data for research and development of image classification in the presence of label noise. In this embodiment, Food101N is divided into a training sample image set and a verification image set at a ratio of 3:1. This ratio is not fixed and can be adjusted according to the size of the dataset, the complexity of the model, and the actual application scenario requirements to improve the performance and generalization ability of the network, thereby better solving the image classification problem.

[0040] As an optional implementation, in step S1, the training sample images are weakly enhanced by resizing, random horizontal flipping, center cropping, and ImageNet image standardization to obtain weakly enhanced images; the training sample images are strongly enhanced by resizing, random horizontal flipping, random vertical flipping, center cropping, RandAugmentwogeo data enhancement, and ImageNet image standardization to obtain strongly enhanced images. Further, the resizing size can be set to 256*256, and the center cropping size can be set to 224*224; in this step, the weakly enhanced image and the strongly enhanced image are marked as and Weak data enhancement is a small degree of transformation of the training sample image, which usually does not change the basic characteristics of the original data. By making small perturbations to the data, the network model can learn more local features, thereby improving network performance; strong data enhancement is a large degree of transformation of the training sample image, which usually changes some basic characteristics of the data to increase the diversity of the training data, thereby improving the robustness of the network model.

[0041] Then, execute step S2, construct an image classification network, input the weakly enhanced image and the strongly enhanced image into the image classification network, and set the warm-up training round to t and the maximum training round to T, such as setting the warm-up training round t=10 and the maximum training round T=100, so that the image classification network starts training. Using images from two perspectives to train the image classification network can improve the performance and generalization ability of the image classification network.

[0042] As an optional implementation, the image classification network in this embodiment uses ResNet50 as the base training network, and the initialization parameters choose to load the ImageNet pre-training parameters, and the parameters are not frozen. ResNet50 is a deep residual network. By introducing residual blocks, the gradient vanishing and network degradation problems in deep networks are solved, so that the network can be deeper and have better performance. The image classification network includes a feature extraction layer and a fully connected layer connected after the feature extraction layer; the output dimension of the fully connected layer is the same as the number of image categories, each neuron represents a category, and its output is the possibility that the input image belongs to this category. In this embodiment, the number of neurons in the fully connected layer is 101.

[0043] Next, execute step S3, establish a global memorization queue for the image classification network, divide the first training round to the current training round into n stages, and record the loss value of the average prediction result on the weakly enhanced image and the strongly enhanced image in each stage. This step also includes: using the average prediction result of the nth stage as the soft label of the weakly enhanced image and the strongly enhanced image; maintaining a multi-stage global memorization queue through training, and using the label learning method to replace the noise label with the soft label to reduce the impact of the noise label on the network performance.

[0044] Specifically, the average prediction result of the nth stage for:

[0045] ;

[0046] in, is the current training round; Representative The prediction results of all weakly enhanced images and strongly enhanced images in rounds are As soft labels for weakly enhanced images and strongly enhanced images.

[0047] The average prediction results from the 1st stage to the n-1th stage for:

[0048] = ;

[0049] Among them, m∈[1,n-1].

[0050] In this embodiment, n=4 is set, that is, the first training round to the current training round are divided into 4 stages, and the average prediction results from the first stage to the fourth stage are shown in formula (1) to formula (4) respectively:

[0051]

[0052]

[0053]

[0054]

[0055] in, Represents the average prediction result of the four stages. In the next round of training, As soft labels for weakly enhanced images and strongly enhanced images.

[0056] Further, it is determined whether the current training round is greater than the warm-up training round t. If not, the cross entropy loss function is used to warm up the image classification network, and then the next round of training is started by returning to step S3. In this embodiment, t=10, that is, the first 10 rounds of training adopt warm-up training; in the initial stage of training, the weights of the image classification network model will change rapidly, which may easily lead to premature overfitting of the model data and affect the generalization ability of the model. By setting the initial training stage to adopt warm-up training, a smaller learning rate can be used first, and then the learning rate can be gradually increased, which helps to preserve the stability of the distribution, slow down the overfitting phenomenon, and can also help the model find a better parameter space faster and accelerate the convergence of the model.

[0057] If so, execute step S4. If the current training round is greater than the warm-up training round t and less than the maximum training round T, determine whether the loss values ​​of the weakly enhanced image and the strongly enhanced image in the current training round are greater than the loss values ​​of the nth stage. In the subsequent training stage, the loss value of the average prediction result of the last stage is compared with the loss value of the prediction result of the current round to distinguish clean samples from noise samples, and is used for the next round of training. This can avoid classifying difficult samples whose loss has not yet decreased as noise label samples in the early stage of training, thereby improving the accuracy of the model.

[0058] Specifically, the loss value of the average prediction result on the weakly enhanced image and the strongly enhanced image at each stage and the loss function of the loss value of the weakly enhanced image and the strongly enhanced image in the current training round are calculated. for: ;

[0059] in, is the true label of the weakly enhanced image or the strongly enhanced image; is the predicted label of the weakly enhanced image or the strongly enhanced image; is the cross entropy loss function.

[0060] As in this embodiment, the average prediction result of the fourth stage is calculated The loss value is: , ;in, is the true label of the weakly enhanced image in the i-th round; i represents the current training round; It is the average prediction result of the weakly enhanced image in the fourth stage; It is the average prediction result of the weakly enhanced image in the fourth stage.

[0061] The loss values ​​of weakly enhanced images and strongly enhanced images in the current training round are calculated as follows: , ;in, is the prediction result of the weakly enhanced image in the i-th round; is the prediction result of the strongly enhanced image in the i-th round.

[0062] If the loss values ​​of the weakly enhanced image and the strongly enhanced image in the current training round are both greater than the loss values ​​in the nth stage, step S5 is executed to mark the weakly enhanced image and the strongly enhanced image as noise samples. When , the weakly enhanced image and the strongly enhanced image are marked as noise samples. The noise samples will no longer participate in the subsequent training.

[0063] If the loss values ​​of the weakly enhanced image and the strongly enhanced image in the current training round are not all greater than the loss values ​​of the nth stage, then execute step S6 to mark the weakly enhanced image and the strongly enhanced image as clean samples; wherein, clean samples include easy samples and difficult samples. It should be noted that noise samples refer to image samples with wrong labels or labels that do not match the actual categories in the image data set. They may be caused by human labeling errors or unclear concepts. For example, the concepts of two categories partially overlap, resulting in an image having two category attributes, but only being labeled as one category; clean samples refer to image samples with accurate labels; easy samples refer to clean samples that are easily identified; difficult samples refer to clean samples that are easily confused with noise samples and are not easily identified.

[0064] Specifically, step S6 includes: if the loss values ​​of the weakly enhanced image and the strongly enhanced image in the current training round are both less than the loss values ​​in the nth stage, the weakly enhanced image and the strongly enhanced image are marked as easy samples; that is, when When , the weakly enhanced images and strongly enhanced images are marked as easy samples.

[0065] If the loss value of one of the weakly enhanced image and the strongly enhanced image in the current training round is less than the loss value of the nth stage, the weakly enhanced image and the strongly enhanced image are marked as difficult samples. ,or When the weakly enhanced image and the strongly enhanced image are marked as difficult samples. This embodiment can better distinguish easy samples, difficult samples and noise samples by comparing the loss value of the average prediction results of each stage with the loss value of the prediction result of the current round, and avoid misclassifying difficult samples as noise samples, thereby ensuring the accuracy of the image classification network model.

[0066] After marking the weakly enhanced image and the strongly enhanced image as clean samples, the method of this embodiment further includes: returning to step S3, using the total loss value of the easy sample and the difficult sample to train the image classification network for the next round. The image classification network is trained for the next round using the total loss value of the easy sample and the difficult sample to further improve the performance of the network.

[0067] Specifically, the loss function for calculating the total loss value of easy samples and difficult samples is for:

[0068] ;

[0069] ;

[0070] ;

[0071] in, is the easy sample parameter, which indicates the weight of easy samples; is the difficult sample parameter, which indicates the weight of the difficult sample; is the loss function of easy samples; is the total number of easy samples; is the loss function of difficult samples; is the total number of difficult samples; is the weakly enhanced image of the i-th round; is the prediction result of the weakly enhanced image in the i-th round; is the strongly enhanced image of the i-th round; is the prediction result of the strongly enhanced image in the i-th round; q is a hyperparameter; is the loss value of the weakly enhanced image in the i-th round; is the loss value of the strongly enhanced image in the i-th round; is the true label of the weakly enhanced image or the strongly enhanced image in the i-th round. and They are the generalized cross entropy loss functions for easy samples and difficult samples respectively. Using the generalized cross entropy loss function of the previous round to train the next round can reduce the impact of noise samples on the image classification network and improve the generalization ability of the network.

[0072] As an optional implementation, this embodiment also adopts a multi-step learning rate decay strategy to decay the learning rate of the image classification network; after multiple simulation experiments, the warm-up training round t is set to 10, the maximum training round T is set to 100, and The model training effect is best when both are set to 1.0, the learning rate decay factor is set to 0.1, and the learning rate is decayed in the 50th and 80th rounds respectively.

[0073] If the current training round is equal to the maximum training round T, the training is stopped and the trained image classification network is obtained. The model weights of the trained image classification network can be reloaded into the newly constructed image classification network, and the prediction results of the network can be evaluated on the test set.

[0074] The embodiment is only a specific example and does not represent only one implementation mode of the present invention.

[0075] The above description is only the preferred embodiment of the present invention. It is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application belong to the protection scope of the present invention.

Claims

1. A noisy image classification method based on sample selection and label distribution learning, characterized in that: The following steps are involved: Perform weak data enhancement and strong data enhancement on the training sample images respectively to obtain weakly enhanced images and strongly enhanced images; Constructing an image classification network, inputting the weakly enhanced image and the strongly enhanced image into the image classification network, and setting the warm-up training rounds to t and the maximum training rounds to T, so that the image classification network starts training; Establishing a global memorization queue of the image classification network, dividing the first training round to the current training round into n stages, and recording the loss value of the average prediction result on the weakly enhanced image and the strongly enhanced image in each stage; If the current training round is greater than the warm-up training round t and less than the maximum training round T, determining whether the loss values ​​of the weakly enhanced image and the strongly enhanced image in the current training round are both greater than the loss value of the nth stage; If the loss values ​​of the weakly enhanced image and the strongly enhanced image in the current training round are both greater than the loss value in the nth stage, the weakly enhanced image and the strongly enhanced image are marked as noise samples; If the loss values ​​of the weakly enhanced image and the strongly enhanced image in the current training round are not all greater than the loss value in the nth stage, the weakly enhanced image and the strongly enhanced image are marked as clean samples; wherein the clean samples include easy samples and difficult samples; Calculate the loss value of the average prediction result on the weakly enhanced image and the strongly enhanced image at each stage, and the loss function of the loss value of the weakly enhanced image and the strongly enhanced image in the current training round for: ; in, is the true label of the weakly enhanced image or the strongly enhanced image; is a predicted label of the weakly enhanced image or the strongly enhanced image; The average prediction result of the nth stage for: ； in, is the current training round; Representative prediction results of all the weakly enhanced images and strongly enhanced images in rounds; n The average prediction result of the stage As soft labels for the weakly enhanced image and the strongly enhanced image; The method further comprises: in the step of establishing a global memorization queue of the image classification network, dividing the first training round to the current training round into n stages, and record the loss value of the average prediction result on the weakly enhanced image and the strongly enhanced image in each stage. n - Average prediction results of 1 stage for: = ； Among them, m∈[1,n-1].

2. The noisy image classification method based on sample selection and label distribution learning according to claim 1, characterized in that: If the loss values ​​of the weakly enhanced image and the strongly enhanced image in the current training round are not all greater than the loss value in the nth stage, marking the weakly enhanced image and the strongly enhanced image as clean samples includes: If the loss values ​​of the weakly enhanced image and the strongly enhanced image in the current training round are both smaller than the loss values ​​in the nth stage, the weakly enhanced image and the strongly enhanced image are marked as easy samples; If the loss value of one of the weakly enhanced image and the strongly enhanced image in the current training round is less than the loss value in the nth stage, the weakly enhanced image and the strongly enhanced image are marked as difficult samples.

3. The noisy image classification method based on sample selection and label distribution learning according to claim 1, characterized in that: After marking the weakly enhanced image and the strongly enhanced image as clean samples, the method further includes: The total loss value of the easy samples and the difficult samples is used to perform the next round of training on the image classification network.

4. The noisy image classification method based on sample selection and label distribution learning according to claim 1, characterized in that: Calculate the loss function of the total loss value of the easy sample and the difficult sample for: ； ； ； in, is an easy sample parameter; is the difficult sample parameter; is the loss function of the easy sample; is the total number of easy samples; is the loss function of the difficult sample; is the total number of difficult samples; is the weakly enhanced image of the i-th round; is the prediction result of the weakly enhanced image in the i-th round; is the strongly enhanced image of the i-th round; is the prediction result of the strongly enhanced image in the i-th round; q is a hyperparameter; is the loss value of the weakly enhanced image in the i-th round; is the loss value of the strongly enhanced image in the i-th round; is the true label of the weakly enhanced image or the strongly enhanced image in the i-th round.

5. The noisy image classification method based on sample selection and label distribution learning according to claim 1, characterized in that: The weak data enhancement and strong data enhancement are respectively performed on the training sample images to obtain the weakly enhanced images and the strongly enhanced images, including: performing weak data enhancement on the training sample images by means of size scaling, random horizontal flipping, center interception and ImageNet image standardization to obtain the weakly enhanced images; performing strong data enhancement on the training sample images by means of size scaling, random horizontal flipping, random vertical flipping, center interception, RandAugmentwogeo data enhancement, and ImageNet image standardization to obtain the strongly enhanced images.

6. The noisy image classification method based on sample selection and label distribution learning according to claim 1, characterized in that: The method further includes: if the current training round is less than or equal to the warm-up training round t, performing warm-up training on the image classification network using a cross entropy loss function.

7. The noisy image classification method based on sample selection and label distribution learning according to claim 1, characterized in that: The method further includes: if the current training round is equal to the maximum training round T, stopping the training to obtain a trained image classification network.

Images