Medical image recognition system based on label noise robust learning

Abstract

The invention discloses a medical image recognition system based on label noise robust learning, and provides a hard sample label refinement strategy based on a confidence perception weighted prototype and an effective noise sample joint correction method to process divided subsets in correction before training so as to obtain training data with higher quality; in progressive hard sample reinforcement learning, data are input according to the learning difficulty of samples for training, the ability of model learning discrimination feature representation is improved by integrating cross entropy loss, consistency loss and credible contrast loss, and the influence of medical tag noise is reduced. According to the method, by integrating label refinement and a progressive hard sample reinforcement learning technology, the accuracy and robustness of medical image recognition under the condition of noise label data are effectively improved, overfitting of the model to noise samples is relieved, and the method can be used for correctly building a disease auxiliary diagnosis model under the condition that the noise label samples exist in training data.

Description

Technical Field

[0002] This invention relates to the field of medical label noise learning technology, specifically a medical image recognition system based on label noise robust learning. Background Technology

[0004] In recent years, with the rapid development of medical imaging technology and the increasing demand for medical diagnosis, deep neural network-based medical image recognition has played an increasingly important role in disease-aided diagnosis and treatment. However, its success depends on data with correct labels. In medical images, the annotation process is highly dependent on the subjective judgment of professional physicians, and coupled with the complexity and diversity of medical images themselves, the uncertainty of annotation results is high, leading to noisy labels in the data. At the same time, other labor-saving annotation methods, such as mining clinical reports, are more likely to introduce samples with noisy labels into the dataset. In these cases, it is usually difficult to guarantee that all collected data labels are correct. The presence of label noise can seriously affect the performance of the model, causing the model to learn incorrect features during training, thereby affecting the accuracy and reliability of diagnostic results. Therefore, how to build a robust medical image recognition model in the presence of label noise is a significant challenge. Summary of the Invention

[0006] To effectively address the challenges of distinguishing between hard and noisy samples, limited data quality improvement capabilities, and low sample utilization in noise-labeled learning, this invention proposes a medical image recognition system based on label-noise robust learning, which improves the accuracy and robustness of medical image recognition in noisy label data environments.

[0007] To achieve the above objectives, the following technical solution is adopted:

[0008] A medical image recognition system based on label noise robust learning includes a pre-training correction module and a progressive hard sample augmentation learning module.

[0009] The pre-training correction module includes:

[0010] Data collection and processing unit: Collects several medical images and assigns labels to each image, resulting in a medical image dataset with noise labels. The medical images in the dataset are preprocessed and then processed using a deep neural network. Obtain the depth features of the weakly enhanced and strongly enhanced views of each medical image;

[0011] Confidence Probability Set Unit: Based on the aforementioned deep features, the classification confidence of the sample is obtained, and the classification predictions of several training rounds are stored in the memory library.

[0012] Two-branch sample splitting unit: Utilizing classification confidence and a memory library to split datasets with noisy labels. Divided into clean sample subsets hard sample subset and noise sample subset ;

[0013] Hard sample label refinement unit: normalized clean sample subset The deep features are used to calculate the confidence-aware class prototype using normalized features; based on The similarity score between hard samples and class prototypes is used to characterize class prototype-based labels by integrating the similarity information of samples in the neighborhood cluster; the labels of hard samples are refined by the pseudo labels and class prototype labels output by the model.

[0014] Noise sample correction unit: utilizing a clean sample subset The labels of noisy samples are corrected based on the similarity of deep features and classification confidence, and the hard sample subset after refining the labels.

[0015] The progressive hard sample augmentation learning module includes:

[0016] Data sampling unit: Loading a deep neural network with shared weights And deep neural networks that update parameters using the exponential moving average formula In the model The difficulty of learning from the updated labeled dataset is calculated, and data is dynamically selected; the augmentation transformation of the data is then input into the deep neural network.

[0017] Consistency constraint unit: compares the prediction results of weakly augmented views and strongly augmented views based on sampled data, and obtains the view through information divergence. Figure 1 Sexual damage ;

[0018] Trustworthy contrastive learning unit: Determines whether a sample is trustworthy, and performs contrastive learning on trustworthy samples to obtain the trustworthy contrastive loss. ;

[0019] Classification unit: Calculates the ensemble prediction cross-entropy loss based on the model's historical output as the supervision loss. Comprehensive monitoring of losses ,See Figure 1 Sexual damage Comparison loss with credibility Calculate the total loss function. Update the network again. The weights are used to complete the classification task of medical images.

[0020] The aforementioned medical image recognition system based on label noise robust learning, wherein the data collection and processing unit: assuming For the first An input medical image, This represents the corresponding noise label, and the collected dataset is... The model network consists of three components: (1) instances Mapping to high-dimensional representation Deep feature encoder (2) A classifier head (A fully connected layer followed by a softmax function), which receives As input, output class prediction The predicted label is (3) Mapping to a low-dimensional representation Linear or nonlinear projection functions Weak Enhanced View and enhanced view The depth features are respectively and .

[0021] The aforementioned medical image recognition system based on label noise robust learning, wherein the confidence probability set unit uses a dataset with noise labels. Training deep neural network models Classification confidence is obtained based on deep features. and Predicting categories This indicates that both sizes are set to 1. memory library and , Store recent Weak view prediction in each training round. Store recent Strong view prediction for each training round.

[0022] The medical image recognition system based on label noise robust learning, wherein the dual-branch sample partitioning unit integrates the partitioning results of two branches.

[0023] The first branch measures the high probability consistency of predictions across different views using classification confidence. A sample is considered clean if its two views yield consistent and high-confidence results from the network.

[0024]

[0025] Hard subsets primarily consist of hard samples from similar classes, for which it is often difficult to determine whether the labels are correct. Therefore, inconsistent predictions from the two views are used for selection.

[0026]

[0027] The noisy subset mainly consists of samples with noisy labels. Because the features and labels of these samples lack correlation and consistency, network models struggle to effectively fit this data during training. Therefore, network models often make predictions inconsistent with the given labels.

[0028]

[0029] in A dynamic instance threshold is used to estimate the learning state in real time by using the model's predicted probability, and then determine an appropriate threshold based on this state.

[0030]

[0031] The second branch considers the consistency of network model changes through the memory library, defining clean samples and their noisy samples as follows:

[0032]

[0033]

[0034] in and The score represents the inconsistency between the two views. Defined as corresponding After sorting in descending order, the first The value at that location, The initial value is the initial noise rate, and then it changes to 0.8 times the original value every 30 rounds. The larger the value, the more frequent the label changes during training, the greater the inconsistency in these changes, and the greater the uncertainty in the model's predictions. Representative at The probability that the middle label is c. .

[0035]

[0036]

[0037]

[0038] Combining the results from both branches, a finer partition is obtained using the following formula to obtain a clean sample subset. hard sample subset and noise sample subset .

[0039]

[0040] The aforementioned medical image recognition system based on label noise robust learning, wherein the hard sample label refinement unit: through... Normalized clean sample subset depth features The class prototype is computed in a confidence-aware manner using all training samples in a clean subset.

[0041]

[0042]

[0043] in represent The probability score of belonging to category c, and the confidence weight. Representative sample Standardized score on category c. Then for a hard sample... Based on its low-dimensional vector and class prototype similarity score Calculate hard samples In the cluster formed by the neighboring samples + The probability that each member belongs to category c, specifically, let Then the class prototype pseudo-tag :

[0044]

[0045] Based on the pseudo-labels of hard samples and prototype neighborhood pseudo-labels The hard sample labels are refined as follows:

[0046]

[0047]

[0048] The aforementioned label-based noise-robust learning medical image recognition system includes a noise sample correction unit that utilizes joint data from clean and hard subsets. The noise labels are updated using a combination of model predictions and feature space information from the noise samples:

[0049]

[0050] in For joint data A subset of category c For noise collection middle The low-dimensional feature vector of the weakly augmented view. For noise collection middle The low-dimensional feature vectors of the weakly augmented view. Higher cosine similarity, or higher similarity between samples and joint data... The larger the bhattacharyya coefficient between them, the more likely they are to have the same label. The noise labels in period t are iteratively corrected using the moving average formula, and the final corrected labels are obtained after the argmax operation.

[0051]

[0052]

[0053]

[0054] The aforementioned medical image recognition system based on label noise robust learning, wherein the data sampling unit is: and Shared parameters , Depend on Updated It is the momentum coefficient. Perform backpropagation; input the processed sample into... In the process, their predicted probabilities are sorted; a dynamic scaling parameter is set that can gradually increase with the increase of training rounds, and an appropriate number (denoted as ) is selected according to this scaling parameter. (sample).

[0055] The aforementioned medical image recognition system based on label noise robust learning includes a consistency constraint unit that utilizes Kullback-Leibler divergence to implement a dual-view constraint on model consistency, thereby obtaining a compact feature representation.

[0056]

[0057] The aforementioned medical image recognition system based on label noise robust learning, wherein the trusted contrastive learning unit generates sample pairs by selecting the majority of trusted samples based on the model consistency of the two-view samples.

[0058]

[0059] The contrastive loss for each sample pair is then defined as:

[0060]

[0061] set up and If they are sets of weakly enhanced views and strongly enhanced views respectively, then... Defined as:

[0062]

[0063] The aforementioned medical image recognition system based on label noise robust learning, wherein the classification unit uses an ensemble prediction cross-entropy loss based on the model's historical output to reduce the influence of noisy labels. Let the... The model predicts for each training epoch. Generate set predictions for each sample The previous historical forecasts are aggregated using an exponential moving average:

[0064]

[0065]

[0066] in This represents the momentum coefficient. Then, both the original noise labels and the integrated predictions are used as targets.

[0067]

[0068] The total training loss is:

[0069]

[0070] in and These are the weights for consistency loss and reliable comparison loss, respectively.

[0071] The beneficial effects of this invention are:

[0072] This invention constructs a medical image recognition system based on label noise robust learning. The invention partitions the dataset into two representative branches based on perceptual consistency of the data, mitigating sample selection bias, and utilizes label refinement and correction strategies to improve the quality of training data, ensuring the reliability of samples input into the model. This invention employs a progressive learning training method and designs an effective loss function to train the medical image classification model, improving the efficiency and reliability of disease-aided diagnosis. The results of this invention are applicable to scenarios where training data labels are incorrect due to limited medical resources. By using deep learning technology, it effectively addresses the label noise problem in medical images, reducing reliance on professional doctors, improving the quality of training data, and providing effective support for computer-aided diagnosis. Attached Figure Description

[0074] Figure 1 This is a schematic diagram of the overall process of the medical image recognition system based on label noise robust learning according to the present invention;

[0075] Figure 2 This is a flowchart illustrating the specific process of the hard sample label refinement unit of the present invention.

[0076] Figure 3 This is a schematic diagram of the noise sample correction unit of the present invention;

[0077] Figure 4 The graph shows a comparison of the test accuracy of the medical image recognition system based on label noise robust learning of this invention with two other methods on skin datasets with different noise rates.

[0078] Figure 5 The figure shows the iterative comparison curves of the test accuracy of the medical image recognition system based on label noise robust learning of the present invention with two other methods at a noise rate of 0.1.

[0079] Figure 6 The figure shows the iterative comparison curves of the test accuracy of the medical image recognition system based on label noise robust learning of the present invention with two other methods at a noise rate of 0.3. Detailed Implementation

[0081] The present invention will be described in detail below with reference to specific embodiments.

[0082] A medical image recognition system based on label noise robust learning is characterized by including a pre-training correction module and a progressive hard sample augmentation learning module, the overall process of which is shown in the figure below. Figure 1 As shown; the pre-training correction module includes:

[0083] A1 Data Collection and Processing Unit: Collects several medical images and assigns labels to each image, resulting in a medical image dataset with noise labels. The medical images in the dataset are preprocessed and then processed using a deep neural network. Obtain the depth features of the weakly enhanced and strongly enhanced views of each medical image;

[0084] Where, assuming For the first An input medical image, This represents the corresponding noise label, and the collected dataset is... The model network consists of three components: (1) instances Mapping to high-dimensional representation Deep feature encoder (2) A classifier head (A fully connected layer followed by a softmax function), which receives As input, output class prediction The predicted label is (3) Mapping to a low-dimensional representation Linear or nonlinear projection functions Weak Enhanced View and enhanced view The depth features are respectively and 。;

[0085] A2 Confidence Probability Set Unit: Based on the aforementioned deep features, the classification confidence of the sample is obtained. and Predicting categories This indicates that the classification predictions from several training rounds are simultaneously stored in a memory database, with both databases set to a size of [size missing]. memory library and , Store recent Weak view prediction in each training round. Store recent Strong view prediction for each training round.

[0086] A3 Two-Branch Sample Splitting Unit: The classification confidence score and in-memory database are input into the two-branch sample splitting unit to split the dataset with noisy labels. Divided into clean sample subsets hard sample subset and noise sample subset The specific steps of step A3 are as follows:

[0087] A31. The first branch measures the high probability consistency of predictions across different views using classification confidence, and is divided into... , and Three subsets. A sample is considered clean if its two views can yield consistent and high-confidence results from the network.

[0088]

[0089] Hard subsets primarily consist of hard samples from similar classes, for which it is often difficult to determine whether the labels are correct. Therefore, inconsistent predictions from the two views are used for selection.

[0090]

[0091] The noisy subset mainly consists of samples with noisy labels. Because the features and labels of these samples lack correlation and consistency, network models struggle to effectively fit this data during training. Therefore, network models often make predictions inconsistent with the given labels.

[0092]

[0093] in A dynamic instance threshold is used to estimate the learning state in real time by using the model's predicted probability, and then determine an appropriate threshold based on this state.

[0094]

[0095] A32. The second branch considers the consistency of network model changes through the memory library, defining clean samples and noisy samples as follows:

[0096]

[0097]

[0098] in and The score represents the inconsistency between the two views. Defined as corresponding After sorting in descending order, the first The value at that location, The initial value is the initial noise rate, and then it changes to 0.8 times the original value every 30 rounds. The larger the value, the more frequent the label changes during training, the greater the inconsistency in these changes, and the greater the uncertainty in the model's predictions. Representative at The probability that the middle label is c. .

[0099]

[0100]

[0101]

[0102] A33. Combine the results of the two branches

[0103] The following formula is used to perform fine partitioning to obtain a clean sample subset. hard sample subset and noise sample subset .

[0104]

[0105] A4 Hard Sample Label Refinement Unit: Normalized Clean Sample Subset The deep features are used to calculate the confidence-aware class prototype using normalized features; based on The similarity score between hard samples and class prototypes is used to characterize class prototype-based labels by integrating similarity information from samples in the neighborhood cluster; the labels of hard samples are refined by using pseudo-labels and class prototype labels output by the model, such as... Figure 2 As shown;

[0106] The specific steps of step A4 are as follows:

[0107] A41. Through Normalized clean sample subset depth features The class prototype is computed in a confidence-aware manner using all training samples in a clean subset.

[0108]

[0109]

[0110] in represent The probability score of belonging to category c, and the confidence weight. Representative sample Standardized score in category c

[0111] A42. For a hard sample Based on its low-dimensional vector and class prototype similarity score Calculate hard samples In the cluster formed by the neighboring samples + The probability that each member belongs to category c, specifically, let Then the class prototype pseudo-tag :

[0112]

[0113] A43. Based on the pseudo-labels of hard samples and prototype neighborhood pseudo-labels The hard sample labels are refined as follows:

[0114]

[0115]

[0116] A5 Noise Sample Correction Unit: Utilizing a Clean Sample Subset The labels of noisy samples are corrected based on the similarity of deep features and classification confidence, and the hard sample subset after refining the labels.

[0117] The specific steps of step A5 are as follows:

[0118] A51. Utilizing the joint data of clean and hard subsets The noise labels are updated using a combination of model predictions and feature space information from the noise samples:

[0119]

[0120] in For joint data A subset of category c For noise collection middle The low-dimensional feature vector of the weakly augmented view. For noise collection middle The low-dimensional feature vectors of the weakly augmented view. Higher cosine similarity, or higher similarity between samples and joint data... The larger the bhattacharyya coefficient between them, the more likely they are to have the same label.

[0121] A52. Use the moving average formula to iteratively correct the noise label in period t, and obtain the final corrected label after the argmax operation.

[0122]

[0123]

[0124]

[0125] The progressive hard sample augmentation learning module includes:

[0126] A6 Data Sampling Unit: Loads a weight-shared deep neural network And deep neural networks that update parameters using the exponential moving average formula Using the data sampling unit, the updated labeled dataset is input into the model. Calculate the difficulty of learning from samples and dynamically select data; input the augmentation transformation of the data into the deep neural network;

[0127] in, and Shared parameters , Depend on Updated It is the momentum coefficient. Perform backpropagation; input the processed sample into... In the process, their predicted probabilities are sorted; a dynamic scaling parameter is set that can gradually increase with the increase of training rounds, and an appropriate number (denoted as ) is selected according to this scaling parameter. (sample).

[0128] A7 Consistency Constraint Unit: In the consistency constraint unit, the prediction results of the weakly augmented view and the strongly augmented view based on the sampled data are compared, and the view is obtained through information divergence. Figure 1 Sexual damage ;

[0129]

[0130] A8 Trusted Contrastive Learning Unit: In the trusted contrastive learning unit, it determines whether a sample is trustworthy and performs comparative learning on trustworthy samples to obtain the trusted contrastive loss. ;

[0131] The specific steps of step A8 are as follows:

[0132] A81. Generate sample pairs by selecting the majority of reliable samples based on the model consistency of the two-view sample:

[0133]

[0134] The contrastive loss for each sample pair is then defined as:

[0135]

[0136] A82. Let... and If they are sets of weakly enhanced views and strongly enhanced views respectively, then... Defined as:

[0137]

[0138] In the classification unit, A9 calculates the ensemble prediction cross-entropy loss based on the model's historical output as the supervision loss. Comprehensive monitoring of losses ,See Figure 1 Sexual damage Comparison loss with credibility Calculate the total loss function. Update the network again. The weights are used to complete the classification task of medical images.

[0139] The specific steps of step A9 are as follows:

[0140] A91. Use ensemble prediction cross-entropy loss based on model history output to reduce the impact of noisy labels.

[0141] Let the first The model predicts for each training epoch. Generate set predictions for each sample The previous historical forecasts are aggregated using an exponential moving average:

[0142]

[0143]

[0144] in This represents the momentum coefficient. Then, both the original noise labels and the integrated predictions are used as targets.

[0145]

[0146] A92. Update the weights using the total training loss:

[0147]

[0148] in and These are the weights for consistency loss and reliable comparison loss, respectively.

[0149] To better illustrate the effects of the present invention, a specific implementation result of the present invention is described below:

[0150] Figure 4 The figure compares the test accuracy of the medical image recognition system based on label noise robust learning of this invention with two other methods on skin datasets with different noise rates. Method 1 applies negative learning to classify noisy data, while Method 2 combines an inconsistent update strategy for dual-network peer training. As can be seen from the figure, the accuracy of this invention is the highest compared to the other two methods.

[0151] Figure 5 This is an iterative comparison curve of the test accuracy of the medical image recognition system based on label noise robust learning of this invention with two other methods at a noise rate of 0.1. Methods 1 and 2 are described in the same way. Figure 4 .

[0152] Figure 6 This is an iterative comparison curve of the test accuracy of the medical image recognition system based on label noise robust learning of this invention with two other methods at a noise rate of 0.3. Methods 1 and 2 are described in the same way. Figure 4 .

[0153] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A medical image recognition system based on label noise robust learning, characterized in that, It includes a pre-training correction module and a progressive hard-sample augmentation learning module. The pre-training correction module includes a data collection and processing unit: collecting several medical images and acquiring the label for each medical image to obtain a medical image dataset with noise labels. The medical images in the dataset are preprocessed and then processed using a deep neural network. Obtain the depth features of the weakly enhanced and strongly enhanced views of each medical image; Confidence probability set unit: Based on the depth features, obtain the classification confidence of the sample. and Simultaneously, the classification predictions from several training rounds are performed. Stored in memory; Two-branch sample partitioning unit: The dataset with noisy labels is partitioned using classification confidence and the memory library. Divided into clean sample subsets hard sample subset and noise sample subset Hard sample label refinement unit: normalized clean sample subset The deep features are used to calculate the confidence-aware class prototype using normalized features. ;according to The similarity score between the hard sample and the class prototype is used to characterize the class prototype-based label by integrating the similarity information of samples in the neighborhood cluster. The labels of hard samples are refined using pseudo-labels and class prototype labels output by the model; Noise sample correction unit: utilizes a subset of clean samples. The hard sample subset, after refining the labels, is used to correct the labels of noisy samples based on the similarity of deep features and classification confidence. The progressive hard sample augmentation learning module includes: a data sampling unit that loads a weight-sharing deep neural network. And deep neural networks that update parameters using the exponential moving average formula In the model The system calculates the learning difficulty of samples from the updated labeled dataset and dynamically selects data. The augmentation transformation of the data is then input into a deep neural network. A consistency constraint unit compares the prediction results of weakly and strongly augmented views of the sampled data and obtains the view consistency loss through information divergence. ; Credible contrastive learning unit: determines whether a sample is credible, and performs contrastive learning on credible samples to obtain the credible contrastive loss. Classification unit: Calculates the ensemble prediction cross-entropy loss based on the model's historical output as the supervised loss. Comprehensive monitoring of losses View consistency loss Comparison loss with credibility Calculate the total loss function. Update the network again. The weights are used to complete the classification task of medical images.

2. The system according to claim 1, characterized in that, The data collection and processing unit mentioned above: assuming For the first An input medical image, This represents the corresponding noise label, and the collected dataset is... ; The model network consists of three components: (1) instances Mapping to high-dimensional representation Deep feature encoder (2) A classifier head (A fully connected layer followed by a softmax function), which receives As input, output class prediction The predicted label is (3) Mapping to a low-dimensional representation Linear or nonlinear projection functions Weak Enhanced View and enhanced view The depth features are respectively and .

3. The system according to claim 1, characterized in that, The confidence probability set unit: uses a dataset with noise labels. Training deep neural network models Classification confidence is obtained based on deep features. and Predicting categories This indicates that both sizes are set to 1. memory library and , Store recent Weak view prediction in each training round. Store recent Strong view prediction for each training round.

4. The system according to claim 1, characterized in that, The dual-branch sample partitioning unit integrates the partitioning results of two branches; the first branch measures the high probability consistency of predictions across different views using classification confidence. A sample is considered clean if its two views yield consistent and high-confidence results from the network. Hard subsets mainly consist of hard samples from similar classes, for which it is often difficult to determine whether the labels are correct; therefore, inconsistent predictions from the two views are used for selection. The noisy subset mainly consists of samples with noisy labels; due to the lack of correlation and consistency between the features and labels of these samples, the network model struggles to effectively fit this data during training. Therefore, the network model often makes predictions inconsistent with the given labels. ;in A dynamic instance threshold is used to estimate the learning state in real time by using the model's predicted probability, and to determine an appropriate threshold based on this state. The second branch considers the consistency of network model changes through the memory library, defining clean samples and their noisy samples as follows: ; ;in and The score represents the inconsistency between the two views. Defined as corresponding After sorting in descending order, the first The value at that location, The initial value is the initial noise rate, and then it becomes 0.8 times the original value every 30 rounds. The larger the value, the more frequent the label changes during training, the greater the inconsistency in these changes, and the greater the uncertainty in the model's predictions. Representative at The probability that the middle label is c. ; ; ; ; Combining the results from both branches, a finer partition is obtained using the following formula to obtain a clean sample subset. hard sample subset and noise sample subset .

5. The system according to claim 1, characterized in that, The hard sample label refinement unit: through Normalized clean sample subset depth features The class prototype is computed in a confidence-aware manner using all training samples in a clean subset. ; ;in represent The probability score of belonging to category c, and the confidence weight. Representative sample Standardized score on category c; then for a hard sample Based on its low-dimensional vector and class prototype similarity score Calculate hard samples In the cluster formed by the neighboring samples + The probability that each member belongs to category c, specifically, let Then the class prototype pseudo-tag : ; Based on the pseudo-labels of hard samples and prototype neighborhood pseudo-labels The hard sample labels are refined as follows: ; .

6. The system according to claim 1, characterized in that, The noise sample correction unit utilizes joint data from clean and hard subsets. The noise labels are updated using a combination of model predictions and feature space information from the noise samples: ;in For joint data A subset of category c For noise collection middle The low-dimensional feature vector of the weakly augmented view. For noise collection middle The low-dimensional feature vectors of the weakly augmented view. Higher cosine similarity, or higher similarity between samples and joint data... The larger the bhattacharyyacoefficient between them, the more likely they are to have the same label. The noise labels in period t are iteratively corrected using a moving average formula, and the final corrected labels are obtained after an argmax operation. ; ; .

7. The system according to claim 1, characterized in that, The data sampling unit mentioned above: and Shared parameters , Depend on Updated It is the momentum coefficient. Perform backpropagation; input the processed sample into... In the process, their predicted probabilities are sorted; a dynamic scaling parameter is set that can gradually increase with the increase of training rounds, and an appropriate number (denoted as ) is selected according to this scaling parameter. (sample).

8. The system according to claim 1, characterized in that, The consistency constraint unit utilizes Kullback-Leibler divergence to implement a two-view constraint on model consistency, thereby obtaining a compact feature representation. .

9. The system according to claim 1, characterized in that, The aforementioned trusted contrastive learning unit generates sample pairs by selecting the majority of trusted samples based on the model consistency of the two-view samples. The contrastive loss for each sample pair is then defined as: ;set up and If they are sets of weakly enhanced views and strongly enhanced views respectively, then... Defined as: .

10. The system according to claim 1, characterized in that, The classification unit described above uses ensemble prediction cross-entropy loss based on model history output to reduce the impact of noisy labels; let the first... The model predicts for each training epoch. Generate set predictions for each sample The previous historical forecasts are aggregated using an exponential moving average: ;in Represents the momentum coefficient; Then, both the original noise labels and the ensemble predictions are used as targets. The total training loss is: ;in and These are the weights for consistency loss and reliable comparison loss, respectively.

Images