A dual-network cooperation-based anti-noise label image recognition method

Abstract

The application discloses an anti-noise label image recognition method based on double-network cooperation, and the steps include: 1, noise label data set acquisition and pretreatment; 2, using noise data set to preheat the network; 3, calculating the confidence of each sample and obtaining the category adaptive threshold; 4, screening samples and constructing multiple subsets; 5, training any model based on the semi-supervised learning paradigm, thereby obtaining the final double-network image recognition model for classifying input pictures. The application can effectively resist the influence of noise labels, reduce the fitting of the model to noise labels, extract the real semantic features of samples and improve the robustness and classification accuracy of the model under the condition of training using a data set containing a large number of noise labels.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and more specifically to image recognition technology under noise labels. Background Technology

[0002] In the field of computer vision, deep learning technology has made significant progress, especially in tasks such as image classification, object detection, and image segmentation. However, the success of deep neural networks largely depends on large-scale, high-quality annotated datasets, which require substantial investment of manpower, resources, and time to acquire. In contrast, using inexpensive alternatives (such as crowdsourcing or web scraping) can significantly reduce the cost of acquiring training data, promoting the maturity and low-cost deployment of image recognition technology. However, using this data also brings other problems: most of this training data comes from crowdsourcing or is directly scraped from the internet using keywords, so the quality of its labels cannot be guaranteed. Compared to image annotations performed by professionals, the labels corresponding to this data are likely to be incorrect, i.e., noisy labels. Previous studies have shown that deep neural networks easily fit these noisy labels, resulting in compromised model robustness and reduced generalization performance. This problem becomes particularly serious in practical applications, such as medical image analysis, autonomous driving, and video surveillance, where datasets typically contain a high proportion of label noise. Therefore, noisy label learning is receiving increasing attention from researchers.

[0003] To combat the negative impact of noisy labels, existing methods mainly fall into three categories: 1. Sample selection-based methods: These methods select clean samples from noisy datasets to obtain clean and noisy sets, and then use semi-supervised learning for subsequent training. 2. Label correction-based methods: These methods correct the noisy labels in the noisy dataset to obtain truly clean labels for model training. 3. Regularization-based methods: These methods mitigate the side effects of label noise by preventing deep neural networks from overfitting all training samples, such as designing noise-robust loss functions and regularization methods. However, these methods often achieve good results in low-noise-rate scenarios but perform poorly in extreme noise-rate scenarios. Therefore, there is still significant room for improvement in image recognition tasks under noisy label scenarios. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a noise-resistant label image recognition method based on dual-network collaboration. This method aims to effectively combat the influence of noise labels and reduce the model's fitting to noise labels when trained on a dataset containing a large number of noise labels, thereby extracting the true semantic features of the samples and improving the robustness and classification accuracy of the model.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0006] The present invention provides a noise-resistant tag image recognition method based on dual-network cooperation, characterized by the following steps:

[0007] Step 1: Obtain the dataset containing noisy labels, and denote any sample-label pair in the dataset as (x... i ,y i ), where x i Let y represent the i-th sample. i x represents i Corresponding noise label; y i ∈{1,2,…,C}, where C represents the total number of categories;

[0008] Step 2: Construct two networks with the same structure to form a dual network. Each network consists of a feature extractor e, a classification head f, and a mapping head h. Train the two networks separately using a dataset with noisy labels to obtain two preheated models with different parameters.

[0009] Step 3: Calculate the model assignment for the i-th sample x. i The confidence level is calculated and the adaptive threshold τ for the c-th class is obtained. c ;

[0010] Step 4: Filter samples and construct multiple subsets, including the noise set of each model itself. and Clean Set and separate difficult subsets ;

[0011] Step 5: Train the two models based on the semi-supervised learning paradigm to obtain the two optimal image recognition networks and form the final dual-network image recognition model for classifying input images.

[0012] The noise-resistant tag image recognition method based on dual-network cooperation described in this invention is characterized in that step 3 includes the following steps:

[0013] Step 3.1: Use equation (1) to calculate the assignment of any model to x. i confidence level s i :

[0014] (1)

[0015] In equation (1), Indicates y i The one-hot vector, Describes any one of the models in the bimodel for x. i The predicted probability distribution;

[0016] Step 3.2: Based on the category corresponding to the noise label of each sample, cluster the confidence scores of samples corresponding to the same category into one class, thereby obtaining the confidence score set under any C categories assigned by the model. ,in, Let represent the set of confidence scores for the c-th class under any given model assignment;

[0017] Step 3.3: Calculate the adaptive threshold τ assigned to the c-th class by any model using equation (2). c :

[0018] (2)

[0019] In equation (2), and The confidence sets for the c-th category are respectively. The mean and maximum value; denoted as temperature coefficient, and µ as hyperparameter.

[0020] Step 4 includes the following steps:

[0021] Step 4.1: For any model, if s i ∈ >τ c Then x i The sample is identified as clean and added to the clean set of its own model. If so, otherwise, x i Identify samples as noise and add them to the noise set of your own model. middle;

[0022] Step 4.2: For the other model, follow the same procedure as in Step 4.1 for x. i The model is then judged to obtain its clean set and noise set.

[0023] If the two models are related to x i Different judgment results will result in x i Additionally, it was added to the difficult subset. middle;

[0024] Step 5 includes the following steps:

[0025] Step 5.1, when x i ∈ Then, the noise robustness loss function of any model can be constructed using equation (3). :

[0026] (3)

[0027] In equation (3), For the difficult collection The number of samples in the dataset, where q is a hyperparameter and q∈(0,1];

[0028] Step 5.2: Apply strong data augmentation and weak data augmentation to x respectively. i After processing, a strongly enhanced view is obtained. and weak enhancement view ;

[0029] Will The data are input into two separate models for processing, and the predicted probability distribution of the first network for the strongly augmented view is output separately. The prediction probability distribution of the second network for the strongly augmented view ;

[0030] Will The data are input into two separate models for processing, and the second network outputs the prediction probability distribution of the weakly augmented view. The prediction probability distribution of the second network for the weakly augmented view ;

[0031] Construct the dual consistency loss function of the bi-model using equation (4). :

[0032] (4)

[0033] In equation (4), KL represents the Kullback-Leibler divergence; N represents the number of samples in the dataset;

[0034] Step 5.3, when x i ∈ Using samples x from the clean set of any model i Weakly enhanced view and enhanced view Input into another model for processing, thereby utilizing equation (5) to process the noise label y i After weighting, soft labels are obtained. :

[0035] (5)

[0036] In equation (5), Let represent the mean of the predicted probability distributions of the strong and weak augmented views output by the m-th model, and m∈{1,2};

[0037] Then use formula (6) for soft labels Perform a sharpening operation to obtain x i New soft label :

[0038] (6)

[0039] In equation (6), T is the temperature coefficient to be adjusted; for The value of the c-th position;

[0040] Step 5.4, when x i ∈ Using samples x from the noise set of any model i Weakly enhanced view and enhanced view The input is processed in another model, and x is calculated using equation (7). i pseudo-tags :

[0041] (7)

[0042] Step 5.5, x i New soft label and x i pseudo-tags Let any one of the labels be denoted as x. i New label , will x j New soft label and x j New pseudo-labels Let any one of the labels be denoted as x. j New label Thus, equations (8) and (9) are used for any pair of samples (x) i x j The new tag corresponding to ) , Perform a MixUp operation to obtain a mixed sample. Mixed tags ;

[0043] (8)

[0044] (9)

[0045] In equations (8) and (9), Indicates mixed weights, and ,in, Represents the weights, and follows the parameter . The Beta distribution;

[0046] Step 5.6, if x i ∈ Then the mixed samples and its mixed labels Add to the clean set In the middle, if x i ∈ Then the mixed samples and its mixed labels Add mixed noise set In this way, a clean mixture set of any model can be constructed. and mixed noise set ;

[0047] Step 5.7: For the other model, process it in the same way as steps 5.3-5.6 to construct the mixed clean set and mixed noise set of the other model;

[0048] Step 5.8, when x i ∈ Then, the cross-entropy loss function of any model can be constructed using equation (10). :

[0049] (10)

[0050] In equation (10), for The value of the c-th position in the middle. for The value of the c-th position;

[0051] When x i ∈ Then, the mean squared error loss function of any model can be constructed using equation (11). :

[0052] (11)

[0053] Construct a semi-supervised loss using equation (12) :

[0054] (12)

[0055] In equation (12), for Weighting coefficients;

[0056] Step 5.9, when x i ∈ Using any sample x from the noise set of any model i Weakly enhanced view and enhanced view The feature extractor e and mapping head h of another model are processed to obtain x. iWeak augmented view features and enhanced view features ;

[0057] When x j ∈ Using any other sample x from the noise set of any model j Weakly enhanced view and enhanced view The feature extractor e and mapping head h of another model are processed to obtain x. j Weak augmented view features and enhanced view features Thus, the contrastive loss of any model can be constructed using equations (13), (14), and (15). :

[0058] (13)

[0059] (14)

[0060] (15)

[0061] In equations (13), (14), and (15), for and The cosine similarity between them, where t is the temperature coefficient; It is an indicator function, when season =1, otherwise, let =0;

[0062] Step 5.10: Construct the total loss function for any model using equation (16). ;

[0063] (16)

[0064] In equation (16), , and They are , and Weighting coefficients;

[0065] Step 5.11: Optimize each noise-resistant label image recognition model using the gradient descent method, so that... By minimizing the minimum, two optimal image recognition networks are obtained.

[0066] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the noise-resistant tag image recognition method, and the processor is configured to execute the program stored in the memory.

[0067] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs the steps of the noise-resistant tag image recognition method.

[0068] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0069] 1. This invention introduces a class-adaptive threshold for sample selection. This design leverages the established consensus that samples exhibit varying degrees of fitting difficulty across different classes. By analyzing the probability distribution of all samples and their corresponding classes, a class-adaptive threshold is determined to select reliable samples. This design helps separate clean and noisy samples, thereby reducing the impact of confirmation bias and improving the model's accuracy in image classification.

[0070] 2. In addition to the clean set and noisy set used in previous methods, this invention constructs a hard set, which includes samples that the two networks disagree on—samples that one network considers clean while the other considers noisy. These samples are close to the decision boundary and have the potential to improve the generalization ability of the DNN. Furthermore, the noise-robust loss introduced on these hard samples allows the model to cautiously utilize potentially clean labels rather than replacing them with pseudo-labels, avoiding information loss and confirmation bias that pseudo-labels may introduce, thereby further improving the robustness of the model and the accuracy of classification.

[0071] 3. This invention proposes a dual consistency loss, which constrains the consistency of predictions from two networks for different augmented views of the same sample. This allows the two networks to learn from each other, thus sharing and utilizing training experience from different sources during training. In other words, through mutual learning, errors caused by biased selections can be mutually reduced by the peer networks, resulting in better generalization and significantly improving the model's image recognition capabilities. Attached Figure Description

[0072] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0073] In this embodiment, as Figure 1 As shown, a noise-resistant label image recognition method based on dual-network cooperation includes the following steps:

[0074] Step 1: Obtain the dataset containing noisy labels, and denote any sample-label pair in the dataset as (x... i ,y i ), where x i Let y represent the i-th sample. i x represents i Corresponding noise label; y i ∈{1,2,…,C}, where C represents the total number of categories; in this embodiment, the selected dataset containing noisy labels is obtained by artificially injecting noisy labels into the clean dataset CIFAR-10 / 100, where C is 10 in the CIFAR10 dataset and C is 100 in the CIFAR100 dataset. Specifically, the sample labels of each category are randomly shuffled according to the noise rate to obtain labels of other categories, thereby obtaining symmetrical noise.

[0075] Step 2: Construct two networks with identical structures to form a dual network. Each network consists of a feature extractor e, a classification head f, and a mapping head h. Train both networks separately using a dataset containing noisy labels to obtain two pre-warmed models with different parameters. In this embodiment, the noisy CIFAR dataset is used to train the model E according to the general paradigm of image classification, i.e., using the cross-entropy loss function. w Rounds. For the CIFAR10 dataset, the number of warm-up rounds is 10. For the CIFAR100 dataset, the number of warm-up rounds is 100.

[0076] Step 3: Calculate the confidence score for each sample and obtain the class-adaptive threshold.

[0077] Step 3.1: Use equation (1) to calculate the assignment of any model to x. i confidence level s i :

[0078] (1)

[0079] In equation (1), Indicates y i The one-hot vector, Describes any one of the models in the bimodel for x. i The predicted probability distribution;

[0080] Step 3.2: Based on the category corresponding to the noise label of each sample, cluster the confidence scores of samples corresponding to the same category into one class, thereby obtaining the confidence score set under any C categories assigned by the model. ,in, Let represent the set of confidence scores for the c-th class under any given model assignment;

[0081] Step 3.3: Calculate the adaptive threshold τ assigned to the c-th class by any model using equation (2). c :

[0082] (2)

[0083] In equation (2), and The confidence sets for the c-th category are respectively. The mean and maximum value; Here, µ is the temperature coefficient, and µ is the hyperparameter; in this embodiment, It is 2. It is 0.3.

[0084] Step 4: Filter the samples and construct multiple subsets:

[0085] Step 4.1: For any model, if s i ∈ >τ c Then x i The sample is identified as clean and added to the clean set of its own model. If so, otherwise, x i Identify samples as noise and add them to the noise set of your own model. middle;

[0086] Step 4.2: For the other model, follow the same procedure as in Step 4.1 for x. i The model is then judged to obtain its clean set and noise set.

[0087] If the two models are related to x i Different judgment results will result in x i Additionally, it was added to the difficult subset. middle.

[0088] Step 5: Train any model based on the semi-supervised learning paradigm:

[0089] Step 5.1, when x i ∈ Then, the noise robustness loss function of any model can be constructed using equation (3). :

[0090] (3)

[0091] In equation (3), For the difficult collection The number of samples in the dataset, q is a hyperparameter, and q∈(0,1]; in this embodiment, q=0.7. When using L'Hôpital's rule, if q If q=0, the GCE loss function is equivalent to the CE loss function; if q=1, the GCE loss function is equivalent to the MAE loss function.

[0092] Step 5.2: Apply strong data augmentation and weak data augmentation to x respectively. i After processing, a strongly enhanced view is obtained. and weak enhancement view In this embodiment, strong data augmentation uses CIFAR10-Policy in Auto-augment, while weak data augmentation uses random pruning and random reversal.

[0093] Will The data are input into two separate models for processing, and the predicted probability distribution of the first network for the strongly augmented view is output separately. The prediction probability distribution of the second network for the strongly augmented view ;

[0094] Will The data are input into two separate models for processing, and the second network outputs the prediction probability distribution of the weakly augmented view. The prediction probability distribution of the second network for the weakly augmented view ;

[0095] Construct the dual consistency loss function of the bi-model using equation (4). :

[0096] (4)

[0097] In equation (4), KL represents the Kullback-Leibler divergence; N represents the number of samples in the dataset; in this embodiment, N is 50000.

[0098] Step 5.3, when x i ∈ Using samples x from the clean set of any model i Weakly enhanced view and enhanced view Input into another model for processing, thereby utilizing equation (5) to process the noise label y i After weighting, soft labels are obtained. :

[0099] (5)

[0100] In equation (5), Let represent the mean of the predicted probability distributions of the strong and weak augmented views output by the m-th model, and m∈{1,2};

[0101] Then use formula (6) for soft labels Perform a sharpening operation to obtain x i New soft label :

[0102] (6)

[0103] In equation (6), T is the temperature coefficient to be adjusted; for The value of the c-th position in the middle.

[0104] Step 5.4, when x i ∈ Using samples x from the noise set of any model i Weakly enhanced view and enhanced view The input is processed in another model, and x is calculated using equation (7). i pseudo-tags :

[0105] (7)

[0106] Step 5.5, x i New soft label and x i pseudo-tags Let any one of the labels be denoted as x. i New label , will x j New soft label and x j New pseudo-labels Let any one of the labels be denoted as x. j New label Thus, equations (8) and (9) are used for any pair of samples (x) i x j The new tag corresponding to ) , Perform a MixUp operation to obtain a mixed sample. Mixed tags ;

[0107] (8)

[0108] (9)

[0109] In equations (8) and (9), Indicates mixed weights, and ,in, Represents the weights, and follows the parameter . The Beta distribution.

[0110] Step 5.6, if x i ∈ Then the mixed samples and its mixed labels Add to the clean set In the middle, if x i ∈ Then the mixed samples and its mixed labels Add mixed noise set In this way, a clean mixture set of any model can be constructed. and mixed noise set .

[0111] Step 5.7: For the other model, process it in the same way as steps 5.3-5.6 to construct the mixed clean set and mixed noise set of the other model;

[0112] Step 5.8, when x i ∈ Then, the cross-entropy loss function of any model can be constructed using equation (10). :

[0113] (10)

[0114] In equation (10), for The value of the c-th position in the middle. for The value of the c-th position.

[0115] When x i ∈ Then, the mean squared error loss function of any model can be constructed using equation (11). :

[0116] (11)

[0117] Construct a semi-supervised loss using equation (12) :

[0118] (12)

[0119] In equation (12), for The weighting coefficients; in this embodiment, =(EE w ) / 16, where E is the current round number. And when (EE) w ) / 16 > 1 hour Take 1.

[0120] Step 5.9, when x i ∈ Using any sample x from the noise set of any model i Weakly enhanced view and enhanced view The weakly augmented view features are processed by the feature extractor e and mapping head h of another model. and enhanced view features Similarly, for any sample x j Weakly enhanced view features can also be obtained. and enhanced view features Therefore, the contrastive loss of any model can be constructed using equations (13), (14), and (15). :

[0121] (13)

[0122] (14)

[0123] (15)

[0124] In equations (13), (14), and (15), for and The cosine similarity between them, where t is the temperature coefficient. It is an indicator function, when hour, =1, otherwise, =0, in this embodiment, t is 0.05.

[0125] Step 5.10: Construct the total loss function for any model using equation (16). ;

[0126] (16)

[0127] In equation (16), , and They are , and The weighting coefficients; in this embodiment, for the CIFAR10 dataset, , and The values ​​are 0.025, 0.01, and 0.025, respectively. For the CIFAR100 dataset, , and The values ​​are 0.025, 0.025, and 0.05, respectively.

[0128] Step 5.11: Optimize each noise-resistant label image recognition model using the gradient descent method, so that... By minimizing the input image, two optimal image recognition networks are obtained and constitute the final dual-network image recognition model, which is used to classify the input image.

[0129] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0130] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. A noise-resistant label image recognition method based on dual-network cooperation, characterized in that, Includes the following steps: Step 1: Obtain the dataset containing noisy labels, and denote any sample-label pair in the dataset as (x... i ,y i ), where x i Let y represent the i-th sample. i x represents i Corresponding noise label; y i ∈{1,2,…,C}, where C represents the total number of categories; Step 2: Construct two networks with the same structure to form a dual network. Each network consists of a feature extractor e, a classification head f, and a mapping head h. Train the two networks separately using a dataset with noisy labels to obtain two preheated models with different parameters. Step 3: Calculate the model assignment for the i-th sample x. i The confidence level is calculated and the adaptive threshold τ for the c-th class is obtained. c ; Step 3.1: Use equation (1) to calculate the assignment of any model to x. i confidence level s i : (1) In equation (1), Indicates y i The one-hot vector, Describes any one of the models in the bimodel for x. i The predicted probability distribution; Step 3.2: Based on the category corresponding to the noise label of each sample, cluster the confidence scores of samples corresponding to the same category into one class, thereby obtaining the confidence score set under any model-assigned C categories. ,in, Let represent the set of confidence scores for the c-th class under any given model assignment; Step 3.3: Calculate the adaptive threshold τ assigned to the c-th class by any model using equation (2). c : (2) In equation (2), and The confidence sets for the c-th category are respectively. The mean and maximum value; Here, µ is the temperature coefficient, and µ is the hyperparameter. Step 4: Filter samples and construct multiple subsets, including the noise set of each model itself. and Clean Set and separate difficult subsets ; Step 4.1: For any model, if s i ∈ >τ c Then x i The sample is identified as clean and added to the clean set of its own model. If so, otherwise, x i Identify samples as noise and add them to the noise set of your own model. middle; Step 4.2: For the other model, follow the same procedure as in Step 4.1 for x. i The model is then judged to obtain its clean set and noise set. If the two models are related to x i Different judgment results will result in x i Additionally, it was added to the difficult subset. middle; Step 5: Train the two models based on the semi-supervised learning paradigm to obtain the two optimal image recognition networks and form the final dual-network image recognition model for classifying input images; Step 5.1, when ∈ Then, the noise robustness loss function of any model can be constructed using equation (3). : (3) In equation (3), For the difficult collection The number of samples in the dataset, where q is a hyperparameter and q∈(0,1]; Represents any one of the model pairs in the bimodel. The Middle Sample The predicted probability distribution express Corresponding noise label One-hot vector Step 5.2: Use strong data augmentation and weak data augmentation respectively for... After processing, a strongly enhanced view is obtained. and weakly enhanced view ; Will The data are input into two separate models for processing, and the predicted probability distribution of the first network for the strongly augmented view is output separately. The prediction probability distribution of the second network for the strongly augmented view ; Will The data are input into two separate models for processing, and the second network outputs the prediction probability distribution of the weakly augmented view. The prediction probability distribution of the second network for the weakly augmented view ; Construct the dual consistency loss function of the bi-model using equation (4). : (4) In equation (4), KL represents the Kullback-Leibler divergence; N represents the number of samples in the dataset; Step 5.3, when ∈ Using the clean set of any model, the first Sample Weakly enhanced view and enhanced view Input it into another model for processing, thereby utilizing equation (5) to... Noise tags After weighting, soft labels are obtained. : (5) In equation (5), Let represent the mean of the predicted probability distributions of the strong and weak augmented views output by the m-th model, and m∈{1,2}; Assign any model Confidence level; Then use formula (6) for soft labels Perform a sharpening operation to obtain New soft label : (6) In equation (6), T is the temperature coefficient to be adjusted; for The value of the c-th position; Step 5.4, when ∈ Using samples from the noise set of any model Weakly enhanced view and enhanced view The input is processed in another model, and then calculated using equation (7). pseudo-tags : (7) Step 5.5, New soft label and pseudo-tags Any tag in the middle is denoted as New label ,Will New soft label and New pseudo-labels Any tag in the middle is denoted as New label Thus, equations (8) and (9) are used for any pair of samples ( , The new tag corresponding to ) , Perform a MixUp operation to obtain a mixed sample. Mixed tags ; (8) (9) In equations (8) and (9), Indicates mixed weights, and ,in, Represents the weights, and follows the parameter . The Beta distribution; Step 5.6, if ∈ Then the mixed samples and its mixed labels Add to the clean set In the middle, if ∈ Then the mixed samples and its mixed labels Add mixed noise set In this way, a clean mixture set of any model can be constructed. and mixed noise set ; Step 5.7: For the other model, process it in the same way as steps 5.3-5.6 to construct the mixed clean set and mixed noise set of the other model; Step 5.8, when the k-th sample ∈ Then, the cross-entropy loss function of any model can be constructed using equation (10). : (10) In equation (10), For the k-th sample Corresponding noise label The value of the c-th position in the middle. for The value of the c-th position; When the v-th sample ∈ Then, the mean squared error loss function of any model can be constructed using equation (11). : (11) Construct a semi-supervised loss using equation (12) : (12) In equation (12), for Weighting coefficients; Step 5.9, when ∈ Using any sample from the noise set of any model Weakly enhanced view and enhanced view The input features are processed by the feature extractor e and the mapping head h of another model to obtain... Weak augmented view features and enhanced view features ; when ∈ Using any other sample from the noise set of any model Weakly enhanced view and enhanced view The input features are processed by the feature extractor e and the mapping head h of another model to obtain... Weak augmented view features and enhanced view features Thus, the contrastive loss of any model can be constructed using equations (13), (14), and (15). : (13) (14) (15) In equations (13), (14), and (15), for and The cosine similarity between them, where t is the temperature coefficient; It is an indicator function, when season =1, otherwise, let =0; It is a noise collection Medium sample size; Step 5.10: Construct the total loss function for any model using equation (16). ; (16) In equation (16), , and They are , and Weighting coefficients; Step 5.11: Optimize each noise-resistant label image recognition model using the gradient descent method, so that... By minimizing the minimum, two optimal image recognition networks are obtained.

2. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the noise-resistant tag image recognition method of claim 1, and the processor is configured to execute the program stored in the memory.

3. A computer-readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to perform the steps of the noise-resistant tag image recognition method of claim 1.

Images