Image recognition network training method and device resisting label noise, equipment and medium

Abstract

The application relates to an image recognition network training method and device resisting label noise, equipment and a medium. When iteratively training an image recognition network, the image recognition network of the last iteration is taken as a teacher network, and the image recognition network of the current iteration is taken as a student network. Cross-entropy loss is calculated according to the prediction result output by the student network and the corresponding label. Self-non-true distillation loss is calculated according to the prediction results output by the student network and the teacher network. A total loss function is obtained according to the cross-entropy loss and the self-non-true distillation loss. Finally, the learnable parameters in the image recognition network are adjusted by using the total loss function until convergence, and a trained image recognition network is obtained. The method can effectively resist the interference of noise labels on network training.

Description

Technical Field

[0001] This application relates to the field of image classification task technology, and in particular to a method, apparatus, device and medium for training image recognition networks that resists label noise. Background Technology

[0002] Currently, neural networks have achieved remarkable success in numerous machine learning methods, such as object detection, language translation, and image classification. The success of these methods typically relies on high-quality datasets. However, obtaining accurately labeled, high-quality datasets in complex real-world scenarios is expensive and difficult. To improve efficiency and reduce costs, researchers often choose to collect data through web scraping or crowdsourcing. Datasets obtained through these methods inevitably contain a large number of noisy labels. For example, the ImageNet dataset, containing 15 million images, was manually labeled over three years using a crowdsourcing platform. Due to limitations in expertise and oversights, errors inevitably occur during the manual labeling process, resulting in some incorrectly labeled data in the collected dataset. Furthermore, in the medical field, complex and difficult-to-interpret medical data may be encountered. Even experienced pathologists cannot provide completely reliable diagnoses, which inevitably leads to some incorrectly labeled data in the labeling of medical-related datasets. In more complex real-world scenarios, due to data quality issues, data encoding problems, and communication problems, the actual collected dataset may contain only a small number of clean labels. The presence of these noisy labels negatively impacts network training.

[0003] On datasets containing noisy labels, training a neural network directly without any processing will make it difficult for it to learn the true distribution of the data. For example, in image classification problems, researchers often use the classic cross-entropy loss function to train networks. This often produces good results when the dataset is free of noisy labels. However, when noisy labels are present, the network's accuracy on clean test sets drops significantly. This is because deep neural networks are often over-parameterized, with their number of parameters far exceeding the size of the training dataset. In scenarios involving noisy label learning, they are fully capable of overfitting to noisy labels, resulting in poor generalization performance. Therefore, avoiding the interference of noisy labels on network training, thereby improving the generalization and robustness of deep learning models, is a major challenge. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, device, and medium for training image recognition networks that resists label noise and can effectively avoid interference from noise labels during network training, in order to address the aforementioned technical problems.

[0005] A method for training an image recognition network that resists label noise, the method comprising: Obtain an image sample training set, which includes multiple image samples and their corresponding labels; The image samples in the image sample training set are input into the image recognition network for multiple iterations of training. In each iteration, the image recognition network from the previous iteration is used as the teacher network, and the image recognition network from the current iteration is used as the student network. The cross-entropy loss is calculated based on the prediction results output by the student network and the corresponding labels. The self-false distillation loss is calculated based on the prediction results output by the student network and the teacher network. The total loss function is obtained based on the cross-entropy loss and the self-false distillation loss. The learnable parameters in the image recognition network are adjusted using the total loss function until convergence, resulting in a trained image recognition network.

[0006] In one embodiment, the label corresponding to the image sample is a truth label or a noise label.

[0007] In one embodiment, calculating the self-false distillation loss based on the prediction results output by the student network and the teacher network includes: From the prediction results output by the student network and the teacher network, the results corresponding to the labels are masked, and then the prediction distribution probability is calculated separately. After performing Softmax calculations on the predicted distribution probabilities of the student network and the teacher network respectively, the KL divergence loss between the two is calculated as the self-inauthentic distillation loss.

[0008] In one embodiment, the total loss function is expressed as:

[0009] In the above formula, This refers to the teacher network. This refers to the student network. This represents the image sample. This represents the cross-entropy loss. This indicates the loss from self-non-true distillation. This represents the predicted distribution probability of the student network. This represents the predicted distribution probability of the teacher network after the labels are masked. This represents the predicted distribution probability of the student network after the labels are masked. Indicates the preset weight. It is a temperature scaling factor.

[0010] In one embodiment, the self-non-true distillation loss is:

[0011] In the above formula, This represents the label of the corresponding image sample input into the image recognition network. Indicates the prediction result. express The Values.

[0012] This application also provides an image recognition network training device that resists label noise, the device comprising: The training set acquisition module is used to acquire an image sample training set, which includes multiple image samples and their corresponding labels. The iterative training module is used to input image samples from the image sample training set into the image recognition network for multiple iterative trainings. In each iterative training, the image recognition network from the previous iterative training is used as the teacher network, and the image recognition network from the current iterative training is used as the student network. The loss function construction module is used to calculate the cross-entropy loss based on the prediction results output by the student network and the corresponding labels, calculate the self-false distillation loss based on the prediction results output by the student network and the teacher network, and obtain the total loss function based on the cross-entropy loss and the self-false distillation loss. The trained image recognition network module is used to adjust the learnable parameters in the image recognition network using the total loss function until convergence, thus obtaining the trained image recognition network.

[0013] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps: Obtain an image sample training set, which includes multiple image samples and their corresponding labels; The image samples in the image sample training set are input into the image recognition network for multiple iterations of training. In each iteration, the image recognition network from the previous iteration is used as the teacher network, and the image recognition network from the current iteration is used as the student network. The cross-entropy loss is calculated based on the prediction results output by the student network and the corresponding labels. The self-false distillation loss is calculated based on the prediction results output by the student network and the teacher network. The total loss function is obtained based on the cross-entropy loss and the self-false distillation loss. The learnable parameters in the image recognition network are adjusted using the total loss function until convergence, resulting in a trained image recognition network.

[0014] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Obtain an image sample training set, which includes multiple image samples and their corresponding labels; The image samples in the image sample training set are input into the image recognition network for multiple iterations of training. In each iteration, the image recognition network from the previous iteration is used as the teacher network, and the image recognition network from the current iteration is used as the student network. The cross-entropy loss is calculated based on the prediction results output by the student network and the corresponding labels. The self-false distillation loss is calculated based on the prediction results output by the student network and the teacher network. The total loss function is obtained based on the cross-entropy loss and the self-false distillation loss. The learnable parameters in the image recognition network are adjusted using the total loss function until convergence, resulting in a trained image recognition network.

[0015] The aforementioned image recognition network training method, apparatus, device, and medium that resists label noise utilizes a specific approach. During iterative training, the image recognition network from the previous iteration is used as the teacher network, and the image recognition network from the current iteration is used as the student network. Cross-entropy loss is calculated based on the prediction results output by the student network and the corresponding labels. Self-false distillation loss is calculated based on the prediction results output by both the student and teacher networks. A total loss function is derived from the cross-entropy loss and self-false distillation loss. Finally, the learnable parameters in the image recognition network are adjusted using the total loss function until convergence, resulting in a trained image recognition network. This method effectively resists the interference of noisy labels on network training. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an image recognition network training method that resists label noise in one embodiment. Figure 2 This is a schematic diagram illustrating the principle of self-non-true distillation in this method in one embodiment; Figure 3 This is a structural block diagram of an image recognition network training device that resists label noise in one embodiment; Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0018] Training deep neural networks using large datasets is essential, but the problem of mislabeled tags cannot be completely avoided. Label-noise learning studies how to learn better from given noisy training data. Therefore, it has gradually gained widespread attention from academia and industry and has become a very popular research field. It is important to emphasize that practical research in label-noise learning often needs to consider various different noise environments, which makes the research more complex. Specifically, the proposed method needs to be experimentally validated on various types of synthetic simulated noise data and real-world noise data at different noise levels.

[0019] In recent years, the memory effect in deep neural networks has provided a new direction for addressing label-noise learning. Although over-parameterized deep neural networks are capable of memorizing all data, the memory effect indicates that they tend to prioritize memorizing clean data before gradually memorizing mislabeled data. A classic method to combat label noise using this property is early stopping. Its core idea is to stop training when the model's performance on the validation set no longer improves, thus avoiding overfitting in the later stages of iteration. Early stopping can reduce the negative impact of noisy labels to some extent. However, because early stopping simply utilizes the data memorization tendency of the memory effect—prioritizing clean data—the network can still overfit some noisy labels even in the early training phase, affecting generalization performance. Furthermore, many current methods are limited to certain complex noise environments and cannot maintain good robustness across a wide range of noise settings.

[0020] In existing technologies, methods for addressing label noise can be broadly categorized into model-based and model-free approaches. Model-based methods focus on the generation process of noisy labels, attempting to reveal the transformation relationship from clean to noisy labels, thereby establishing a noise transition matrix. In this type of method, the accurate estimation of the noise transition matrix significantly impacts the effectiveness of label noise learning. However, accurately estimating the noise transition matrix is ​​difficult when dealing with large categories or heavy noise. Model-free methods, on the other hand, treat noisy data as anomalies and incorporate overfitting prevention techniques to enhance the robustness of label noise. However, a drawback of this type of method is that its generalization performance significantly degrades when dealing with complex data. Additionally, sample selection methods are also effective, utilizing the memory effect of deep neural networks to select clean labels for training. However, these methods can lead to the accumulation of incorrect selections during training and perform poorly when there is significant overlap between the distribution areas of clean and noisy data. Furthermore, directly discarding data deemed "noisy" results in the loss of a considerable amount of valuable information.

[0021] To address the aforementioned issues, this application, such as Figure 1 The present invention provides a method for training an image recognition network that resists label noise, comprising the following steps: Step S100: Obtain the image sample training set, which includes multiple image samples and their corresponding labels.

[0022] Step S110: Input the image samples in the image sample training set into the image recognition network for multiple iterations of training. In each iteration, the image recognition network from the previous iteration is used as the teacher network, and the image recognition network from the current iteration is used as the student network.

[0023] Step S110: Calculate the cross-entropy loss based on the prediction results output by the student network and the corresponding labels; calculate the self-false distillation loss based on the prediction results output by the student network and the teacher network; and obtain the total loss function based on the cross-entropy loss and the self-false distillation loss.

[0024] Step S120: Adjust the learnable parameters in the image recognition network using the total loss function until convergence, and obtain the trained image recognition network.

[0025] In this embodiment, self-distillation is introduced into the network training process, utilizing the teacher network from the previous stage to inform the currently trained network, thereby enhancing the memorization of clean data. Building upon this, non-true distillation is further proposed, where the label categories are masked in the logical output before extracting knowledge from the teacher network, focusing only on the remaining categories to correct accumulated erroneous knowledge.

[0026] In this embodiment, the network being trained is an image recognition network, which represents a multi-class classification problem scenario. It represents the number of categories. Let the feature space... ,in, It's a dimension, a label space. . yes The joint probability distribution. In an ideal classification task, from Draw an independent and identically distributed sample set from the data. Model training is performed, where This refers to the sample size. It's expected that the model trained on this sample set will be able to operate on the true distribution. The performance is good. However, in the real world, labels can be contaminated, so they cannot be directly obtained from the distribution. Samples were sampled from the middle. Using... To represent the joint distribution of contamination, use Indicates from A set of independent and identically distributed samples drawn from the dataset, wherein... This indicates a label with noise.

[0027] The goal of label noise learning is to achieve this even when using a noisy sample set. It can also learn a method that can be applied to the real distribution. A robust classifier that accurately classifies data. Let... For having learnable parameters The classifier, Represents each sample The logical output is the prediction result. In fact, this method can be widely applied to supervised training of networks in various scenarios to resist the interference of noisy labels on network training. In this paper, we will take the image object classification scenario as an example to illustrate this method.

[0028] In this embodiment, the image recognition network can be: LeNet-5, a classic convolutional neural network composed of convolutional layers, pooling layers, and fully connected layers; VGGNet, characterized by its very small convolutional kernel size, stacking multiple convolutional and pooling layers to extract image features; GoogLeNet (Inception), employing the Inception module structure, capturing features at different scales by parallelizing multiple convolutional kernels and pooling layers of different sizes; ResNet, a deep residual network that solves the gradient vanishing problem in deep network training by introducing residual connections, resulting in a very deep network structure; DenseNet, a densely connected network that connects the feature maps of previous layers with the feature maps of the current layer in each layer to form a densely connected network structure; and MobileNet, a lightweight convolutional neural network that uses techniques such as depthwise separable convolution to reduce the number of parameters and computational cost.

[0029] In step S100, the labels corresponding to the image samples used for training are either ground truth labels, which are labels that correctly classify the target, or noise labels, which are labels that misclassify the target.

[0030] Since network training is essentially a forgetting process—meaning the network gradually forgets some knowledge learned from clean data during overfitting to mislabeled data—this embodiment introduces a self-distillation method into training to address this issue. Self-distillation utilizes the network from previous stages as a teacher network, guiding the current network to review its previously learned knowledge, thereby strengthening the learning of clean data and reducing knowledge forgetting. Specifically, during the target recognition network training process, the previously trained target recognition network serves as the teacher network, and the currently trained target recognition network serves as the student network.

[0031] Specifically, in the self-distillation process, the current network... Designated as the student network, while the network from the previous training phase... As a teacher network, among which This indicates the training phase. (Through...) Output and Alignment reinforces and strengthens the knowledge learned in previous stages. Through self-distillation, the network can effectively and continuously review previous knowledge. However, the negative impact of mislabeled data accumulates during training. Simply reviewing knowledge from the previous stage is insufficient. The network also needs to promptly correct the erroneous model learning direction caused by mislabeled data.

[0032] Therefore, further, a self-non-true distillation is proposed based on self-distillation. That is, in this embodiment, alignment... and When outputting the logic output, the label categories are masked, retaining only the non-labeled (i.e., non-true) categories. Although the network cannot determine whether the label is mislabeled, masking the label categories in the logic output is beneficial. If the image label is incorrect, this operation can directly mask the interference of the erroneous label. Conversely, even if the image label is correct, the erroneous label will be retained. However, because the negative impact of mislabeled data gradually accumulates during training, the current network's prediction probability will have a large bias. By aligning the non-true categories of the previous network, the prediction bias of the current network can be corrected in a timely manner.

[0033] Specifically, in the and When aligning the outputs, cross-entropy loss and self-false distillation loss are used. The cross-entropy loss is calculated using the prediction results from the current network (student network) and their corresponding class labels. The self-false distillation loss is calculated by first masking the corresponding label results from the prediction results of both the student and teacher networks, then performing Softmax calculations on the prediction probability distributions of both networks, and finally calculating the KL divergence loss between them as the self-false distillation loss. The principle of self-false distillation is as follows: Figure 2 As shown.

[0034] In this embodiment, the prediction result is a sequence of predicted probabilities for targets of various categories in the image sample. After deleting the category corresponding to the label, the softening probability distribution is calculated.

[0035] In this embodiment, the total loss function is expressed as: (1) In formula (1), Indicates a teacher network, Indicates student network, Represents image samples, Represents cross-entropy loss, This indicates self-inauthentic distillation loss. This represents the predicted distribution probability of the student network. This represents the predicted distribution probability of the teacher network after the labels are masked. This represents the predicted distribution probability of the student network after the labels are masked. Indicates the preset weight. It is a temperature scaling factor.

[0036] Furthermore, the loss from self-improper distillation is expressed as: (2) In formula (2), This represents the label of the corresponding image sample input into the image recognition network. Indicates the prediction result. express The Values.

[0037] Preferably, In Set to 1 and adjust. In For simplicity, this invention is set up... and adjust The value of .

[0038] In the field of labeled noise learning, the main datasets currently used include real-world noise datasets and simulated noise datasets. For real-world noise datasets such as Clothing1M and CIFAR10N, they can be used directly to verify the effectiveness of the method. However, for simulated noise datasets, noise needs to be manually added to existing clean datasets such as CIFAR-10 and CIFAR-100.

[0039] To verify the effectiveness of this method, we plan to manually destroy the clean labels of four commonly used datasets: MNIST, F-MNIST, CIFAR-10, and CIFAR-100. Specifically, we will consider four types of simulated noise to generate noisy labels: (1) Symmetric noise (sym.): Each class has the same probability of being incorrectly flipped to any other class. (2) Asymmetric noise (asym.): Similar classes are incorrectly flipped to each other. (3) Pairflip noise (pair.): Each class can only be incorrectly flipped to its adjacent class. (4) Instance-dependent noise (ins.): This type of noise is affected by image features, which is more consistent with the noise generation process in the real world and is also more challenging.

[0040] Tables 1 and 2 present the comparative results of our proposed method with other methods on real-world noise datasets and simulated noise datasets, respectively. Our proposed method demonstrates good classification accuracy, achieving near-optimal results under various noise settings. This provides strong evidence for the effectiveness of our proposed method in processing both real-world and simulated noise datasets.

[0041] Table 1: Comparison results of our method with other methods on real-world noise datasets Clothing1M (I), CIFAR10N-1 (II), CIFAR10N-2 (III), CIFAR10N-3 (IV), and CIFAR10N-W (V). Table 2: Comparison results of this method with other methods on simulated noise datasets F-MNIST, CIFAR-10, and CIFAR-100. Among the above-mentioned image recognition network training methods that resist label noise, an innovative self-false distillation method is provided to address the label noise learning problem. This method has achieved remarkable technical results in practice, as specifically demonstrated below: (1) Preventing knowledge forgetting and error accumulation: In the field of label noise learning, network training is a process of knowledge forgetting and error accumulation. During the training process, it is necessary to continuously review the previously learned knowledge and correct the accumulated errors in a timely manner.

[0042] (2) Introduction of Self-Distillation: This method introduces a self-distillation approach, which is embedded into the training process to guide the current student network to review and consolidate knowledge from clean data. This method effectively helps the neural network better understand and utilize known correct information, thereby improving the model's learning efficiency and performance.

[0043] (3) Propose a non-true distillation method: This method also includes a non-true distillation method, which corrects the erroneous orientation of mislabeled data by masking the label categories in the model's logical output and aligning only the non-true categories. This technique helps the model better adapt to label noise and improves its robustness to data with noisy labels.

[0044] (4) Validation on multiple noisy datasets: This method has been thoroughly validated experimentally on multiple noisy datasets, demonstrating the superior performance of the self-non-true distillation method. This method has proven its excellent applicability and effectiveness in practical applications, and can significantly improve the performance of deep learning models on challenging data with noisy labels.

[0045] In summary, this method addresses the issues of knowledge forgetting and error accumulation, introduces self-distillation and non-true distillation to improve model performance, and its effectiveness and superiority are ensured through validation on multiple experimental datasets. This innovative solution is of great significance to the field of label-noise learning and provides strong support for deep learning models to perform exceptionally well in handling complex real-world data.

[0046] It should be understood that, although Figure 1The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0047] In one embodiment, such as Figure 3 The present invention provides an image recognition network training device that resists label noise, comprising: a training set acquisition module 200, an iterative training module 210, a loss function construction module 220, and a trained image recognition network acquisition module 230, wherein: The training set acquisition module 200 is used to acquire an image sample training set, which includes multiple image samples and their corresponding labels.

[0048] The iterative training module 210 is used to input image samples from the image sample training set into the image recognition network for multiple iterative trainings. In each iterative training, the image recognition network from the previous iterative training is used as the teacher network, and the image recognition network from the current iterative training is used as the student network.

[0049] The loss function construction module 220 is used to calculate the cross-entropy loss based on the prediction results output by the student network and the corresponding labels, calculate the self-false distillation loss based on the prediction results output by the student network and the teacher network, and obtain the total loss function based on the cross-entropy loss and the self-false distillation loss.

[0050] The trained image recognition network module 230 is used to adjust the learnable parameters in the image recognition network using the total loss function until convergence, thus obtaining the trained image recognition network.

[0051] Specific limitations regarding the training device for image recognition networks resistant to label noise can be found in the limitations of the training method for image recognition networks resistant to label noise mentioned above, and will not be repeated here. Each module in the aforementioned training device for image recognition networks resistant to label noise can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0052] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 4 The computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for training an image recognition network that resists tag noise. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0053] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0054] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps: Obtain an image sample training set, which includes multiple image samples and their corresponding labels; The image samples in the image sample training set are input into the image recognition network for multiple iterations of training. In each iteration, the image recognition network from the previous iteration is used as the teacher network, and the image recognition network from the current iteration is used as the student network. The cross-entropy loss is calculated based on the prediction results output by the student network and the corresponding labels. The self-false distillation loss is calculated based on the prediction results output by the student network and the teacher network. The total loss function is obtained based on the cross-entropy loss and the self-false distillation loss. The learnable parameters in the image recognition network are adjusted using the total loss function until convergence, resulting in a trained image recognition network.

[0055] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Obtain an image sample training set, which includes multiple image samples and their corresponding labels; The image samples in the image sample training set are input into the image recognition network for multiple iterations of training. In each iteration, the image recognition network from the previous iteration is used as the teacher network, and the image recognition network from the current iteration is used as the student network. The cross-entropy loss is calculated based on the prediction results output by the student network and the corresponding labels. The self-false distillation loss is calculated based on the prediction results output by the student network and the teacher network. The total loss function is obtained based on the cross-entropy loss and the self-false distillation loss. The learnable parameters in the image recognition network are adjusted using the total loss function until convergence, resulting in a trained image recognition network.

[0056] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0057] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0058] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for training an image recognition network that resists label noise, characterized in that, The method includes: Obtain an image sample training set, which includes multiple image samples and their corresponding labels; The image samples in the image sample training set are input into the image recognition network for multiple iterations of training. In each iteration, the image recognition network from the previous iteration is used as the teacher network, and the image recognition network from the current iteration is used as the student network. The cross-entropy loss is calculated based on the prediction results output by the student network and the corresponding labels. The self-false distillation loss is calculated based on the prediction results output by the student network and the teacher network. The total loss function is obtained based on the cross-entropy loss and the self-false distillation loss. The self-false distillation loss is: In the above formula, This represents the label of the corresponding image sample input into the image recognition network. Indicates the prediction result. express The One value; The learnable parameters in the image recognition network are adjusted using the total loss function until convergence, resulting in a trained image recognition network.

2. The image recognition network training method according to claim 1, characterized in that, The labels corresponding to the image samples are either truth labels or noise labels.

3. The image recognition network training method according to claim 2, characterized in that, The self-false distillation loss calculated based on the prediction results output by the student network and the teacher network includes: From the prediction results output by the student network and the teacher network, the results corresponding to the labels are masked, and then the prediction distribution probability is calculated separately. After performing Softmax calculations on the predicted distribution probabilities of the student network and the teacher network respectively, the KL divergence loss between the two is calculated as the self-inauthentic distillation loss.

4. The image recognition network training method according to claim 3, characterized in that, The total loss function is expressed as: In the above formula, This refers to the teacher network. This refers to the student network. This represents the image sample. This represents the cross-entropy loss. This indicates the loss from self-non-true distillation. This represents the predicted distribution probability of the student network. This represents the predicted distribution probability of the teacher network after the labels are masked. This represents the predicted distribution probability of the student network after the labels are masked. Indicates the preset weight. It is a temperature scaling factor.

5. A training device for an image recognition network that resists label noise, characterized in that, The device includes: The training set acquisition module is used to acquire an image sample training set, which includes multiple image samples and their corresponding labels. The iterative training module is used to input image samples from the image sample training set into the image recognition network for multiple iterative trainings. In each iterative training, the image recognition network from the previous iterative training is used as the teacher network, and the image recognition network from the current iterative training is used as the student network. The loss function construction module is used to calculate the cross-entropy loss based on the prediction results output by the student network and the corresponding labels, calculate the self-false distillation loss based on the prediction results output by the student network and the teacher network, and obtain the total loss function based on the cross-entropy loss and the self-false distillation loss; the self-false distillation loss is: In the above formula, This represents the label of the corresponding image sample input into the image recognition network. Indicates the prediction result. express The One value; The trained image recognition network module is used to adjust the learnable parameters in the image recognition network using the total loss function until convergence, thus obtaining the trained image recognition network.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

Images