A marking enhancement method and system for single positive multiple marking

Abstract

The present invention discloses a label enhancement method and system for single-positive multi-label. First, in the data preprocessing stage, a multi-label image classification dataset is input. The dataset is divided into a training set and a test set by positive labels, and the training set data is preprocessed; then, a category prior estimation method is used to estimate the category prior probability of each label, and the samples are labeled with pseudo labels to obtain a label distribution; then, a neural network model is constructed, and the training set is used to train the model to obtain the optimal parameters, and the optimal prediction model is obtained by minimizing the objective function; finally, the test set is classified according to the obtained optimal prediction model to obtain the target label classification result. This method solves the label incompleteness and noise problems in traditional methods by using the estimated category prior probability to generate pseudo labels, and improves the label quality and quantity of the training dataset. This method has broad application prospects in practical applications and can improve the performance and generalization ability of the model.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and mainly relates to a label enhancement method and system for single positive multiple labels. Background Art

[0002] Single-positive multi-label learning is an important class of machine learning tasks, where each example may be assigned multiple labels, not just a single one. This task is widely used in real-world applications, such as image classification, natural language processing, and recommender systems. However, due to issues such as incomplete labeling, label noise, and sample imbalance, traditional single-positive multi-label learning methods perform poorly in complex real-world scenarios.

[0003] To improve the performance of single-positive multi-label learning tasks, researchers have proposed label augmentation methods. Label augmentation methods improve the quality and quantity of labels for samples in the training dataset to enhance model performance and generalization. Traditional label augmentation methods include label propagation, active learning, and transfer learning. While these methods can improve model performance to a certain extent, some issues remain.

[0004] First, traditional label enhancement methods usually obtain label distribution based on correlation assumptions and then unify the sample categories through threshold division. This lacks the utilization of prior knowledge of different categories; second, label propagation methods may introduce more erroneous labels when dealing with incomplete or noisy labels, affecting the generalization performance of the model; finally, active learning and transfer learning methods also have certain limitations in obtaining high-quality and diverse labeled data. Summary of the Invention

[0005] The present invention addresses the limitations of traditional label enhancement methods in the prior art and provides a label enhancement method and system for single-positive multi-label applications. First, in the data preprocessing stage, a multi-label image classification dataset is input. The dataset is divided into a training set and a test set by positive labels, and the training set data is preprocessed. A category prior estimation method is then used to estimate the category prior probability of each label, and the samples are pseudo-labeled to obtain a label distribution. A neural network model is then constructed, and the model is trained using the training set to obtain optimal parameters. The optimal prediction model is obtained by minimizing the objective function. Finally, the test set is classified based on the obtained optimal prediction model to obtain the target label classification result. This method solves the label incompleteness and noise problems of traditional methods by using the estimated category prior probabilities to generate pseudo-labels, thereby improving the label quality and quantity of the training dataset. This method has broad application prospects in practical applications and can improve the performance and generalization ability of the model.

[0006] In order to achieve the above objectives, the technical solution adopted by the present invention is: a label enhancement method for single positive multiple labels, which pseudo-labels unknown samples according to the estimated category prior probability, uses a neural network to train the model, and improves the prediction ability of the model by proposing and optimizing the loss function, ultimately achieving label enhancement of single positive multiple labels.

[0007] As an improvement of the present invention, the method specifically includes:

[0008] S1, data preprocessing stage: input a multi-label image classification dataset, the dataset is positively labeled and divided into a training set and a test set, and the training set data is preprocessed;

[0009] S2, model training stage: use the category prior estimation method to estimate the category prior probability of each label, give the sample a pseudo label to obtain the label distribution; then build a neural network model, use the training set to train the model to obtain the optimal parameters, and obtain the optimal prediction model by minimizing the objective function;

[0010] S3, prediction stage: classify the test set according to the optimal prediction model obtained in step S2 to obtain the target label classification result.

[0011] As an improvement of the present invention, the category prior estimation method in step S2 estimates the category prior probability of each tag as follows: the single positive tag corresponding to each instance is uniformly sampled from its corresponding related tag set, which is expressed as follows:

[0012]

[0013] Among them, p(l k =1|y) represents the probability that each positive tag in the tag vector is selected as a single positive tag for a known instance of the tag vector y, I(·) is the indicator function, Represents the number of positive labels and decomposes the prior probability of a single positive label:

[0014] p(l k =1) =∫p(l k =1,y)dy

[0015] =∫p(l k =1|y)p(y)dy

[0016] Where p(l k =1|y) can be decomposed into:

[0017] p(l k =1|y)=p(l k =1|y k =0,y)p(y k =0)+p(lk =1|y k =1,y)p(y k =1)

[0018] In the known y k = 0, the kth mark cannot be sampled as a single positive mark, that is, p(l k =1|y k =0,y)=0, so it can be simplified to:

[0019] p(l k =1|y)=p(l k =1|y k =1,y)p(y k =1)

[0020] so,

[0021]

[0022] Use the category prior of the observed single positive label vector to estimate the category prior of the true label at the location, let is a hyperparameter, then:

[0023] p(l k =1) = p(y k =1)α

[0024] Using the single positive labeled pair p(l k =1) Estimate:

[0025]

[0026] Then, the class prior probability of each class is estimated as:

[0027]

[0028] As another improvement of the present invention, the pseudo-labeling of samples in step S2 is specifically as follows: after obtaining the prior probability of each category, the number of positive labels of each category in the data set can be estimated. By sorting the output of the network, the The large output is marked with a positive label, otherwise it is marked with a negative label;

[0029] Initialize the pseudo-label vector of each instance to a single positive label vector to warm up the model, and then update the pseudo-label in each update step:

[0030]

[0031] where rank(f k (x i)) represents the model for instance x i The ranking of the output of the kth token in the entire dataset,

[0032]

[0033] Where m is the currently sorted token and n is the total number of tokens.

[0034] As another improvement of the present invention, the target loss function in the neural network model in step S2 is specifically:

[0035]

[0036] in, is the target loss function, It is a pseudo-marker.

[0037] In order to achieve the above purpose, the present invention also adopts a technical solution: a marking enhancement system for single positive multiple markings, including a computer program, which implements the steps of any of the above methods when executed by a processor.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] (1) This method generates pseudo labels by using the estimated category prior probability, which solves the labeling incompleteness and noise problems in traditional methods and improves the labeling quality and quantity of the training dataset. This method has broad application prospects in practical applications and can improve the performance and generalization ability of the model.

[0040] (2) This method estimates and utilizes the prior probability of each category to generate personalized data augmentation strategies for different categories, overcoming the limitations of traditional labeling methods. Category prior estimates can be derived from historical data, expert knowledge, or other relevant information to help guide the data augmentation process. By flexibly utilizing category prior estimates, this method can enhance the labeling quality and quantity of samples from different categories in the training dataset.

[0041] (3) This method can solve the problem of expensive manual labeling and can realize multi-label learning problems when the number of labels is limited. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flowchart of the steps of a single positive multiple marker marker enhancement method of the present invention;

[0043] Figure 2 This is a model architecture diagram of a single-positive multi-label marker enhancement method of the present invention. DETAILED DESCRIPTION

[0044] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0045] Example 1

[0046] A label enhancement algorithm for single positive multiple labels. Traditional label enhancement (LE) algorithms usually use the correlation between categories to propagate labels. After obtaining the label distribution, all instances are labeled with the same threshold. However, this often ignores prior information and introduces noise. At the same time, due to the lack of labeled data, the category prior probability cannot be directly obtained. This method considers the prior information between different labels, reduces the risk of noise introduction, and generates pseudo labels for samples based on the estimated category prior probability to improve the accuracy of the model and achieve label enhancement. Figure 1 ,2, the method of the present invention comprises the following steps:

[0047] S1, data preprocessing stage: input a multi-label image classification dataset, the dataset is positively labeled and divided into a training set and a test set, and the training set data is preprocessed;

[0048] Collect a partially labeled training set, where each instance has a positive label. Split the dataset into training and test sets. We used four large-scale multi-label image classification (MLIC) datasets. For each MLIC dataset, we reserved 20% of the training set for validation.

[0049] S2, model training phase: Use the category prior estimation method to estimate the category prior probability of each label, and give the samples pseudo labels to obtain the label distribution; then build a neural network model, use the training set to train the model to obtain the optimal parameters, and obtain the optimal prediction model by minimizing the objective function;

[0050] Based on the single positive label in the training set, the class prior probability of each label is estimated using the class prior estimation method. The class prior estimation method used in the present invention is based on the assumption that, from the perspective of data generation, the single positive label corresponding to each instance is uniformly sampled from its corresponding set of related labels, which can be expressed as:

[0051]

[0052] Where p(l k =1|y) represents the probability that each positive tag in the tag vector is selected as a single positive tag for a known instance of the tag vector y, I(·) is the indicator function, Indicates the number of positive labels.

[0053] Based on this assumption, the prior probability of a single positive label is decomposed as:

[0054] p(l k =1) =∫p(l k =1,y)dy

[0055] =∫p(l k =1|y)p(y)dy

[0056] Where p(l k =1|y) can be decomposed into:

[0057] p(l k =1|y)=p(l k =1|y k =0,y)p(y k =0)+p(l k =1|y k =1,y)p(y k =1)

[0058] According to the above formula, the category prior estimation method is based on the assumption that each instance corresponds to a single positive label, and the category prior probability of each label is calculated by the formula, that is, when y is known k = 0, the kth mark cannot be sampled as a single positive mark, that is, p(l k =1|y k =0,y)=0, so it can be simplified to:

[0059] p(l k =1|y)=p(l k =1|y k =1,y)p(y k =1)

[0060] so,

[0061]

[0062] According to the simplified results above, we can use the category prior of the observed single positive label vector to estimate the category prior of the actual label of the position, let is a hyperparameter, then:

[0063] p(l k =1) = p(y k =1)α

[0064] Using the single positive labeled pair p(l k =1) Estimate:

[0065]

[0066] Then the class prior probability of each class can be estimated as:

[0067]

[0068] After obtaining the prior estimate, generate pseudo labels based on the estimated category prior probability, sort the model output in descending order, treat the first part of the labels as positive labels, and the last part of the labels as negative labels. Generate pseudo labels based on category prior estimates: After obtaining the prior probability of each category, you can estimate the number of positive labels for each category in the data set. By sorting the output of the network, the The output with a large value is labeled as a positive label, otherwise it is labeled as a negative label. First, the pseudo label vector of each instance is initialized to a single positive label vector to warm up the model, and then the pseudo label is updated in each update step.

[0069]

[0070] where rank(f k (x i )) represents the model for instance x i The ranking of the output of the kth token in the entire dataset,

[0071]

[0072] According to the above ranking results, the data is pseudo-labeled and the pseudo-labeled dataset is used for training to improve the accuracy of the model:

[0073]

[0074] S3, prediction stage: According to the above model, the model is trained using the training set, and optimized through a computer cluster to finally obtain the optimal prediction model, thereby classifying the test set and obtaining the target label classification result.

[0075] The proposed algorithm is compared with five existing single-positive multi-label learning verification methods, and the results are shown in the table below:

[0076] Average accuracy VOC COCO NUS CUB AN 85.546±0.294 64.326±0.204 42.494±0.338 18.656±0.090 AN-LS 87.548±0.137 67.074±0.196 43.616±0.342 16.446±0.269 WAN 87.138±0.240 65.552±0.171 45.785±0.192 14.622±1.300 EPR 85.228±0.444 63.604±0.249 45.240±0.338 19.842±0.423 ROLE 88.088±0.167 67.022±0.141 41.949±0.205 14.798±0.613 EM 88.674±0.077 70.636±0.094 47.254±0.297 20.692±0.527 EM-APL 88.860±0.080 70.758±0.215 47.778±0.181 21.202±0.792 SMILE 86.311±0.450 63.331±0.112 43.611±0.172 18.611±0.144 LAGC 88.021±0.121 70.422±0.062 46.211±0.155 21.840±0.237 The present invention 90.063±0.032 74.862±0.046 51.088±0.050 22.583±0.110

[0077] As can be seen from the above table, the experimental results of the method of the present invention are superior, and compared with other comparison algorithms, the experimental results of our algorithm are comprehensively ahead.

[0078] In summary, the present invention discloses a label enhancement method for single positive multiple labels, proposes an unbiased estimation method for the category prior of each label, uses the estimated category prior to pseudo-label unknown labels to achieve label enhancement, uses a neural network to train the model, and improves the prediction ability of the model by proposing and optimizing the loss function, ultimately achieving label enhancement in the single positive multiple label problem.

[0079] It should be noted that the above content merely illustrates the technical idea of ​​the present invention and cannot be used to limit the scope of protection of the present invention. 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 all fall within the scope of protection of the claims of the present invention.

Claims

1. A method for single positive multi-label marker enhancement, characterized in that : Based on the estimated class prior probability, pseudo-label the unknown samples, use the neural network to train the model, and improve the model's prediction ability by proposing and optimizing the loss function, ultimately achieving single positive multi-label label enhancement; specifically, the following steps are included: S1, data preprocessing stage: input multi-label image classification dataset, the dataset is positively labeled and divided into training set and test set, the training set data is preprocessed, and the convolutional neural network is used to capture the features of the image; S2, model training phase: Use the category prior estimation method to estimate the category prior probability of each label, and give the samples pseudo labels to obtain the label distribution; then build a neural network model, use the training set to train the model to obtain the optimal parameters, and obtain the optimal prediction model by minimizing the objective function; The category prior estimation method estimates the category prior probability of each label as follows: the single positive label corresponding to each instance is uniformly sampled from its corresponding related label set, and the prior probability of the single positive label is decomposed and resolved: Use the category prior of the observed single positive label vector to estimate the category prior of the true label at the location, let is a hyperparameter, I(·) is an indicator function, Represents the number of positive labels, k = 1…c, then: p(l k =1)=p(y k =1)a; Using the single positive labeled pair p(l k =1) Estimate: Then, the class prior probability of each class is estimated as: The specific method of pseudo-labeling samples is as follows: after obtaining the prior probability of each category, the number of positive labels of each category in the data set can be estimated. By sorting the output of the network, the The large output is marked with a positive label, otherwise it is marked with a negative label; Initialize the pseudo-label vector of each instance to a single positive label vector to warm up the model, and then update the pseudo-label in each update step: where rank(f k (x i )) represents the model for instance x i The ranking of the output of the kth tag in the entire dataset, Among them, m is the current sorted token, and n is the total number of tokens; S3, prediction stage: classify the test set according to the optimal prediction model obtained in step S2 to obtain the target label classification result.

2. The method for single positive multi-marker label enhancement according to claim 1, characterized in that: The target loss function in the neural network model of step S2 is specifically: in, is the target loss function, It is a pseudo-marker.

3. A single positive multi-marker marker enhancement system, comprising a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 2 are implemented.

Images