Anti-noise label image identification method based on dual-network cooperation
A noise-resistant label image recognition method based on dual-network collaboration is proposed. This method utilizes confidence to filter samples and construct multiple subsets, and combines noise robustness loss and dual consistency loss to optimize the model and address the robustness and generalization performance issues caused by noise labels. This results in high accuracy and robustness in noisy scenarios.
Patent Information
- Application Number
- CN202411126873.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-16
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-08-16
AI Technical Summary
Existing deep learning models suffer from compromised robustness and generalization performance when faced with noisy labeled datasets, especially under high noise rate scenarios. Existing methods perform well under low noise rates but poorly under extreme noise rates.
A noise-resistant label image recognition method based on dual-network collaboration is adopted. By constructing two networks with the same structure, samples are filtered using confidence and adaptive thresholds to build clean sets, noisy sets and hard sets. Noise robust loss, dual consistency loss and semi-supervised learning are introduced to optimize the model to improve robustness and classification accuracy.
By selecting samples and collaborating with networks, the fitting of noisy labels is effectively reduced, improving the robustness and classification accuracy of the model and enhancing its generalization ability in noisy scenarios.
Smart Images

Figure CN119027733A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer vision technology, and in particular to image recognition technology under noise labels. Background Art
[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 depends largely on large-scale, high-quality annotated datasets, which require a lot of manpower, resources, and time to obtain. In contrast, the use of some cheap alternatives (such as crowdsourcing or web crawling) can significantly reduce the cost of obtaining training data and promote the maturity and low-cost implementation of image recognition technology. However, the use of these data will also bring other problems: most of these training data come from crowdsourcing or are directly crawled from the Internet using keywords, so the quality of their labels cannot be guaranteed. Compared with the image annotation performed by professionals, the labels corresponding to these data are likely to be wrong, that is, noisy labels. Previous studies have shown that deep neural networks can easily fit these noisy labels, resulting in impaired model robustness and reduced generalization performance. This problem becomes particularly serious in practical applications, such as medical image analysis, autonomous driving, video surveillance, and other fields, where the datasets usually contain a high proportion of label noise. Therefore, noisy label learning has received more and more attention from researchers.
[0003] In order to combat the negative impact of noisy labels, existing methods are mainly divided into the following three types: 1. Methods based on sample selection: The main idea is to select clean samples from the noisy data set to obtain clean sets and noisy sets, and use the idea of semi-supervised learning for subsequent training. 2. Methods based on label correction: The main idea is to correct the noisy labels in the noisy data set to obtain truly clean labels for training models. 3. Regularization-based methods: The main idea is to mitigate the side effects of label noise by preventing deep neural networks from overfitting all training samples, such as designing noise-robust loss functions, regularization methods, etc. However, these methods often achieve better results in low-noise scenarios, but perform poorly in extreme noise rates. Therefore, there is still a lot of room for improvement in image recognition tasks in label noise scenarios. Summary of the invention
[0004] In order to address the shortcomings of the above-mentioned prior art, the present invention proposes a noise-resistant label image recognition method based on dual network collaboration, so as to effectively combat the influence of noise labels and reduce the model's fitting of noise labels under the condition of training with a data set 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] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:
[0006] The invention discloses a method for anti-noise label image recognition based on dual network collaboration, which comprises the following steps:
[0007] Step 1: Get a dataset containing noisy labels and record any sample-label pair in the dataset as (x i ,y i ), where x i represents the i-th sample, y i Represents x i The corresponding noise label; y i ∈{1,2,…,C}, 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, and use a data set containing noise labels to train the two networks in the dual network separately to obtain two models with different parameters after preheating;
[0009] Step 3: Calculate the i-th sample x assigned to any model i The confidence of the c-th category and obtain the adaptive threshold τ c ;
[0010] Step 4: Filter samples and construct multiple subsets, including each model’s own noise set and clean set and a separate subset of difficulties ;
[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 anti-noise label image recognition method based on dual network collaboration described in the present invention is also characterized in that the step 3 includes the following steps:
[0013] Step 3.1: Use formula (1) to calculate the value assigned to x by any model i The confidence level s i :
[0014] (1)
[0015] In formula (1), Represents y i The one-hot vector of Represents any model pair x in the dual model i The predicted probability distribution of
[0016] Step 3.2: According to the category corresponding to the noise label of each sample, the confidence of samples corresponding to the same category is clustered into one category, so as to obtain the confidence set under C categories assigned by any model. ,in, represents the confidence set of the cth category under any model assignment;
[0017] Step 3.3: Use formula (2) to calculate the adaptive threshold τ assigned to the cth category by any model: c :
[0018] (2)
[0019] In formula (2), and are the confidence sets of the cth category The mean and maximum value of ; is the temperature coefficient and µ is a hyperparameter.
[0020] The step 4 comprises the following steps:
[0021] Step 4.1: For any model, if s i ∈ >τ c , then x i Identify as a clean sample and add it to the clean set of its own model Otherwise, x i Identify as noise samples and add them to the noise set of the model middle;
[0022] Step 4.2: For another model, follow the process of step 4.1 to calculate x i Perform discrimination to obtain the clean set and noise set of its own model;
[0023] If the two models are i The judgment results are different, and x i Add to the difficult subset middle;
[0024] The step 5 comprises the following steps:
[0025] Step 5.1: When x i ∈ , then use formula (3) to construct the noise robust loss function of any model :
[0026] (3)
[0027] In formula (3), For difficult set The number of samples in, q is a hyperparameter, and q∈(0,1];
[0028] Step 5.2: Use strong data enhancement and weak data enhancement for x i Processing is performed to obtain a strongly enhanced view and weakly enhanced views ;
[0029] Will The two models are input for processing and the predicted probability distribution of the first network for the strongly enhanced view is output respectively. And the predicted probability distribution of the second network for the strongly enhanced view ;
[0030] Will The two models are input for processing and the predicted probability distribution of the second network for the weakly enhanced view is output respectively. And the predicted probability distribution of the second network for the weakly enhanced view ;
[0031] Use formula (4) to construct the dual consistency loss function of the dual model :
[0032] (4)
[0033] In formula (4), KL represents the Kullback-Leibler divergence; N represents the number of samples in the data set;
[0034] Step 5.3: When x i ∈ , using samples x from the clean set of any model i Weakly enhanced view of and strong enhanced view Input into another model for processing, so as to use formula (5) to i After weighting, we get the soft label :
[0035] (5)
[0036] In formula (5), represents the mean of the predicted probability distribution of the strong and weak enhanced views output by the mth model, and ; m∈{1,2};
[0037] Then use formula (6) to calculate the soft tag Perform sharpening operation to obtain x i New soft label :
[0038] (6)
[0039] In formula (6), T is the temperature coefficient to be adjusted; for The value of the cth position in ;
[0040] Step 5.4: When x i ∈ , using samples x from the noise set of any model i Weakly enhanced view of and strong enhanced view Input into another model for processing, so as to calculate x using formula (7) i Pseudo-labels :
[0041] (7)
[0042] Step 5.5: x i New soft label and x i Pseudo-labels Any label in is denoted as x i New Tags , x j New soft label and x j New pseudo-labels for Any label in is denoted as x j New Tags , so that we can use equations (8) and (9) to calculate any pair of samples (x i , x j ) corresponding to the new label ( , ) Perform MixUp operation to obtain mixed samples Mixed Tags ;
[0043] (8)
[0044] (9)
[0045] In formula (8) and formula (9), represents the mixing weight, and ,in, represents the weight, and the obedience parameter is Beta distribution of
[0046] Step 5.6: If x i ∈ , then the mixed sample and mixed tags Add to Mix Clean Set If x i ∈ , then the mixed sample and mixed tags Add to Mix Noise Set In order to construct a mixed clean set of any model and mixed noise set ;
[0047] Step 5.7: For another model, the process from step 5.3 to step 5.6 is also performed to construct a mixed clean set and a mixed noise set of another model;
[0048] Step 5.8: When x i ∈ , then use formula (10) to construct the cross entropy loss function of any model :
[0049] (10)
[0050] In formula (10), for The value of the cth position in for The value of the cth position in ;
[0051] When x i ∈ , then use formula (11) to construct the mean square error loss function of any model :
[0052] (11)
[0053] Use formula (12) to construct the semi-supervised loss :
[0054] (12)
[0055] In formula (12), for The weight coefficient of
[0056] Step 5.9: When x i ∈ , use any sample x in the noise set of any model i Weakly enhanced view of and strong enhanced view Input the feature extractor e and mapping head h of another model for processing to obtain x iWeakly enhanced view features and strong enhanced view features ;
[0057] When x j ∈ , use any other sample x in the noise set of any model j Weakly enhanced view of and strong enhanced view Input the feature extractor e and mapping head h of another model for processing to obtain x j Weakly enhanced view features and strong enhanced view features , and thus use equations (13), (14) and (15) to construct the contrast loss of any model :
[0058] (13)
[0059] (14)
[0060] (15)
[0061] In formula (13), formula (14) and formula (15), for and The cosine similarity between , t is the temperature coefficient; is the indicator function, when season =1, otherwise, let =0;
[0062] Step 5.10: Use formula (16) to construct the total loss function of any model ;
[0063] (16)
[0064] In formula (16), , and They are , and The weight coefficient of
[0065] Step 5.11: Use the gradient descent method to optimize and solve each noise-resistant label image recognition model. Reach the minimum, thus obtaining two optimal image recognition networks.
[0066] The electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the anti-noise label image recognition method, and the processor is configured to execute the program stored in the memory.
[0067] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the anti-noise label image recognition method when the computer program is executed by a processor.
[0068] Compared with the prior art, the present invention has the following beneficial effects:
[0069] 1. This invention introduces a class-adaptive threshold for sample selection. This design takes advantage of the established consensus that samples show different fitting difficulties in different categories. By analyzing the probability distribution of all samples and their corresponding categories, a class-adaptive threshold is determined to select reliable samples. This design helps to separate clean and noisy samples, thereby reducing the impact of confirmation bias and improving the recognition accuracy of the model for image classification.
[0070] 2. In addition to the clean set and noisy set used in previous methods, the present invention also constructs a difficult set, which contains those samples that the two networks discriminate inconsistently, that is, samples that one network considers clean and the other network considers noisy. These samples are close to the decision boundary and have the ability to improve the generalization ability of DNN. In addition, the noise robust loss introduced on these difficult samples enables the model to carefully use possible clean labels instead of replacing them with pseudo labels, avoiding information loss and confirmation bias that may be caused by pseudo labels, thereby further improving the robustness of the model and the accuracy of classification.
[0071] 3. The present invention proposes a dual consistency loss, which constrains the consistency of the predictions of the two networks for different enhanced views of the same sample. This allows the two networks to learn from each other, so that they can share and utilize training experience from different sources during the training process. That is, through mutual learning, the errors caused by biased selection can be mutually reduced by the peer networks, thereby achieving better generalization and significantly improving the model's image recognition ability. BRIEF DESCRIPTION OF THE DRAWINGS
[0072] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION
[0073] In this embodiment, Figure 1 As shown, a noise-resistant label image recognition method based on dual network collaboration includes the following steps:
[0074] Step 1: Get a dataset containing noisy labels and record any sample-label pair in the dataset as (x i ,y i ), where x i represents the i-th sample, y i Represents x i The corresponding noise label; y i ∈{1,2,…,C}, C represents the total number of categories; in this embodiment, the selected dataset containing noise labels is obtained by artificially injecting noise labels into the clean dataset CIFAR-10 / 100 dataset, where C in the CIFAR10 dataset is 10 and C in the CIFAR100 dataset is 100. Specifically, the sample labels of each category are randomly shuffled into the labels of other categories according to the noise rate, so as to obtain symmetric noise.
[0075] Step 2: construct two networks with the same structure to form a dual network, each network is composed of a feature extractor e, a classification head f and a mapping head h, and use a data set containing noise labels to train the two networks in the dual network respectively to obtain two models with different parameters after preheating; in this embodiment, the noisy CIFAR data set is used according to the general paradigm of image classification, that is, the cross entropy loss function is used to train the model E w 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 of each sample and obtain the category adaptive threshold:
[0077] Step 3.1: Use formula (1) to calculate the value assigned to x by any model i The confidence level s i :
[0078] (1)
[0079] In formula (1), Represents y i The one-hot vector of Represents any model pair x in the dual model i The predicted probability distribution of
[0080] Step 3.2: According to the category corresponding to the noise label of each sample, the confidence of samples corresponding to the same category is clustered into one category, so as to obtain the confidence set under C categories assigned by any model. ,in, represents the confidence set of the cth category under any model assignment;
[0081] Step 3.3: Use formula (2) to calculate the adaptive threshold τ assigned to the cth category by any model: c :
[0082] (2)
[0083] In formula (2), and are the confidence sets of the cth category The mean and maximum value of ; is the temperature coefficient, µ is a hyperparameter; in this embodiment, is 2, is 0.3.
[0084] Step 4: Filter samples and construct multiple subsets:
[0085] Step 4.1: For any model, if s i ∈ >τ c , then x i Identify as a clean sample and add it to the clean set of its own model Otherwise, x i Identify as noise samples and add them to the noise set of the model middle;
[0086] Step 4.2: For another model, follow the process of step 4.1 to calculate x i Perform discrimination to obtain the clean set and noise set of its own model;
[0087] If the two models are i The judgment results are different, and x i Add 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 use formula (3) to construct the noise robust loss function of any model :
[0090] (3)
[0091] In formula (3), For difficult set The number of samples in , 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: Use strong data enhancement and weak data enhancement for x i Processing is performed to obtain a strongly enhanced view and weakly enhanced views In this embodiment, strong data enhancement uses CIFAR10-Policy in Auto-augment, and weak data enhancement uses random cropping and random inversion.
[0093] Will The two models are input for processing and the predicted probability distribution of the first network for the strongly enhanced view is output respectively. And the predicted probability distribution of the second network for the strongly enhanced view ;
[0094] Will The two models are input for processing and the predicted probability distribution of the second network for the weakly enhanced view is output respectively. And the predicted probability distribution of the second network for the weakly enhanced view ;
[0095] Use formula (4) to construct the dual consistency loss function of the dual model :
[0096] (4)
[0097] In formula (4), KL represents the Kullback-Leibler divergence; N represents the number of samples in the data set; in this embodiment, N is 50,000.
[0098] Step 5.3: When x i ∈ , using samples x from the clean set of any model i Weakly enhanced view of and strong enhanced view Input into another model for processing, so as to use formula (5) to i After weighting, we get the soft label :
[0099] (5)
[0100] In formula (5), represents the mean of the predicted probability distribution of the strong and weak enhanced views output by the mth model, and ; m∈{1,2};
[0101] Then use formula (6) to calculate the soft tag Perform sharpening operation to obtain x i New soft label :
[0102] (6)
[0103] In formula (6), T is the temperature coefficient to be adjusted; for The value of the cth position in .
[0104] Step 5.4: When x i ∈ , using samples x from the noise set of any model i Weakly enhanced view of and strong enhanced view Input into another model for processing, so as to calculate x using formula (7) i Pseudo labels :
[0105] (7)
[0106] Step 5.5: x i New soft label and x i Pseudo labels Any label in is denoted as x i New Tags , x j New soft label and x j New pseudo-labels for Any label in is denoted as x j New Tags , so that we can use equations (8) and (9) to calculate any pair of samples (x i , x j ) corresponding to the new label ( , ) Perform MixUp operation to obtain mixed samples Mixed Tags ;
[0107] (8)
[0108] (9)
[0109] In formula (8) and formula (9), represents the mixing weight, and ,in, represents the weight, and the obedience parameter is Beta distribution.
[0110] Step 5.6: If x i ∈ , then the mixed sample and mixed tags Add to Mix Clean Set If x i ∈ , then the mixed sample and mixed tags Add to Mix Noise Set In order to construct a mixed clean set of any model and mixed noise set .
[0111] Step 5.7: For another model, the process from step 5.3 to step 5.6 is also performed to construct a mixed clean set and a mixed noise set of another model;
[0112] Step 5.8: When x i ∈ , then use formula (10) to construct the cross entropy loss function of any model :
[0113] (10)
[0114] In formula (10), for The value of the cth position in for The value of the cth position in .
[0115] When x i ∈ , then use formula (11) to construct the mean square error loss function of any model :
[0116] (11)
[0117] Use formula (12) to construct the semi-supervised loss :
[0118] (12)
[0119] In formula (12), for The weight coefficient of =(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 ∈ , use any sample x in the noise set of any model i Weakly enhanced view of and strong enhanced view Input into the feature extractor e and mapping head h of another model for processing to obtain weakly enhanced view features and strong enhanced view features Similarly, for any sample x j Weakly enhanced view features can also be obtained and strong enhanced view features Thus, we can use equations (13), (14) and (15) to construct the contrast loss of any model. :
[0121] (13)
[0122] (14)
[0123] (15)
[0124] In formula (13), formula (14) and formula (15), for and The cosine similarity between , t is the temperature coefficient. is the indicator function, when hour, =1, otherwise, =0, in this embodiment, t is 0.05.
[0125] Step 5.10: Use formula (16) to construct the total loss function of any model ;
[0126] (16)
[0127] In formula (16), , and They are , and The weight coefficient of ; In this embodiment, for the CIFAR10 dataset, , and are 0.025, 0.01, and 0.025 respectively. For the CIFAR100 dataset, , and They are 0.025, 0.025 and 0.05 respectively.
[0128] Step 5.11: Use the gradient descent method to optimize and solve each noise-resistant label image recognition model. To achieve the minimum, we obtain two optimal image recognition networks and form the final dual-network image recognition model, which is used to classify the input images.
[0129] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above 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 on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.
Claims
1. A noise-resistant label image recognition method based on dual network collaboration, characterized in that: The following steps are involved: Step 1: Get a dataset containing noisy labels and record any sample-label pair in the dataset as (x i ,y i ), where x i represents the i-th sample, y i Represents x i The corresponding noise label; y i ∈{1,2,…,C}, 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, and use a data set containing noise labels to train the two networks in the dual network separately to obtain two models with different parameters after preheating; Step 3: Calculate the i-th sample x assigned to any model i The confidence of the c-th category and obtain the adaptive threshold τ c ; Step 4: Filter samples and construct multiple subsets, including each model’s own noise set and clean set and a separate subset of difficulties ; 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.
2. According to claim 1, a method for anti-noise label image recognition based on dual network collaboration is characterized in that: The step 3 comprises the following steps: Step 3.1: Use formula (1) to calculate the value assigned to x by any model i The confidence level s i : (1) In formula (1), Represents y i The one-hot vector of Represents any model pair x in the dual model i The predicted probability distribution of Step 3.2: According to the category corresponding to the noise label of each sample, the confidences of samples corresponding to the same category are clustered into one category, thereby obtaining a confidence set under C categories assigned by any model. ,in, represents the confidence set of the cth category under any model assignment; Step 3.3: Use formula (2) to calculate the adaptive threshold τ assigned to the cth category by any model: c : (2) In formula (2), and are the confidence sets of the cth category The mean and maximum value of ; is the temperature coefficient and µ is a hyperparameter.
3. The method for anti-noise label image recognition based on dual network collaboration according to claim 2 is characterized in that: The step 4 comprises the following steps: Step 4.1: For any model, if s i ∈ >τ c , then x i Identify as a clean sample and add it to the clean set of its own model Otherwise, x i Identify as noise samples and add them to the noise set of the model middle; Step 4.2: For another model, follow the process of step 4.1 to calculate x i Perform discrimination to obtain the clean set and noise set of its own model; If the two models are i The judgment results are different, and x i Add to the difficult subset middle.
4. The method for anti-noise label image recognition based on dual network collaboration according to claim 3 is characterized in that: The step 5 comprises the following steps: Step 5.1: When x i ∈ , then use formula (3) to construct the noise robust loss function of any model : (3) In formula (3), For difficult set The number of samples in, q is a hyperparameter, and q∈(0,1]; Step 5.2: Use strong data enhancement and weak data enhancement for x i Processing is performed to obtain a strongly enhanced view and weakly enhanced views ; Will The two models are input for processing and the predicted probability distribution of the first network for the strongly enhanced view is output respectively. And the predicted probability distribution of the second network for the strongly enhanced view ; Will The two models are input for processing and the predicted probability distribution of the second network for the weakly enhanced view is output respectively. And the predicted probability distribution of the second network for the weakly enhanced view ; Use formula (4) to construct the dual consistency loss function of the dual model : (4) In formula (4), KL represents the Kullback-Leibler divergence; N represents the number of samples in the data set; Step 5.3: When x i ∈ , using samples x from the clean set of any model i Weakly enhanced view of and strong enhanced view Input into another model for processing, so as to use formula (5) to i After weighting, we get the soft label : (5) In formula (5), represents the mean of the predicted probability distribution of the strong and weak enhanced views output by the mth model, and ; m∈{1,2}; Then use formula (6) to calculate the soft tag Perform sharpening operation to obtain x i New soft label : (6) In formula (6), T is the temperature coefficient to be adjusted; for The value of the cth position in ; Step 5.4: When x i ∈ , using samples x from the noise set of any model i Weakly enhanced view of and strong enhanced view Input into another model for processing, so as to calculate x using formula (7) i Pseudo-labels : (7) Step 5.5: x i New soft label and x i Pseudo labels Any label in is denoted as x i New Tags , x j New soft label and x j New pseudo-labels for Any label in is denoted as x j New Tags , so that we can use equations (8) and (9) to calculate any pair of samples (x i , x j ) corresponding to the new label ( , ) Perform MixUp operation to obtain mixed samples Mixed Tags ; (8) (9) In formula (8) and formula (9), represents the mixing weight, and ,in, represents the weight, and the obedience parameter is Beta distribution of Step 5.6: If x i ∈ , then the mixed sample and mixed tags Add to Mix Clean Set If x i ∈ , then the mixed sample and mixed tags Add to Mix Noise Set In order to construct a mixed clean set of any model and mixed noise set ; Step 5.7: For another model, the process from step 5.3 to step 5.6 is also performed to construct a mixed clean set and a mixed noise set of another model; Step 5.8: When x i ∈ , then use formula (10) to construct the cross entropy loss function of any model : (10) In formula (10), for The value of the cth position in for The value of the cth position in ; When x i ∈ , then use formula (11) to construct the mean square error loss function of any model : (11) Use formula (12) to construct the semi-supervised loss : (12) In formula (12), for The weight coefficient of Step 5.9: When x i ∈ , use any sample x in the noise set of any model i Weakly enhanced view of and strong enhanced view Input the feature extractor e and mapping head h of another model for processing to obtain x i Weakly enhanced view features and strong enhanced view features ; When x j ∈ , use any other sample x in the noise set of any model j Weakly enhanced view of and strong enhanced view Input the feature extractor e and mapping head h of another model for processing to obtain x j Weakly enhanced view features and strong enhanced view features , and thus use equations (13), (14) and (15) to construct the contrast loss of any model : (13) (14) (15) In formula (13), formula (14) and formula (15), for and The cosine similarity between , t is the temperature coefficient; is the indicator function, when season =1, otherwise, let =0; Step 5.10: Use formula (16) to construct the total loss function of any model ; (16) In formula (16), , and They are , and The weight coefficient of Step 5.11: Use the gradient descent method to optimize and solve each noise-resistant label image recognition model. Reach the minimum, thus obtaining two optimal image recognition networks.
5. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the anti-noise label image recognition method according to any one of claims 1 to 4, and the processor is configured to execute the program stored in the memory.
6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the anti-noise label image recognition method according to any one of claims 1 to 4 are executed.
Citation Information
Patent Citations
Label noise-containing image classification method based on graph consistency and semi-supervised model
CN114881125A
Robust learning method based on class labels with noise and imbalance
CN115331088A
Self-tag modification method for processing noise tag
CN115861625A
Noise-containing label image identification method based on adaptive category equalization
CN116912568A
Noise-containing label image identification method and system based on category balance sample selection
CN117173494A
