A single-positive multi-label classification method based on label-aware global consistency

Abstract

The present invention discloses a single-positive multi-label classification method based on label-aware global consistency. Given a training data set with N samples, the model components are initialized; B samples are randomly selected from the data set, and all unknown labels of each sample are assumed to be negative labels. The loss function is calculated using the positive and negative pseudo-label thresholds δ and the classification prediction value p of each sample to obtain credible positive and negative pseudo-labels, and the pseudo-label consistency regularization loss is calculated to combine the single positive label and pseudo-label of each sample, and the low-dimensional feature embedding z predicted as the positive label is summarized into each cluster. The loss is calculated using the label-aware matrix with the goal of minimizing the total loss function, and the model parameters are updated. The model is iteratively optimized until the model converges. The present invention can fully explore and utilize the potential relationship between labels when only partial label information is available, greatly reducing the labeling cost of the data.

Description

Technical Field

[0001] The present invention belongs to the field of multi-label classification, and in particular relates to a single-positive multi-label classification method based on label-aware global consistency. Background Art

[0002] Machine learning is one of the key technologies driving intelligent development across various fields. Labeled data is the foundation for training most machine learning models, but obtaining this data is often time-consuming and labor-intensive, becoming a major obstacle to the widespread application of machine learning. The process of acquiring labeled data typically involves data collection, data cleaning, data labeling, and manual review, with data labeling being particularly crucial and complex. For example, in the field of medical image analysis, training a multi-label classification model requires the precise labeling of a large number of medical images, a task that typically requires significant time and effort from professional physicians.

[0003] In practical applications, data is often incomplete, typically only partial labels are available for each sample, rather than complete multi-label annotated data. This situation also arises in fields such as text classification in natural language processing and user interest tag prediction in recommendation systems. Traditional multi-label classification methods typically assume that each sample has multiple positive labels. These methods perform well when sufficient label information is available, but their performance degrades significantly when annotation information is insufficient. Research has shown that existing multi-label learning methods perform poorly when only a single positive label is available, failing to fully exploit the correlation information between labels, resulting in reduced classification performance.

[0004] To address these challenges, this paper proposes a training method based on single positively labeled data. By constructing a label-aware matrix and applying a global consistency criterion, this method effectively exploits label correlations between samples, improves classification accuracy, and reduces reliance on fully labeled data. This method has broad application prospects and practical value in fields such as medical image analysis, natural language processing, and recommender systems. Summary of the Invention

[0005] Purpose of the invention: In order to reduce the cost of data annotation and improve the utilization rate of incompletely annotated data, the present invention provides a single-positive multi-label classification method based on label-aware global consistency.

[0006] Technical solution: The present invention provides a single-positive multi-label classification method based on label-aware global consistency, which specifically includes the following steps:

[0007] (1) Given a training dataset with N samples, merge and initialize the model components;

[0008] (2) Randomly select B samples from the training data set, assume that all unknown labels of each sample are negative labels, and calculate the loss function

[0009] (3) Using the positive and negative pseudo-label thresholds δ and Divide the classification prediction value p of each sample to obtain credible positive and negative pseudo labels, and calculate the pseudo label consistency regularization loss

[0010] (4) Combine the single positive label of each sample with the pseudo label obtained in step (3), summarize the low-dimensional features z predicted as positive labels into each cluster, and use the label perception matrix to calculate the label perception consistency regularization loss

[0011] (5) With the goal of minimizing the total loss function, update the model parameters and return to step (2) to iteratively optimize the model until the model converges.

[0012] Furthermore, the implementation process of step (1) is as follows:

[0013] Training data set Among them, x n For the sample, is a single positive label; the initialization model components include: feature encoder E(·), classifier f(·) and nonlinear mapper g(·), where f(·) and g(·) process the sample features extracted by E(·) and output the classification prediction value p and low-dimensional feature embedding z; the positive and negative pseudo label thresholds δ and Used to classify credible labels for samples; the classifier f(·) uses one fully connected layer as an example, and the nonlinear mapper g(·) uses two fully connected layers and one activation layer.

[0014] Furthermore, the loss function in step (2) for:

[0015] Using BCE loss, and introducing weak enhancement technology to increase the number of visible samples of the model, while ensuring the integrity of the enhanced semantic information, for:

[0016]

[0017] in, For the classifier f(·) to predict the input Aug w (x n )’s class probability vector, Aug w (·) represents a weak enhancement method, is a single positive label; L is the total number of categories, and the subscript i is the category index, that is, express The value of the i-th category, for The value of the i-th category; the weak enhancement method includes horizontal and vertical flipping of the image, and synonym replacement of the text.

[0018] Furthermore, the implementation process of step (3) is as follows:

[0019] Introducing strong enhancement Aug s (·) Strengthen consistency, encourage the prediction results of strong and weak enhancement versions of the same sample to be consistent; By using the positive and negative pseudo-label thresholds δ and Filter out high-quality pseudo labels as the object of consistency regularization, that is, discard Partial label; Calculated by the following formula:

[0020]

[0021] in, For the classifier f(·) to predict the input Aug s (x n )’s class probability vector, Aug s (·) represents a strong enhancement method, For the classifier f(·) to predict the input Aug w (x n ), is a single positive label, L is the total number of categories, and the subscript i is the category index, that is, express The value of the i-th category; when the weak enhancement sample Aug w (x n ) Greater than or equal to the threshold δ or the original label When is positive, the sample x n The i-th class is marked as positive, corresponding to the first term in the formula; when the predicted value Less than threshold When , the sample x n The i-th class is marked as negative, corresponding to the second term in the formula; if the predicted value The missing labels are discarded; the strong enhancement method includes random cropping and flipping of the image, and random deletion of text.

[0022] Furthermore, the implementation process of step (4) is as follows:

[0023] According to the obtained pseudo-label, the low-dimensional feature predicted as the positive label is embedded into the z classification. or There is c(z ni )=i, that is, only the sample predicts the value in the i-th category Greater than or equal to the threshold δ or a single positive label When it is the i-th category, the feature of the sample in the i-th category is embedded into z ni is classified into category i; therefore, the feature embedding category function c(·) is defined as:

[0024]

[0025] For B samples, there are B'=B×L feature embeddings (each sample has L categories); z ni Select and constitute feature embedding as The operator is that the low-dimensional feature embedding of each sample is split into L parts according to the category dimension and the part with positive pseudo labels is filtered out to be used as the operator for subsequent calculations; the filtered feature embedding constitutes a new set, defined as

[0026] The label perception matrix I is defined as follows:

[0027]

[0028] final, Calculated by the following formula:

[0029]

[0030] Among them, exp(z j ·z j' / τ) is used to aggregate feature embeddings of the same category. The role of is to distance the feature embedding of different categories, and τ is the temperature coefficient.

[0031] Furthermore, the total loss function in step (5) is:

[0032]

[0033] Among them, λ PLC and λ LAC is a weight hyperparameter used to adjust the weight of each loss. is the pseudo-label consistency regularization loss, is the label-aware consistency regularization loss; the λ PLC and λ LAC The value is 1.

[0034] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: By constructing a label perception matrix and applying a global consistency criterion, the present invention can make full use of the label correlation between samples, thereby significantly improving the accuracy of multi-label classification; traditional multi-label classification methods usually require each sample to have multiple positive labels, while the method of the present invention can maintain high classification performance when each sample has only a single positive label, reducing the dependence on comprehensive labeled data; the present invention performs well in processing incomplete labeled data, can effectively use limited labeled information for training, and is suitable for practical application scenarios with high labeling costs and incomplete labeling information; the present invention has great applications in medical image analysis, natural language processing and recommendation The present invention has shown significant advantages in fields such as systems, and can be widely used in practical applications that require multi-label classification, and has important practical value; the present invention adopts an iterative optimization algorithm to maximize the global consistency criterion by calculating the gradient of the optimization objective function and updating the model parameters. The optimization process is simple and efficient, and is easy to implement and apply; the present invention can adapt to the addition of new data by dynamically updating the label perception matrix, thereby improving the generalization ability and adaptability of the model, thereby maintaining high performance in a constantly changing data environment; the present invention effectively utilizes data with a single positive label, and the method of the present invention significantly reduces the cost and time of data annotation, improves the efficiency of model training, and makes it more suitable for resource-constrained application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Flowchart of the present invention. DETAILED DESCRIPTION

[0036] The present invention will be further described in detail below with reference to the accompanying drawings.

[0037] like Figure 1 As shown in FIG, the present invention proposes a single-positive multi-label classification method based on label-aware global consistency, which calculates two consistency regularization losses to improve the accuracy of multi-label classification. Specifically, the method includes the following steps:

[0038] Step 1: Given a training data set with N samples Among them, x n For the sample, is a single positive label; initialize the model components: feature encoder E(·), classifier f(·) and nonlinear mapper g(·), where f(·) and g(·) process the sample features extracted by E(·) and output the classification prediction value p and low-dimensional feature embedding z; preset the positive and negative pseudo label thresholds δ and Confident labels for classifying samples.

[0039] In this example, the medical image classification task is used as an example. Initially, 10,000 images are given. These images can be X-rays, CT scans, MRI scans, ultrasound images, etc. They usually contain important information about the patient's health status. Each image can correspond to multiple disease types, i.e. The total number of categories L = 80, including labels such as "fracture" and "arthritis"; each image is labeled with a single positive label to form a training data set In this embodiment, E(·) uses the Q2L model as an example, f(·) uses one fully connected layer as an example, and g(·) uses two fully connected layers and one activation layer as an example; among them, f(·) and g(·) process the image features extracted by E(·) and output the classification prediction value p and the low-dimensional feature embedding z.

[0040] Step 2: Randomly select B samples from the training data set D, assume that all unknown labels of each sample are negative labels, and calculate the loss function

[0041] Each time, a batch of data is randomly taken out from the set D as training data. If the loss is calculated only for the given positive labels, although the introduction of wrong labels can be effectively avoided, it will cause the model to collapse to a trivial solution, that is, the model gradually loses the ability to distinguish different categories during the training process, and may eventually only output a single prediction result. In medical image classification, it manifests as the model always predicting all images as a specific disease, ignoring other possible pathological states. In order to avoid the occurrence of this phenomenon, all unknown labels can be assumed to be negative labels. The reason for this assumption is that in multi-label data annotation, the proportion of negative labels is much higher than that of positive labels, so it can ensure that the vast majority of real negative labels are hit. The classic Binary Cross Entropy (BCE) loss is used, and a weak enhancement technique is introduced to increase the number of visible samples of the model, while ensuring the integrity of the enhanced semantic information, to assist in improving the generalization performance of the model. This loss is defined as "Assume Negative" (AN), that is It can be calculated by the following formula:

[0042]

[0043] Where L is the total number of categories, is an indicator function, which takes 1 when the condition in the brackets is met, otherwise it takes 0. For the classifier f(·) to predict the input Aug w (x n )’s class probability vector, Aug w(·) indicates weak enhancement methods, such as horizontal flipping of images, synonym replacement of text, etc. In images, it can be common transformation methods such as horizontal or vertical flipping.

[0044] Step 3: Use the positive and negative pseudo-label thresholds δ and Divide the classification prediction value p of each sample to obtain credible positive and negative pseudo labels and calculate the loss function

[0045] Inevitably, the more intuitive method of step 2 introduces erroneous supervision information, misclassifying some true positive labels as negative labels, resulting in a decrease in generalization. In medical image classification, the existence of wrong labels is inevitable. If the processing of wrong labels is neglected during training, the model may misdiagnose certain diseases in practical applications. This is especially dangerous in the medical field, because misdiagnosis may have a serious impact on the patient's treatment and health. Therefore, additional regularization terms need to be designed to correct this error. Define "Pseudo-Label Consistency" (PLC) between pseudo labels and model predictions. In addition to weakly enhancing Augment w (·) In addition, we also introduced the strong enhancement Aug s (·) to strengthen consistency, that is, to encourage the prediction results of the strong and weak enhanced versions of the same sample to be consistent. By using the positive and negative pseudo-label thresholds δ and Filter out high-quality pseudo labels as the object of consistency regularization, that is, discard section tags. It can be calculated by the following formula:

[0046]

[0047] in, For the classifier f(·) to predict the input Aug s (x n )’s class probability vector, Aug s (·) represents a strong enhancement method, such as random cropping of images, random deletion of text, etc. In image classification tasks, strong enhancement includes random horizontal flipping, cropping, and other methods. δ and Represents the threshold, which is used to divide the pseudo positive label and the pseudo negative label. Here, when the weakly enhanced sample Aug w (x n ) Greater than or equal to the threshold δ or the original label When is positive, the sample x n The i-th class is marked as positive, corresponding to the first term in the formula; when the predicted value Less than threshold When , the sample x n The i-th class is marked as negative, corresponding to the second term in the formula; if the predicted value The missing labels are discarded.

[0048] In step 4, combining the single positive label of each sample with the pseudo label obtained in step 3, the low-dimensional feature embedding z predicted as the positive label can be summarized into each cluster, and the loss function is calculated using the label perception matrix

[0049] loss Apply consistency regularization to the model to encourage it to output the same probability distribution for the two enhanced versions of a sample, that is, to label the sample with a pseudo label through the weakly enhanced version, and then let the probability distribution output by the strongly enhanced version be close to the pseudo label. By using stronger enhancement, It can alleviate the overfitting problem of false negative labels and improve the model's ability to identify potential positive labels. In general, this regularization can optimize the supervision information of single positive labels and significantly reduce the negative impact of false negative labels. However, as the depth of the neural network increases, the effect of using high-order representation consistency on network parameters will gradually weaken, thereby weakening the loss This may cause the model to lack the ability to correct erroneous pathological inferences during the learning process, thereby overfitting the prediction results to these misdiagnosed diseases. This overfitting phenomenon is particularly dangerous in medical image classification because it may mislead clinical decisions. Misdiagnosis in the medical field may have serious ethical and legal consequences. If the model misdiagnoses due to problems in the training process, medical institutions and related personnel may need to bear corresponding responsibilities. In order to further alleviate the harm caused by erroneous negative labels, driven by the clustering hypothesis, the "Label-Aware Consistency" (LAC) loss is proposed to perform consistency regularization in low-order representations, thereby solving the problem of weakening the influence of high-order representation consistency on the model. When defining the loss function Before, we first embed z classification into the low-dimensional features predicted as positive labels according to the pseudo labels obtained in step 3 (including the single positive label of the sample itself). or There is c(z ni )=i, otherwise That is, only the sample predicts the value on the i-th category Greater than or equal to the threshold δ or a single positive label When it is the i-th category, the feature of the sample in the i-th category is embedded into z ni is classified as category i, otherwise the feature embedding is ignored; therefore, the category function c(·) to which the feature embedding belongs is defined as:

[0050]

[0051] For B samples, there are B'=B×L feature embeddings (each sample has L categories); z ni Select and constitute feature embedding as The operator is that the low-dimensional feature embedding of each sample is split into L parts according to the category dimension and the part with positive pseudo labels is filtered out to be used as the operator for subsequent calculations; the filtered feature embedding constitutes a new set, defined as

[0052] The calculation method of the label perception matrix i is defined as follows:

[0053]

[0054] The label-aware matrix i is used to construct correlations between label-level features. A value of 1 in the matrix corresponds to two features belonging to the same category, and 0 corresponds to features belonging to different categories. In medical images, grouping features corresponding to the same pathology into the same cluster facilitates full exploration of the pathogenesis of that pathology, allowing researchers to better understand the manifestations and patterns of changes in specific pathological features. This helps reveal the pathogenesis and progression of the disease.

[0055] final, Calculated by the following formula:

[0056]

[0057] Among them, exp(z j ·z j' / τ) is the feature embedding of the same category, It is to expand the feature embedding of different categories, and τ is the temperature coefficient.

[0058] loss The main idea is to aggregate the features of the same class in the feature space (the numerator) and separate the features of different classes (the denominator). and label-aware consistency regularization Perform consistency regularization from high-dimensional label space and low-dimensional feature embedding space respectively, and train the model simultaneously in two dimensions to reduce The impact of introducing incorrect labels.

[0059] Step 5: Minimize the total loss function As the target, update the model parameters and return to step 2 to iteratively optimize the model until the model converges.

[0060]

[0061] Among them, λ PLC and λ LAC is a weight hyperparameter used to adjust the weights of the three losses. Here, λ PLC and λ LAC are all set to 1. To minimize the total loss function As the target, the gradient descent method is used to update the model weights.

[0062] At this point, a round of single positive multi-label data training process is completed. If the model has converged, stop training and output the model modules required for classification: feature encoder E(·) and classifier f(·); otherwise, return to step 2 for the next round of training. In this embodiment, if the model performance is flat compared with the previous 10 rounds of training, it means that the model has converged. The output model can be directly used to assist specific medical tasks (determined by the type of training data). Because in clinical diagnosis, doctors need to analyze a large number of medical images to determine the type and severity of the disease. An automated image classification system can assist doctors in making preliminary diagnoses and improve the efficiency and accuracy of diagnosis. In this embodiment, the image to be diagnosed only needs to be input into the feature encoder E(·) and the classifier f(·) to obtain a preliminary diagnosis result.

[0063] The present invention is particularly suitable for fields such as medical image analysis, natural language processing, and recommendation systems. Traditional multi-label learning methods require data to have complete positive and negative labels. Therefore, when faced with single positive label data, it is usually difficult to fully utilize the correlation between labels, resulting in poor classification performance. The present invention significantly improves the effect of multi-label learning by constructing a label perception matrix and applying a global consistency criterion. In practical applications, the method of the present invention effectively improves the recognition accuracy of disease labels in medical images, improves the sentiment analysis and topic classification effects in natural language processing, and enhances the accuracy of personalized recommendations in recommendation systems.

[0064] To verify the effectiveness of the present invention, this embodiment gives the evaluation results of the proposed method on two multi-label image datasets: MS-COCO 2014 (COCO) and CUB-200-2011 (CUB), and compares them with the current state-of-the-art methods.

[0065] The division of the dataset and the number of categories included are shown in Table 1. The relevant information of each dataset will also be introduced in detail in the subsequent comparative experiments.

[0066] Table 1 Training, validation, and test partitions of different datasets and the number of included categories

[0067]

[0068]

[0069] The evaluation metric for the comparative experiments is the mean average precision (mAP) on the test set, which measures the average accuracy across all categories. All reported mAPs use the model weights at the time of achieving the highest mAP on the validation set of the corresponding dataset.

[0070] All methods use the Resnet-50 backbone network, loaded with weights pre-trained on ImageNet, to extract image features. To obtain label-level feature embeddings for comparative learning, the features extracted by Resnet-50 are positionally encoded and transformed and then input into a feature decomposition module. This module converts features into label-level feature embeddings and consists of a Transformer encoder and two Transformer decoders. The feature extraction module, combined with the feature decomposition module, forms the main network structure of the model in this paper, defined as E. Finally, a classification head f(·) and a nonlinear mapping head g(·) are added to generate various prediction values ​​and low-dimensional feature embeddings, respectively. Here, f(·) consists of a set of independent linear classifiers, and g(·) consists of a convolutional neural network.

[0071] The hidden layer dimension of Transform is set to 2048, which is consistent with the number of channels of Resnet-50. The dimension of the output of f(·) is consistent with the number of categories L (different for different datasets, see Table 1 for details). The dimension of the output of g(·) is L×d, where d is 128, to achieve the purpose of dimensionality reduction (2048→128). The size of all images is set to 448×448. During the training process, two image enhancement methods are used, one is weak enhancement (only random horizontal flipping) and the other is strong enhancement (including random horizontal flipping, cropping, etc.). The optimizer used in this paper is the AdamW optimizer with a weight decay of 10 -2 , the learning rate is adjusted using OneCycleLR, and the maximum learning rate is 10 -4 , with an increase ratio of 0.2. Different batch sizes {8, 16, 32, 64} were considered, and the total number of training rounds was 40. However, if the model performance did not improve within a certain number of rounds, the training was terminated early. Set λ PLC =1,λ LAC = 1, threshold δ = {0.5, 0.6, 0.7, 0.8, 0.9}, τ = {0.1, 0.5, 1.0, 1.5, 2.0}. All hyperparameters were selected to achieve the best mAP on the validation set. Table 2 shows the experimental results of the comparison method and the present invention on different datasets.

[0072] Table 2 mAP metrics of different methods on different datasets

[0073]

[0074]

[0075] The MS-COCO 2014 dataset contains 82,081 training images and 40,137 test images, with a label space of 80. The training set is divided into 65,665 training samples and 16,416 validation samples. Table 2 lists the mAP of the best model of the present invention on the COCO dataset, which reaches 75.6, which is higher than the advanced methods. It is improved by 3.5. As can be seen, the improvement on the COCO dataset is very significant, proving the effectiveness of the consistency regularization method proposed in this paper.

[0076] The CUB-200-2011 dataset contains 5994 training images and 5794 test images, with a label space of 312. The split training set consists of 4795 training samples and 1199 validation samples. Table 2 lists the mAP of the best model of the present invention on the CUB dataset, which reaches 23.0, which is higher than the SOTA method. Although the performance of the model itself is limited (all methods have mAP around 20), its ample unlabeled space (with an average number of labels of 31.4) provides room for the consistency regularization method proposed in this paper to play a role.

[0077] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A single positive multi-label classification method based on label-aware global consistency, applied to image processing, characterized by: The following steps are involved: (1) Given a training dataset with N samples, merge and initialize the model components; (2) Randomly select B samples from the training data set, assume that all unknown labels of each sample are negative labels, and calculate the loss function (3) Using the positive and negative pseudo-label thresholds δ and Divide the classification prediction value p of each sample to obtain credible positive and negative pseudo labels, and calculate the pseudo label consistency regularization loss (4) Combine the single positive label of each sample with the pseudo label obtained in step (3), embed the low-dimensional feature z predicted as the positive label into each cluster, and use the label perception matrix to calculate the label perception consistency regularization loss (5) With the goal of minimizing the total loss function, update the model parameters and return to step (2) to iteratively optimize the model until the model converges; The implementation process of step (4) is as follows: According to the obtained pseudo-label, the low-dimensional feature predicted as the positive label is embedded into the z classification. or There is c(z ni )=i, that is, only the sample predicts the value in the i-th category Greater than or equal to the threshold δ or a single positive label When it is the i-th category, the feature of the sample in the i-th category is embedded into z ni is classified into category i; therefore, the feature embedding category function c(·) is defined as: For B samples, there are B'=B×L feature embeddings (each sample has L categories); z ni Select and form feature embedding as The operator is that the low-dimensional feature embedding of each sample is split into L parts according to the category dimension and the part with positive pseudo labels is filtered out to be used as the operator for subsequent calculations; the filtered feature embedding constitutes a new set, defined as The label perception matrix I is defined as follows: final, Calculated by the following formula: Among them, exp(z j ·z j′ / τ) is used to aggregate feature embeddings of the same category. The role of is to distance the feature embedding of different categories, and τ is the temperature coefficient.

2. The single positive multi-label classification method based on label-aware global consistency according to claim 1 is characterized in that: The implementation process of step (1) is as follows: Training data set Among them, x n For the sample, is a single positive label; the initialization model components include: feature encoder E(·), classifier f(·) and nonlinear mapper g(·), where f(·) and g(·) process the sample features extracted by E(·) and output the classification prediction value p and low-dimensional feature embedding z; the positive and negative pseudo label thresholds δ and Confident labels for classifying samples.

3. The single positive multi-label classification method based on label-aware global consistency according to claim 1 is characterized in that: The loss function in step (2) for: Using BCE loss, and introducing weak enhancement technology to increase the number of visible samples of the model, while ensuring the integrity of the enhanced semantic information, for: in, For the classifier f(·) to predict the input Aug w (x n )’s class probability vector, Aug w (·) represents a weak enhancement method, is a single positive label; L is the total number of categories, and the subscript i is the category index, that is, express The value of the i-th category, for The value in the i-th category.

4. The single positive multi-label classification method based on label-aware global consistency according to claim 1 is characterized in that: The implementation process of step (3) is as follows: Introducing strong enhancement Aug s (·) Strengthen consistency, encourage the prediction results of strong and weak enhancement versions of the same sample to be consistent; By using the positive and negative pseudo-label thresholds δ and Filter out high-quality pseudo labels as the object of consistency regularization, that is, discard Partial label; Calculated by the following formula: in, For the classifier f(·) to predict the input Aug s (x n )’s class probability vector, Aug s (·) represents a strong enhancement method, For the classifier f(·) to predict the input Aug w (x n ), is a single positive label, L is the total number of categories, and the subscript i is the category index, that is, express The value of the i-th category; when the weak enhancement sample Aug w (x n ) Greater than or equal to the threshold δ or the original label When is positive, the sample x n The i-th class is marked as positive, corresponding to the first term in the formula; when the predicted value Less than threshold When , the sample x n The i-th class is marked as negative, corresponding to the second term in the formula; if the predicted value The missing labels are discarded.

5. The single positive multi-label classification method based on label-aware global consistency according to claim 1 is characterized in that: The total loss function in step (5) is: Among them, λ PLC and λ LAC is a weight hyperparameter used to adjust the weight of each loss. is the pseudo-label consistency regularization loss, is the label-aware consistency regularization loss.

6. The single positive multi-label classification method based on label-aware global consistency according to claim 2 is characterized in that: The classifier f(·) uses one fully connected layer as an example, and the nonlinear mapper g(·) uses two fully connected layers and one activation layer.

7. The single positive multi-label classification method based on label-aware global consistency according to claim 3 is characterized in that: The weak enhancement method includes horizontal and vertical flipping of the image and synonym replacement of the text.

8. The single positive multi-label classification method based on label-aware global consistency according to claim 4 is characterized in that: The strong enhancement method includes random cropping and flipping of images, and random deletion of text.

9. The single positive multi-label classification method based on label-aware global consistency according to claim 5, characterized in that: The λ PLC and λ LAC The value is 1.

Images