A noise label robust training method and device for an image recognition model

Abstract

The application discloses a noise label robust training method and device for an image recognition model, and belongs to the technical field of image recognition model training, and comprises the following steps: inputting image training samples into a deep neural network for feature extraction; constructing a multi-granularity granular ball structure in a feature space based on a feature vector; performing hierarchical correction on the labels of the image training samples; performing label propagation in the granular ball, and calculating the propagation confidence distribution and consistency score of each image sample; screening a clean sample subset according to the consistency score, and iteratively training the deep neural network to update network parameters. Through the adaptive multi-granularity granular ball division mechanism, the application can fully depict the local structure characteristics of data in the feature space, avoid the problems of excessive fragmentation or insufficient purity caused by traditional fixed-granularity clustering in a high-noise environment, and provide a stable and reliable structure prior for subsequent label correction and information propagation.

Description

Technical Field

[0001] This invention belongs to the technical field of image recognition model training, specifically relating to a method and apparatus for robust training of image recognition models using noise labels. Background Technology

[0002] In image recognition, with the development of deep learning technology, supervised learning-based deep neural networks (DNNs) have made significant progress in tasks such as image classification and object recognition. However, supervised learning methods heavily rely on large-scale, high-quality manually labeled data. In practical applications, due to the high cost of manual annotation, subjective differences among annotators, errors in the automatic data collection process, and the ambiguity of samples in complex scenarios, training data often inevitably contains noisy labels, meaning that the observed labels of samples are inconsistent with their true labels. DNNs have strong fitting capabilities; as the training process progresses, the model gradually memorizes noisy labels, leading to overfitting of incorrect supervision signals and reducing generalization performance and robustness. Therefore, learning with noisy labels (LNL) has become a key research area, crucial for utilizing large-scale, imperfect data from real-world applications.

[0003] A mainstream LNL method is Sample Selection and Relabeling (SSR), which aims to identify potentially clean samples using criteria such as small loss or prediction consistency, and then correct mislabeled samples through model-driven relabeling. However, traditional small-loss-based selection methods tend to discard boundary and noisy samples, leading to significant information loss and decreased generalization performance. Furthermore, relying primarily on the model's high-confidence predictions to correct labels can reinforce erroneous labeling: if the model struggles to identify noisy labels in complex image environments early on, it increases the model's confidence in incorrect predictions, causing it to overfit these errors.

[0004] To address these limitations, this invention, based on model confidence correction, performs secondary correction of complex labels that are difficult for models to recognize from the feature layer from a multi-granular perspective. Its main contributions are as follows: 1) An adaptive multi-granularity clustering mechanism that analyzes features across multiple granularity levels, simultaneously capturing both coarse-grained and fine-grained data details. 2) A hierarchical label correction strategy that first corrects easily distinguishable noisy labels through model prediction, then performs coarse-grained relabeling by majority voting within the granular-ball (GB) to maintain intra-ball coherence and suppress prevalent errors. 3) A granular-ball label propagation component that uses similarity-weighted intra-ball voting, enabling probabilistic label diffusion to amplify reliable signals while diluting isolated noise. Summary of the Invention

[0005] The purpose of this invention is to address the above-mentioned shortcomings in the prior art by providing a method and apparatus for robust training of noisy labels for image recognition models, thereby solving the problems of high manual labeling costs, data acquisition errors, and label noise caused by sample ambiguity.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] Firstly, a method for robust training of noise-labeled image recognition models includes the following steps:

[0008] S1. Obtain image training samples containing noise labels;

[0009] S2. Input the image training samples into the deep neural network to extract features, and obtain the feature vector of the image in the feature space and the corresponding category prediction probability.

[0010] S3. Based on the feature vector, construct a multi-granularity sphere structure in the feature space, and divide the spheres using an adaptive multi-granularity sphere partitioning strategy based on the sphere purity.

[0011] S4. Combining model prediction confidence, particle purity, and particle structure information, perform hierarchical correction on the labels of the image training samples;

[0012] S5. Utilize the structural characteristics of the training samples in the internal image of the grain to perform label propagation within the grain, and calculate the propagation confidence distribution and consistency score for each image sample.

[0013] S6. Based on the consistency score, a clean sample subset is selected, and the deep neural network is iteratively trained using the clean sample subset and the corrected labels to update the network parameters.

[0014] This invention first uses a feature extraction network to perform forward propagation on image training samples containing noisy labels to obtain the vector representation of the samples in the feature space and the corresponding class prediction probability. Then, based on the vector representation, a multi-granularity sphere structure is constructed in the feature space. For each sphere, the sphere center, sphere purity, and similarity matrix between samples within the sphere are calculated. Based on this, hierarchical label correction is performed in combination with the prediction confidence: fine-grained label correction is performed on samples with high prediction confidence, and coarse-grained label correction is performed on samples with low confidence in high-purity spheres. Subsequently, during the particle label propagation stage, information such as sample similarity within the particle, particle purity, and distance from the sample to the center of the particle is used to propagate the corrected labels in a weighted manner, obtaining the propagation confidence distribution and consistency score of each sample in each category, which is used to distinguish relatively clean samples from suspected noise samples. Finally, the above label correction results and sample division are fed back into the subsequent model training. Supervised learning loss is used to update parameters for high-confidence samples, and unsupervised consistency constraints can be optionally combined to optimize the feature representation of all samples, thereby iteratively improving the robustness and generalization ability of the model in noisy label environments in an end-to-end manner.

[0015] Furthermore, S2 specifically includes:

[0016] Image training samples are input into a deep neural network. The feature extraction network maps the input image training samples to the feature space to obtain the feature vector of each image training sample. The feature vectors are normalized by norm to form a feature library. Then, the class prediction probability of the corresponding feature vector is obtained by classifying the class head and using the softmax function.

[0017] Furthermore, S3 specifically includes:

[0018] Multiple particles are obtained using the K-means method on the feature library. For each particle or sub-particle generated by the segmentation, the maximum number of recursive partitions is calculated.

[0019] If the current number of partitions does not exceed the maximum number of recursive partitions and the purity of the particle is lower than the preset purity threshold, then the internal features of the particle are partitioned using K-means to further divide it into two sub-particles; otherwise, the partitioning of the particle is stopped.

[0020] Furthermore, in S3, the maximum number of recursive partitions is calculated, which is expressed as:

[0021]

[0022] In the formula, To determine the maximum number of recursive partitions, The maximum iteration limit set globally. Indicates rounding down; The purity of the particles is expressed as:

[0023]

[0024] In the formula, Indicates category, Indicates the first Sample index in each sphere For the first Each sample category label, This is an indicator function.

[0025] Furthermore, in S4, the hierarchical correction includes fine-grained correction and coarse-grained correction;

[0026] First, fine-grained correction is performed, specifically as follows:

[0027]

[0028] In the formula, For more refined labeling, For the sample The maximum prediction confidence. To refine the threshold; For categories The predicted probability, For the first The labels of each sample;

[0029] If the maximum class prediction probability of the image training samples exceeds the preset fine-grained correction threshold Then the label of the training sample of the image is updated to the category with the highest predicted probability, thus obtaining a fine-grained corrected label;

[0030] Subsequently, coarse-grained correction is performed, specifically as follows: for image training samples that do not meet the fine-grained correction conditions, if the purity of their respective grain spheres is greater than a preset purity threshold, the label of the image training sample is updated to the weighted majority label within that grain sphere; otherwise, its label remains unchanged, specifically expressed as follows:

[0031]

[0032] In the formula, To correct the label, Weighted majority label; The purity threshold, This represents the average confidence level for all samples.

[0033] Furthermore, in S4, the weighted majority label It is represented as:

[0034]

[0035] In the formula, For granules The index is The sample of the first The original predicted scores for each category, The total number of samples.

[0036] Furthermore, S5 specifically includes the following sub-steps:

[0037] S51, For each category For those with correction labels The intrasphere similarity matrices of the image training samples are aggregated to obtain the similarity of each image training sample to each category. Contribution of score propagation:

[0038]

[0039] In the formula, Indicates granules The Middle Each element pairs the category The contribution of score propagation; For granules The middle label is the category. Sample index, This indicates that for each category granules The middle contains correction tags A subset of the sample index; For granules Inner The sample and the first Similarity between samples For granules Inner The propagation weight of each sample;

[0040] S52. Normalize the propagation contribution of each sample across all categories to obtain the propagation confidence distribution vector of that sample:

[0041]

[0042] In the formula, Indicates sample The probability of propagation; The total number of categories for the classification task. Indicates the first The original score vector of each sample, It is a constant;

[0043] S53. Calculate the consistency score based on the propagation confidence distribution vector.

[0044] Furthermore, in step S53, a consistency score is calculated based on the propagation confidence distribution vector, which is expressed as:

[0045]

[0046] in:

[0047]

[0048]

[0049] In the formula, To get the correct score, For the corrected label The number of votes, The maximum propagation score in category C. For the sample Correction tags The probability value in the propagation probability distribution. For the sample Belongs to the The probability value after the class is propagated.

[0050] Furthermore, in step S6, a clean sample subset is selected based on the consistency score, including: if the correct score is obtained... If the score is greater than or equal to the scoring threshold, it means the propagation result is consistent with the corrected label, and the sample is a clean sample; if the score is correct... If the score is below the scoring threshold, the propagation result deviates from the corrected label, and the sample is considered to contain noisy labels.

[0051]

[0052]

[0053] In the formula, An index containing clean samples suitable for training. An index containing samples labeled as noise. This is the scoring threshold.

[0054] Secondly, a noise-label robust training device for an image recognition model includes:

[0055] The image acquisition module is used to acquire image training samples containing noise labels;

[0056] The feature extraction module is used to input image training samples into a deep neural network for feature extraction, and obtain the feature vector of the image in the feature space and the corresponding class prediction probability.

[0057] The particle-sphere construction module is used to construct a multi-granularity particle-sphere structure in the feature space based on the feature vector, and to divide the particles based on the particle purity using an adaptive multi-granularity particle-sphere partitioning strategy.

[0058] The hierarchical correction module is used to perform hierarchical correction on the labels of the image training samples by combining the model prediction confidence, particle purity and particle structure information;

[0059] The sphere label propagation module is used to propagate labels within the sphere by utilizing the structural characteristics of the training samples in the image inside the sphere and the corrected labels, and to calculate the propagation confidence distribution and consistency score for each image sample.

[0060] The model update module is used to select a clean sample subset based on the consistency score, and to use the clean sample subset and the corrected labels to iteratively train the deep neural network to update the network parameters.

[0061] The noise-label robust training method and apparatus for image recognition models provided by this invention have the following beneficial effects:

[0062] This invention integrates an adaptive multi-granularity clustering mechanism, a hierarchical label correction strategy, and sphere label propagation. It accurately characterizes the local structure of data from a multi-granularity perspective, combines model prediction confidence to achieve coarse-to-fine granular collaborative label correction, and utilizes intra-sphere sample similarity weighted propagation to amplify reliable signals and dilute isolated noise. This method can automatically distinguish between clean samples and suspected noise samples, achieving end-to-end iterative optimization. Evaluations on datasets such as Animal-10N and CIFAR-10 demonstrate its excellent label correction performance and model generalization ability under varying noise levels.

[0063] This invention, for image recognition, is based on three core modules: sphere construction, hierarchical label correction, and sphere label propagation. It constructs a systematic solution for noisy labeling scenarios within the existing deep neural network framework. Through an adaptive multi-granularity sphere partitioning mechanism, this invention can fully characterize the local structural features of data in the feature space, avoiding the excessive fragmentation or insufficient purity problems caused by traditional fixed-granularity clustering in high-noise environments. This provides a stable and reliable structural prior for subsequent label correction and information propagation. Attached Figure Description

[0064] Figure 1 This is a diagram of the overall architecture of a noise-label robust training method for an image recognition model in an embodiment of the present invention. In the diagram, CAGM stands for Confidence-Aware Granular-Ball Mitigator.

[0065] Figure 2 This is a diagram showing the effect of coarse-grained and fine-grained correction on the Animal-10N real noise dataset in an embodiment of the present invention. Figure 2 In the diagram, (a) represents the global distribution of model prediction confidence scores for all samples. Figure 2 (b) in the example is fine-grained correction. Figure 2 (c) in the example is coarse-grained correction.

[0066] Figure 3 These are visualizations of the effects of using various noise correction methods individually on the CIFAR-10 dataset under different noise environments, as described in this invention. Figure 3 (a) in the dataset is the original noisy dataset with 40% noise. Figure 3 (b) in the figure represents the clean subset obtained after multi-granularity correction under 40% noise. Figure 3 In the example, (c) represents the clean subset obtained after coarse-grained correction under 40% noise. Figure 3 In this context, (d) represents the clean subset obtained after fine-grained correction under 40% noise. Figure 3 (e) in the dataset represents the original noisy dataset with 60% noise. Figure 3 In the example, (f) represents the clean subset obtained after multi-granularity correction under 60% noise. Figure 3 In this context, (g) represents the clean subset obtained after coarse-grained correction under 60% noise. Figure 3 In this context, (h) represents the clean subset obtained after fine-grained correction under 60% noise. Figure 3 In this context, (i) represents the original noise dataset with 80% noise. Figure 3 In this context, (j) represents the clean subset obtained after multi-granularity correction under 80% noise. Figure 3 In this context, (k) represents the clean subset obtained after coarse-grained correction under 80% noise. Figure 3 (l) is the clean subset obtained after fine-grained correction under 80% noise.

[0067] Figure 4 This is a pie chart showing the performance indicators of different correction methods under extreme noise conditions in embodiments of the present invention.

[0068] Figure 5 This is a flowchart of a noise-label robust training method for an image recognition model in an embodiment of the present invention. Detailed Implementation

[0069] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0070] This embodiment provides a robust training method for noise-labeled image recognition models. This method centers on particle sphere computation, introducing model prediction confidence and multi-granularity particle sphere structure information to perform hierarchical label correction and label propagation on noisy samples. This effectively suppresses the adverse effects of noise labels while fully utilizing image training samples. In this invention, image training samples are first feature-encoded using a DNN to obtain prediction confidence, which characterizes the distribution of samples in the feature space and the reliability of the model's prediction results.

[0071] like Figure 1 As shown, this embodiment first uses a DNN to extract the original features of the image to form a feature library, and then calculates the class probability using the softmax function. Subsequently, based on the feature representation of the samples, the samples are clustered in the feature space to construct multiple sphere units, each containing several samples that are close to each other in the feature space. After completing the sphere construction, the sample labels are hierarchically corrected by combining the model prediction confidence and the sphere structure information. After completing the coarse-grained and fine-grained label corrections respectively, the corrected label information is propagated using the similarity relationship between samples within the sphere, referencing... Figure 5 Specifically, it includes the following:

[0072] S1. Obtain image training samples containing noise labels;

[0073] S2. Input the image training samples into the deep neural network to extract features, and obtain the feature vector of the image in the feature space and the corresponding category prediction probability.

[0074] For details, please refer to Figure 1 The framework shown first uses image training samples containing noise labels during the noise evaluation phase of each training round. ,in For the first One sample, For the first The label of each sample The total number of samples in the dataset is input into the deep neural network classification model. This is then processed by the feature extraction network. The input samples are mapped to the feature space to obtain the feature vector representation of each sample. , This is the feature extractor obtained by removing the fully connected layers from a CNN model. The dimension of the original feature embedding. Given the set of real numbers; then these eigenvectors are passed through... The norm is normalized to form a feature library. ,in, , for The feature vector after normalization by norm; at the same time, the classification head (Linear layer) processes raw features to generate , Classified by CNN model , The total number of categories for the classification task; finally, the softmax function is used to calculate the total number of categories. Calculate the category prediction probability .

[0075] S3. Construct a multi-granularity sphere structure in the feature space based on the feature vector, and divide the spheres based on the purity of the spheres using an adaptive multi-granularity sphere partitioning strategy.

[0076] During the formation of granules, in the feature library Obtained using the K-means method For each granular-ball (GB) or the sub-granular-balls generated by the division (a low-purity sphere is recursively divided into two sub-spheres), calculate the maximum number of recursive divisions.

[0077] If the current number of partitions does not exceed the maximum number of recursive partitions and the particle purity is lower than the preset purity threshold. Or the sample size of the sub-balls is less than the minimum sample size. If the internal features of the particle are correctly identified, K-means partitioning is used to further divide it into two sub-particles, where each sub-particle inherits the adaptive constraints of its parent particle, thus maintaining purity information constraints throughout the hierarchical structure; otherwise, the partitioning of the particle is stopped.

[0078] In one specific embodiment, let Represents the final sphere index, where:

[0079] Indicates the first Sample index in a sphere This indicates the number of spheres. For each GB, its center... and distance The calculation is as follows:

[0080]

[0081]

[0082] In the formula, For granules The Middle Feature vector of each sample To the center of the sphere Euclidean distance, For granules The Middle The feature vector corresponding to each sample This represents the Euclidean norm.

[0083] GB purity It is a key metric for label consistency, and its definition is as follows:

[0084]

[0085] In the formula, Indicates the category, referring to the range of values ​​for all categories. For the first Each sample category label, For indicator functions, This refers to the purity of the particles.

[0086] Intrasphere similarity matrix Represented as:

[0087]

[0088] In the formula, The characteristic matrix is ​​represented as follows: ; for Transpose of;

[0089] To adapt to data distributions with varying noise intensities, this embodiment employs an adaptive multi-granularity sphere partitioning strategy. For each initial sphere, the maximum number of recursive partitioning iterations is adaptively set based on its purity.

[0090]

[0091] In the formula, To determine the maximum number of recursive partitions, The maximum iteration limit set globally. This indicates rounding down; higher purity spheres correspond to larger ones. This allows for deeper subdivision to obtain a finer local structure; while lower purity spheres limit their subdivision depth to avoid over-fragmentation of small spheres in high-noise regions.

[0092] S4. Combining model prediction confidence, particle purity, and particle structure information, hierarchical correction is performed on the labels of image training samples.

[0093] Hierarchical correction includes fine-grained correction and coarse-grained correction;

[0094] First, fine-grained correction is performed. Specifically, if the maximum class prediction probability of the image training sample exceeds the preset fine-grained threshold, the label of the image training sample is updated to the class with the maximum prediction probability, thereby obtaining the fine-grained corrected label.

[0095] In one specific embodiment, for training samples Each sample If the maximum predicted probability exceeds the refinement threshold Then the label is updated to the predicted category:

[0096]

[0097] In the formula, For more refined labeling, For the sample The maximum prediction confidence. To refine the threshold; For categories The predicted probability;

[0098] Thus, this embodiment can directly utilize the model's discriminative power to cover noisy labels for high-confidence instances, thereby establishing an initial refined label set. These refined labels laid the foundation for subsequent label propagation, enabling high-confidence corrections to spread throughout the entire dataset.

[0099] Then, coarse-grained correction is performed, which is as follows: for image training samples that do not meet the fine-grained correction conditions, if the purity of the sphere to which the image training sample belongs is greater than the preset purity threshold, the label of the image training sample is updated to the weighted majority label within the sphere; otherwise, the label remains unchanged.

[0100] After fine-grained refinement, the labels are refined. The scoring matrix is ​​calculated by passing the ball within the ball using a ball-tag propagation method. These scores are used to guide coarse-grained adjustments;

[0101] In one specific embodiment, for each sphere, the weighted majority label is calculated as the category with the highest average score across all samples within the sphere:

[0102]

[0103] In the formula, For granules The index is The sample of the first The original predicted scores for each category, The total number of samples, This is a weighted majority label.

[0104] Subsequently, after meeting the purity requirements granules In this process, samples that meet the following conditions are selectively corrected. Tags:

[0105]

[0106] In the formula, To correct the labels; The purity threshold, This represents the average confidence level for all samples.

[0107] The coarse-grained label correction strategy described above reflects the following principles: for samples with low model confidence but located inside high-purity spheres, it tends to trust most of the structural information inside the spheres, that is, to correct them by using the spheres to represent the category; for samples with high model confidence or insufficient sphere purity, it retains the fine-grained correction results or the original labels, thereby avoiding over-reliance on unreliable sphere structures.

[0108] like Figure 2 As shown, Figure 2 In (a), the global distribution of model prediction confidence for all samples is shown. The gray dashed line indicates the fine-grained correction threshold. The samples with higher confidence to the right are subject to fine-grained label correction, while the coarse-grained correction is shown to the left of the dashed line, which supplements the label correction for samples with lower confidence. Figure 2 (b) in the example is fine-grained correction. Figure 2 (c) in the example is a coarse-grained correction, showing how to effectively relabel a hamster (label 8) as a guinea pig (label 9) when the fine-grained method misses out.

[0109] S5. Utilize the structural characteristics of the training samples in the internal image of the grain to perform label propagation within the grain, and calculate the propagation confidence distribution and consistency score for each image sample.

[0110] To further optimize label correction and improve the reliability of sample selection, a sphere label propagation mechanism is introduced. This mechanism utilizes the structural characteristics of spheres to propagate label information between samples. The obtained raw prediction scores are then aggregated into a score matrix. Then, through normalized thresholding... Calculating the reliability index for distinguishing clean samples from noisy samples involves the following sub-steps:

[0111] S51, For each category ,Sure China has correction labels A subset of the sample indexes, denoted as The intrasphere similarity matrix of the sample subset is aggregated to obtain the training sample of each image for each category. Contribution of score propagation:

[0112]

[0113] In the formula, Indicates granules Each sample pair is a category The contribution of score propagation Indicates granules The Middle Each element pairs the category The contribution of score propagation; For granules The middle label is the category. Sample index, This indicates that for each category Sample Index The middle contains correction tags A subset of the sample index; For granules Inner The sample and the first Similarity between samples For granules Inner The propagation weight of each sample.

[0114] This can be equivalently represented in matrix form as follows: ,in yes The submatrix whose column indices are ,and yes The subvector whose element index is The summation form highlights the diffusion of label confidence, where the confidence originates from the labeling as a category. The contributions of similar and reliable samples were aggregated together, and then... The sum of the predicted scores of the middle sample across all categories is the matchable sphere. The propagation score matrix:

[0115]

[0116] In the formula, For granules The propagation score matrix, Indicates granules The propagation contribution vector of each sample to the score of category 1;

[0117] S52. Normalize the propagation contribution of each sample across all categories; let Indicates the first The original score vector of each sample. Specifically, for each sample, this sample vector is transformed into a probability distribution through row-wise normalization:

[0118]

[0119] In the formula, Indicates sample The probability of propagation, ; The total number of categories for the classification task. Indicates the first The original score vector of each sample, It is a small constant added to the denominator to prevent division by zero when the original fraction is all zeros;

[0120] S53. Calculate the consistency score based on the propagation confidence distribution vector:

[0121] To further quantify the consistency between the propagation results and the corrected labels, a sample consistency score can be constructed based on the propagation probability distribution to assist in subsequent sample selection and model training. For each sample, the voting value of the corrected label is extracted as follows: , indicating that it is assigned to The propagation probability quality. Maximum propagation score. , representing the maximum propagation score in category C. Correct score. Defined as the following ratio:

[0122]

[0123] In the formula, To get the correct score, For the corrected label The number of votes, To achieve the maximum propagation score, For the sample Correction tags The probability value in the propagation probability distribution. For the sample Belongs to the The probability value after the class is propagated.

[0124] S6. Select a clean sample subset based on the consistency score, and use the clean sample subset and the corrected labels to iteratively train the deep neural network to update the network parameters.

[0125] A clean subset of samples was selected based on the consistency score, including: those with correct scores. If the score is greater than or equal to the scoring threshold, it means the propagation result is consistent with the corrected label, and the sample is a clean sample; if the score is correct... If the score is below the scoring threshold, the propagation result deviates from the corrected label, and the sample is considered to contain noisy labels.

[0126]

[0127]

[0128] In the formula, An index containing clean samples suitable for training. An index containing samples labeled as noise. This is the scoring threshold. In practice, The value is fixed at 1 to ensure that only samples with full confidence alignment are selected into the clean subset, thereby minimizing the risk of including residual noisy instances during training.

[0129] This embodiment conducted a systematic ablation experiment on a multi-granularity label correction strategy, and examined the impact of three configurations—using only fine-grained correction, using only coarse-grained correction, and using multi-granularity joint correction—on indicators such as the error label correction effect and the size of the constructed clean subset.

[0130] This embodiment uses the t-SNE visualization method to analyze CIFAR-10 features under different label noise levels. For example... Figure 3 As shown, at all noise levels, the full correction method consistently generates a larger clean subset and achieves more accurate label correction compared to methods using only coarse-grained correction or only fine-grained correction.

[0131] Figure 4The pie charts illustrate the proportion of label selection results in the final training epochs for the three correction strategies on the CIFAR-10 and CIFAR-100 datasets under 80% and 90% symmetric noise conditions. In noisy label learning, these metrics are defined as follows: True Positive (TP) represents a correctly identified clean label, False Positive (FP) represents a noisy label misclassified as a clean label, True Negative (TN) represents a correctly identified noisy label, and False Negative (FN) represents a clean label misclassified as a noisy label. The visualizations show that the full correction method consistently achieves a higher TP ratio and a lower FP ratio, especially in high-noise environments. In this case, fine-grained correction leads to higher FP due to its oversensitivity to noise, while coarse-grained correction, due to its conservative nature, has a higher TN but a lower TP. This balanced distribution in multi-granularity correction further validates its robustness and effectiveness, as it minimizes misclassification while maximizing the preservation of reliable labels, thus improving the overall performance of the model in challenging noisy environments.

[0132] This embodiment also provides a noise-label robust training device for image recognition models, including:

[0133] The image acquisition module is used to acquire image training samples containing noise labels;

[0134] The feature extraction module is used to input image training samples into a deep neural network for feature extraction, and obtain the feature vector of the image in the feature space and the corresponding class prediction probability.

[0135] The particle-sphere construction module is used to construct a multi-granularity particle-sphere structure in the feature space based on the feature vector, and to divide the particles based on the particle purity using an adaptive multi-granularity particle-sphere partitioning strategy.

[0136] The hierarchical correction module is used to perform hierarchical correction on the labels of the image training samples by combining the model prediction confidence, particle purity and particle structure information;

[0137] The sphere label propagation module is used to propagate labels within the sphere by utilizing the structural characteristics of the training samples in the image inside the sphere and the corrected labels, and to calculate the propagation confidence distribution and consistency score for each image sample.

[0138] The model update module is used to select a clean sample subset based on the consistency score, and to use the clean sample subset and the corrected labels to iteratively train the deep neural network to update the network parameters.

[0139] Although specific embodiments of the invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of this patent. Various modifications and variations that can be made by a person skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of this patent.

Claims

1. A method for robust training of image recognition models using noise labels, characterized in that, Includes the following steps: S1. Obtain image training samples containing noise labels; S2. Input the image training samples into the deep neural network to extract features, and obtain the feature vector of the image in the feature space and the corresponding category prediction probability. S3. Based on the feature vector, construct a multi-granularity particle structure in the feature space, and divide the particles based on the particle purity using an adaptive multi-granularity particle partitioning strategy. S4. Combining model prediction confidence, particle purity, and particle structure information, perform hierarchical correction on the labels of the image training samples; S5. Utilize the structural characteristics of the training samples in the internal image of the grain to perform label propagation within the grain, and calculate the propagation confidence distribution and consistency score for each image sample. Specifically, it includes the following sub-steps: S51, For each category For those with correction labels The intrasphere similarity matrices of the image training samples are aggregated to obtain the similarity of each image training sample to each category. Contribution of score propagation: In the formula, Indicates granules The Middle Each element pairs the category The contribution of score propagation; For granules The middle label is the category. Sample index, This indicates that for each category granules The middle contains correction tags A subset of the sample index; For granules Inner The sample and the first Similarity between samples For granules Inner The propagation weight of each sample; S52. Normalize the propagation contribution of each sample across all categories to obtain the propagation confidence distribution vector of that sample: In the formula, Indicates sample The probability of propagation; The total number of categories for the classification task. Indicates the first The original score vector of each sample, It is a constant; S53. Calculate the consistency score based on the propagation confidence distribution vector; S6. Based on the consistency score, a clean sample subset is selected, and the deep neural network is iteratively trained using the clean sample subset and the corrected labels to update the network parameters.

2. The method for robust training of noise labels for image recognition models according to claim 1, characterized in that, S2 specifically includes: Image training samples are input into a deep neural network. The feature extraction network maps the input image training samples to the feature space to obtain the feature vector of each image training sample. The feature vectors are normalized by norm to form a feature library. Then, the class prediction probability of the corresponding feature vector is obtained by classifying the class head and using the softmax function.

3. The method for robust training of noise labels for image recognition models according to claim 2, characterized in that, S3 specifically includes: Multiple spheres are obtained using the K-means method on the feature library. For each sphere or sub-sphere generated by the segmentation, the maximum number of recursive partitions is calculated. If the current number of partitions does not exceed the maximum number of recursive partitions and the purity of the particle is lower than the preset purity threshold, then the internal features of the particle are partitioned using K-means to further divide it into two sub-particles; otherwise, the partitioning of the particle is stopped.

4. The method for robust training of noise labels for image recognition models according to claim 3, characterized in that, In S3, the maximum number of recursive partitions is calculated, which is expressed as: In the formula, To determine the maximum number of recursive partitions, The maximum iteration limit set globally. Indicates rounding down; The purity of the particles is expressed as: In the formula, Indicates category, Indicates the first Sample index in each sphere For the first Each sample category label, This is an indicator function.

5. The method for robust training of image recognition models with noise labels according to claim 4, characterized in that, In S4, the hierarchical correction includes fine-grained correction and coarse-grained correction; First, fine-grained correction is performed, specifically as follows: In the formula, For more refined labeling, For the sample The maximum prediction confidence. To refine the threshold; For categories The predicted probability, For the first The labels of each sample; If the maximum class prediction probability of the image training samples exceeds the preset fine-grained correction threshold Then the label of the training sample of the image is updated to the category with the highest predicted probability, thus obtaining a fine-grained corrected label; Subsequently, coarse-grained correction is performed, specifically as follows: for image training samples that do not meet the fine-grained correction conditions, if the purity of their respective grain spheres is greater than a preset purity threshold, the label of the image training sample is updated to the weighted majority label within that grain sphere; otherwise, its label remains unchanged, specifically expressed as follows: In the formula, To correct the label, Weighted majority label; The purity threshold, This represents the average confidence level for all samples.

6. The method for robust training of image recognition models with noise labels according to claim 5, characterized in that, In S4, the weighted majority label It is represented as: In the formula, For granules The index is The sample pair The original predicted scores for each category, The total number of samples.

7. The method for robust training of image recognition models with noise labels according to claim 1, characterized in that, In step S53, a consistency score is calculated based on the propagation confidence distribution vector, which is expressed as follows: in: In the formula, To get the correct score, For the corrected label The number of votes, The maximum propagation score in category C. For the sample Correction tags The probability value in the propagation probability distribution. For the sample Belongs to the The probability value after the class is propagated.

8. The method for robust training of image recognition models with noise labels according to claim 1, characterized in that, In step S6, a clean sample subset is selected based on the consistency score, including: if the score is correct. If the score is greater than or equal to the scoring threshold, it means the propagation result is consistent with the corrected label, and the sample is a clean sample; if the score is correct... If the score is below the scoring threshold, the propagation result deviates from the corrected label, and the sample is considered to contain noisy labels. In the formula, An index containing clean samples suitable for training. An index containing samples labeled as noise. This is the scoring threshold.

9. A noise-label robust training device for image recognition models, characterized in that, include: The image acquisition module is used to acquire image training samples containing noise labels; The feature extraction module is used to input image training samples into a deep neural network for feature extraction, and obtain the feature vector of the image in the feature space and the corresponding class prediction probability. The particle-sphere construction module is used to construct a multi-granularity particle-sphere structure in the feature space based on the feature vector, and to divide the particles based on the particle purity using an adaptive multi-granularity particle-sphere partitioning strategy. The hierarchical correction module is used to perform hierarchical correction on the labels of the image training samples by combining the model prediction confidence, particle purity and particle structure information; The sphere label propagation module is used to propagate labels within the sphere by utilizing the structural characteristics of the training samples in the image inside the sphere and the corrected labels, and to calculate the propagation confidence distribution and consistency score for each image sample. Specifically, it includes: For each category For those with correction labels The intrasphere similarity matrices of the image training samples are aggregated to obtain the similarity of each image training sample to each category. Contribution of score propagation: In the formula, Indicates granules The Middle Each element pairs the category The contribution of score propagation; For granules The middle label is the category. Sample index, This indicates that for each category granules The middle contains correction tags A subset of the sample index; For granules Inner The sample and the first Similarity between samples For granules Inner The propagation weight of each sample; Normalize the propagation contribution of each sample across all classes to obtain the propagation confidence distribution vector of that sample: In the formula, Indicates sample The probability of propagation; The total number of categories for the classification task. Indicates the first The original score vector of each sample, It is a constant; Calculate the consistency score based on the propagation confidence distribution vector; The model update module is used to select a clean sample subset based on the consistency score, and to use the clean sample subset and the corrected labels to iteratively train the deep neural network to update the network parameters.

Images