A multi-label classification model construction method, a classification method, a classification device, electronic equipment and a storage medium

Abstract

This invention discloses a method for constructing a multi-label classification model, a classification method, a classification device, an electronic device, and a storage medium, belonging to the field of multi-label learning technology. This invention utilizes PU learning technology, designs a new loss function and an adaptive temperature coefficient module, thereby optimizing the multi-label classification model. The designed loss function can achieve good classification accuracy using only partially labeled positive labels and unknown labels during model training. The adaptive temperature coefficient module can dynamically adjust the dispersion of predicted scores, alleviating the problem of overly smooth predicted scores in the early stages of model training, thus accelerating model convergence and improving classification accuracy.

Description

Technical Field

[0001] This invention belongs to the field of machine learning technology, and in particular relates to a method for constructing a multi-label classification model with missing labels based on PU learning, a classification method, a classification device, an electronic device, and a storage medium. Background Technology

[0002] Traditional supervised learning primarily relies on single-label learning, where each image corresponds to a unique label. However, real-world target samples are often quite complex, and traditional single-label annotation methods cannot meet the needs of labeling complex samples. This is because, in the real world, a sample typically belongs to multiple categories simultaneously, which is the so-called multi-label classification problem.

[0003] However, in traditional multi-label learning, most methods rely on fully labeled multi-label datasets. This leads to high dataset labeling costs and, due to differences in the understanding of different data labelers, can result in mislabeling. For example, in medical data, the limited knowledge of medical-related subjects among data labelers can easily lead to mislabeling. Therefore, missing-label multi-label learning has attracted increasing attention from researchers. Its goal is to train a model using only partially labeled data, thereby achieving good classification accuracy. In the training set of traditional missing-label multi-label learning, labels are divided into known and unknown labels. Known labels include both positive and negative labels, and the same applies to unknown labels. Missing-label multi-label learning uses known labels in the dataset to identify and predict unknown labels through methods such as reweighting, semantic relevance, or generating pseudo-labels. Missing-label multi-label learning not only reduces the labeling cost of multi-label datasets but also makes its applications more relevant to real-life scenarios. Research on missing-label multi-label learning methods has, to some extent, promoted the application of multi-label learning methods in practical life.

[0004] Current missing-label multi-label learning methods rely on using positive and negative labels from the known labels to train the model. However, due to the labeling method of multi-label datasets—that is, only the categories present in the image are labeled positively, and the remaining categories are assumed not to exist in the image and labeled negatively—a large number of unlabeled positive labels are included in the negative labels, thus affecting model optimization and ultimately impacting the final classification accuracy. Furthermore, missing-label multi-label learning methods based on generating pseudo-labels also struggle to avoid generating new erroneous labels during the pseudo-label generation process, which also negatively impacts model optimization. In addition, in traditional missing-label multi-label learning methods, using partially labeled positive and negative samples for model training incurs significant dataset labeling costs. Summary of the Invention

[0005] 1. The problem to be solved

[0006] The purpose of this invention is to provide a method, classification method, classification device, electronic device, and storage medium for constructing a multi-label classification model with missing labels based on PU learning, thereby solving the problem of relatively low classification accuracy in traditional multi-label learning methods with missing labels. Based on PU learning technology, this invention designs a loss function and an adaptive temperature coefficient module, achieving model training and good classification accuracy using only partially labeled positive labels and unknown labels.

[0007] 2. Technical Solution

[0008] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0009] The present invention provides a method for constructing a multi-label classification model, comprising the following steps:

[0010] S1. Input the training image set and extract image features;

[0011] S2. Using the semantic decoupling module, a category-specific semantic representation vector is obtained by integrating the semantic information of each category in the image features extracted in S1.

[0012] S3. Feed the obtained category-specific semantic representation vector into the linear classifier and calculate the prediction score for each label;

[0013] S4. Perform image-level blending on the input model to obtain a blended image. Repeat steps S1-S3 to calculate the prediction score of the blended image. Simultaneously, blend the prediction scores of the original images to obtain a mixed prediction score. Input the prediction scores of the blended image and the mixed prediction score into the loss function to obtain the regularization term L of the overall loss. reg ;

[0014] S5. Using the adaptive temperature coefficient module, calculate the standard deviation of the predicted score obtained in S3, and calculate the adaptive temperature coefficient τ accordingly to dynamically adjust the predicted score.

[0015] S6. Input the predicted score of the original image and the label under the PU setting into the loss function to obtain the classification loss term L. var ;

[0016] S7. Based on the regularization term L obtained in S4 reg And the classification loss term L obtained in S6 var We construct the overall loss function, and then optimize the model.

[0017] Furthermore, the regularization term L of the overall loss in S4 reg The process of obtaining it is as follows:

[0018] Step 1: Mix N images I according to coefficient μ to obtain N / 2 mixed images. Input the mixed images into the prediction model to obtain the prediction score of the mixed images.

[0019]

[0020]

[0021] In formulas (3)-(4), the coefficient μ follows a Beta distribution with parameter β; Φ represents the prediction model; Represents the prediction score of the mixed image; and Let represent the first half and the second half of N images I, respectively;

[0022] Step 2: Mix the prediction scores S of the N images I according to the coefficient μ to obtain the mixed prediction scores:

[0023]

[0024] In formula (5), the coefficient μ follows a Beta distribution with parameter β; and Let S represent the first half and the second half of the prediction scores S for N images I, respectively.

[0025] Step 3, and The regularization term for the overall loss is obtained by feeding it into the mixed loss function:

[0026]

[0027] In formula (6), σ(·) represents the sigmoid activation function; express The predicted score of the c-th category in the n-th image; express The predicted score of the c-th category for the n-th image.

[0028] Furthermore, in S6, the classification loss term L var The calculation is as follows:

[0029]

[0030]

[0031] In the above formula, Represents the set of unknown labels under category c in an image collection; Represents the combination of known labels under category c in an image set; s u Indicates that the label belongs to The corresponding predicted score; s p Indicates that the label belongs to The corresponding predicted score; P represents the average predicted probability of each category in an image set; the parameter γ represents the balance exponent, which controls the penalty for different proportions of known labels; γ This represents the rebalancing factor, used to adjust the contribution of unknown labels to the loss function.

[0032] Furthermore, the construction process of the overall loss function in step S7 is as follows:

[0033] Step 1: Adjust the adaptive temperature coefficient τ in step S3. c Substituting into formula (15), for the classification loss term L... var Optimize:

[0034]

[0035] Where, τ c This represents the adaptive temperature coefficient for the c-th category;

[0036] Step 2: Add L to formulas (6) and (15) reg and L var Substituting into the equation yields the overall loss function L. pu-mlc That is, we get:

[0037] L pu-mlc =L var +λL reg (17)

[0038] In formula (17), the parameter λ is used to adjust the weight of the regularization term.

[0039] Furthermore, the adaptive temperature coefficient τ is calculated as follows:

[0040]

[0041] In formula (7), Represents the temperature coefficients for category C; d represents the average predicted score for category c in N images; n,c This represents the predicted score of the c-th category in the n-th image of S; the parameter α represents the penalty coefficient, which controls the adjustment capability of the temperature coefficient.

[0042] Furthermore, in step S1, a deep neural network is used to extract image features from the training image set.

[0043] The present invention provides a multi-label classification method in which the sample to be tested is input into the model constructed above to obtain the predicted value of the sample, and then the presence of a label category in the sample is determined according to the classification threshold.

[0044] A multi-label classification device of the present invention includes a multi-label classification model construction unit for constructing a multi-label classification model; and a label prediction unit for inputting a sample to be tested into the constructed multi-label classification model to obtain a predicted value of the sample, and then determining whether a label category exists in the sample based on a classification threshold; wherein, the multi-label classification model construction unit includes:

[0045] The feature extraction module is used to extract image features from the training image set;

[0046] The semantic decoupling module is used to obtain a category-specific semantic representation vector by integrating the semantic information of each category in the image features extracted by S1;

[0047] The score prediction module is used to feed the obtained category-specific semantic representation vector into the linear classifier to calculate the predicted score for each label;

[0048] The image-level fusion module performs image-level fusion on the input model to obtain a fused image and calculates the prediction score of the fused image. Simultaneously, it fuses the prediction scores of the original images to obtain a fused prediction score. The prediction scores of the fused image and the fused prediction score are input into the loss function to obtain the regularization term L of the overall loss. reg ;

[0049] The adaptive temperature coefficient module is used to calculate the standard deviation of the predicted score obtained in S3, and to calculate the adaptive temperature coefficient τ accordingly.

[0050] The classification loss term calculation module is used to input the predicted score of the original image and the label under the PU setting into the loss function to calculate the classification loss term;

[0051] The model optimization module is used to construct the overall loss function based on the calculated regularization term and classification loss term, thereby optimizing the model.

[0052] The present invention provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the classification method of the present invention.

[0053] An electronic device according to the present invention includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the classification method of the present invention.

[0054] Compared with existing technologies, the present invention can achieve the following beneficial effects:

[0055] (1) Based on PU learning technology, this invention optimizes the design of loss function and adaptive temperature coefficient module to train the model using only partially labeled positive labels and unknown labels, and achieves good classification accuracy. It solves the problem of a large number of mislabeled data in negative labels in traditional multi-label learning with missing labels, and further alleviates the problem of high labeling cost of multi-label datasets.

[0056] (2) Due to the small number of positive labels in the training set under the multi-label setting with missing labels, the prediction probability in the early stage of model training will be over-smoothed. This invention can dynamically adjust the dispersion of the prediction probability by designing an adaptive temperature coefficient, thereby accelerating the convergence of the model and further improving the final model classification accuracy. Attached Figure Description

[0057] Figure 1 This is a model framework diagram of the multi-label classification method with missing labels in this invention. Detailed Implementation

[0058] To further understand the content of the present invention, the present invention will be further described below with reference to specific embodiments. Example

[0059] like Figure 1 As shown in this embodiment, a multi-label classification method for missing labels based on PU learning includes three stages: model building, training, and label prediction. The specific steps are as follows:

[0060] I. Model Building and Training:

[0061] S1. Input the training image set and load a pre-trained deep convolutional neural network to extract image features.

[0062] The input images in the dataset are Where N represents the number of samples in an image set, and c, w, and h represent the number of image channels, width, and height, respectively. In this embodiment, c ∈ {3}, w, and h ∈ {4, 4, and 8}. Represents the real number field; the positive label matrix P∈{0,1} of the input image I. N×C and the unknown label matrix U∈{0,1} N×C C represents the number of labeled categories in the dataset, where 0 represents an unknown label and 1 represents a positive label.

[0063] Features are extracted from image i using a backbone network. Specifically, in this embodiment, the image features are obtained using a deep neural network ResNet101, and the output of the last layer is extracted as the image features.

[0064] S2. Using the semantic decoupling module, a category-specific semantic representation vector is obtained by integrating the semantic information of each category in the image features extracted in S1. The specific operation is as follows:

[0065] Load the semantic embedding vectors extracted by the pre-trained model, and use the semantic decoupling module to obtain the semantic representation vectors of image i specific to C categories:

[0066]

[0067] In formula (1), The semantic decoupling module is described in detail in Section 3.2 of the paper "Learning Semantic-Specific Graph Representation for Multi-Label Image Recognition". This represents the semantic embedding vectors of C categories extracted using the GloVe model. In this embodiment, d s =300. Let be the category-specific semantic representation vector of image i; the category-specific semantic representation vectors of N images are:

[0068] S3. Feed the category-specific semantic representation vector F obtained in S2 into the linear classifier and calculate the prediction score for each label, specifically expressed as follows:

[0069]

[0070] In formula (2), Represents a linear classifier. Let represent the predicted scores for C categories in N images.

[0071] S4. Design an image-level fusion module to perform image-level fusion on the input model to obtain a fused image. Repeat steps S1-S3 to calculate the prediction score of the fused image. Simultaneously, fuse the prediction scores of the original images to obtain a fused prediction score. Input the prediction scores of the fused image and the fused prediction score into the loss function to obtain the regularization term L of the overall loss. reg The regularization term L for the overall loss reg The specific process of obtaining it is as follows:

[0072] Step 1: First, mix the N images I according to the coefficient μ to obtain N / 2 mixed images; input the mixed images into the prediction model to obtain the prediction score of the mixed image:

[0073]

[0074]

[0075] In formulas (3)-(4), the coefficient μ follows a Beta distribution with parameter β, which is set to 0.3 in this embodiment; Φ represents the prediction model; Represents the prediction score of the mixed image; and Let represent the first half and the last half of N images I, respectively.

[0076] Step 2: Mix the prediction scores S of the N images I according to the coefficient μ to obtain the mixed prediction scores:

[0077]

[0078] In formula (5), the coefficient μ follows a Beta distribution with parameter β; and Let S represent the first half and the second half of the prediction scores S for N images I, respectively.

[0079] Step 3, Finally, and By feeding the mixture into the loss function, we obtain the regularization term for the overall loss:

[0080]

[0081] In formula (6), σ(·) represents the sigmoid activation function; express The predicted score of the c-th category in the n-th image; express The predicted score of the c-th category for the n-th image.

[0082] S5. Design an adaptive temperature coefficient module to calculate the standard deviation of the predicted scores obtained in S3, and use it as a temperature coefficient to dynamically adjust the predicted scores, thereby alleviating the over-smoothing of classification probabilities in the early stage of model training. The adaptive temperature coefficient is specifically expressed as follows:

[0083]

[0084] In formula (7), Represents the temperature coefficients for category C; s represents the average predicted score for category c in N images;n,c This represents the predicted score of the c-th category in the n-th image of S; the parameter α represents the penalty coefficient, which is used to control the adjustment capability of the temperature coefficient. It is set manually and takes a value between 0 and 1 (excluding 0).

[0085] S6. Based on the methods in PU learning, a new loss function is proposed. The predicted score of the original image and the label under the PU settings are input into this loss function to obtain L. var The specific operation is as follows:

[0086] Step 1: First, the traditional multi-label classification loss L... mlc Represented as:

[0087]

[0088] In formula (8), This indicates that the indicator function takes a value of 1 when the condition within the parentheses is met, and a value of 0 when the condition within the parentheses is not met; L + L represents the multi-label classification loss function. mlc Positive label items; L - L represents the multi-label classification loss function. mlc Negative label items; y n,c This indicates the label under the PN setting, where +1 represents a positive label and -1 represents a negative label. In formula (8), L... mlc Transformed into expected risk R pn The form is:

[0089]

[0090] In formula (9), π p Indicates a priori positive label class; Represents the set of positive labels; Represents the set of negative labels; Let L represent the expectation function. + and L - The expectation can be specifically expressed as:

[0091]

[0092]

[0093] In formulas (10)-(11), s m Let S represent any predicted score that satisfies the constraints.

[0094] Step 2: Further convert the expected risk form of the multi-label classification loss under the PN setting into the expected risk R under the PU setting. var :

[0095]

[0096] In formula (12), log represents the logarithmic function; This represents the set of unknown labels. Using the loss function in this embodiment allows the model to be optimized using only positive and unknown labels. Labels in the PU settings: +1 represents a positive label, and 0 represents an unknown label. R... var The loss function can be expressed in the form of:

[0097]

[0098] In formula (13), Represents the set of unknown labels under category c in an image collection; Represents the combination of known labels under category c in an image set; s u Indicates that the label belongs to The corresponding predicted score; s r Indicates that the label belongs to The corresponding predicted score.

[0099] Step 3: Finally, design a rebalancing factor to mitigate the impact of positive and unknown labels on the loss function R. var The problems include imbalanced contributions and varying training difficulty for samples of different classes.

[0100]

[0101]

[0102] In formulas (14)-(15), P represents the average predicted probability of each category in an image set; the parameter γ represents the balance index, which controls the penalty for different proportions of known labels, and its value is between 0 and 1 (excluding 0); γ The overall representation is the rebalancing factor, which adjusts the contribution of unknown labels to the loss function.

[0103] S7, take the L obtained in S4 reg As a regularization term, L obtained in S6 var The overall loss function, composed of loss function terms, is used to optimize the model. The overall loss function is specifically structured as follows:

[0104] Step 1: First, substitute the adaptive temperature coefficient τ into formula (15):

[0105]

[0106] In formula (16), τ c This represents the adaptive temperature coefficient for the c-th category.

[0107] Step 2: Combine the L values ​​in formulas (6) and (15) reg and L var Substituting into the equation yields the overall loss function L. pu-mlc That is, we get:

[0108] L pu-mlc =L var +λL reg (17)

[0109] In formula (17), the parameter λ is used to adjust the weight of the regularization term.

[0110] II. Tag Prediction

[0111] The test sample is input into the model trained in S1-S7 to obtain the predicted value of the sample. Then, the presence of a labeled category in the test sample is determined based on the classification threshold. The specific method is as follows:

[0112] (1) Substitute the sample to be tested x into the final classification model obtained through steps S1-S7. The parameters learned by the model are θ. The predicted value is obtained based on the model.

[0113] (2) Set a classification threshold t, and calculate the final output label vector y of the test sample in C categories based on the predicted value of the test sample x and the classification threshold t. t ∈{0,1} 1×C ,in This indicates an indicator function. Based on the calculation result of formula (18), it returns 1 when the condition in parentheses is met, indicating that the i-th category exists in the test sample x. Otherwise, it returns 0, indicating that the i-th category does not exist in the test sample.

[0114]

[0115] In formula (18), 1≤i≤C.

[0116] This embodiment also provides a multi-label classification device, which can be a hardware structure, a software module, or a hardware structure plus a software module. Specifically, the classification device in this embodiment includes a multi-label classification model construction unit for constructing a multi-label classification model; and a label prediction unit for inputting the sample to be tested into the constructed multi-label classification model to obtain the predicted value of the sample, and then determining whether a label category exists in the sample based on a classification threshold. The multi-label classification model construction unit includes:

[0117] The feature extraction module is used to extract image features from the training image set;

[0118] The semantic decoupling module is used to obtain a category-specific semantic representation vector by integrating the semantic information of each category in the image features extracted by S1;

[0119] The score prediction module is used to feed the obtained category-specific semantic representation vector into the linear classifier to calculate the predicted score for each label;

[0120] The image-level fusion module performs image-level fusion on the input model to obtain a fused image and calculates the prediction score of the fused image. Simultaneously, it fuses the prediction scores of the original images to obtain a fused prediction score. The prediction scores of the fused image and the fused prediction score are input into the loss function to obtain the regularization term L of the overall loss. reg ;

[0121] The adaptive temperature coefficient module is used to calculate the standard deviation of the predicted scores obtained in S3, and accordingly calculate the adaptive temperature coefficient τ. c ;

[0122] The classification loss term calculation module is used to input the predicted score of the original image and the label under the PU setting into the loss function to calculate the classification loss term;

[0123] The model optimization module is used to construct the overall loss function based on the calculated regularization term and classification loss term, thereby optimizing the model.

[0124] This embodiment also provides an electronic device, which includes a memory and a processor. The memory stores a program for the multi-label classification method of this embodiment, and the processor can call the program instructions stored in the memory to execute the multi-label classification method of this invention.

[0125] This embodiment also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the multi-label classification method of this embodiment, thereby achieving sample classification.

[0126] In summary, this invention utilizes PU learning technology to design a loss function that can achieve good classification accuracy by using only partially labeled positive and unknown labels during model training. Simultaneously, a rebalancing factor dynamically adjusts the contributions of positive and unknown label terms in the loss function, ensuring that the model optimization process considers the contributions of unknown labels without neglecting the contributions of positive labels. Furthermore, an adaptive temperature coefficient module is designed. By calculating the standard deviation of the predicted scores for each category, a category-specific temperature coefficient is obtained. This temperature coefficient is used to dynamically adjust the dispersion of predicted scores, alleviating the problem of overly smooth predicted scores in the early stages of model training, thereby accelerating model convergence and improving classification accuracy.

Claims

1. A method for constructing a multi-label classification model, characterized in that, Includes the following steps: S1. Input the training image set and extract image features; S2. Using the semantic decoupling module, a category-specific semantic representation vector is obtained by integrating the semantic information of each category in the image features extracted in S1. S3. Feed the obtained category-specific semantic representation vector into the linear classifier and calculate the prediction score for each label; S4. Perform image-level blending on the input model to obtain a blended image. Repeat steps S1-S3 to calculate the prediction score of the blended image. Simultaneously, blend the prediction scores of the original images to obtain a mixed prediction score. Input the prediction scores of the blended image and the mixed prediction score into the loss function to obtain the regularization term L of the overall loss. reg ; S5. Using the adaptive temperature coefficient module, calculate the standard deviation of the predicted score obtained in S3, and calculate the adaptive temperature coefficient τ accordingly to dynamically adjust the predicted score. S6. Input the predicted score of the original image and the label under the PU setting into the loss function to obtain the classification loss term L. var ; S7. Based on the regularization term L obtained in S4 reg And the classification loss term L obtained in S6 var We construct the overall loss function, and then optimize the model.

2. The method for constructing a multi-label classification model according to claim 1, characterized in that, The regularization term L for the overall loss in S4 reg The process of obtaining it is as follows: Step 1: Mix N images I according to coefficient μ to obtain N / 2 mixed images. Input the mixed images into the prediction model to obtain the prediction score of the mixed images. In formulas (3)-(4), the coefficient μ follows a Beta distribution with parameter β; Φ represents the prediction model; Represents the prediction score of the mixed image; and Let represent the first half and the second half of N images I, respectively; Step 2: Mix the prediction scores S of the N images I according to the coefficient μ to obtain the mixed prediction scores: In formula (5), the coefficient μ follows a Beta distribution with parameter β; and Let S represent the first half and the second half of the prediction scores S for N images I, respectively. Step 3, and The regularization term for the overall loss is obtained by feeding it into the mixed loss function: In formula (6), σ(·) represents the sigmoid activation function; express The predicted score of the c-th category in the n-th image; express The predicted score of the c-th category for the n-th image.

3. The method for constructing a multi-label classification model according to claim 2, characterized in that, S6 Classification loss term L var The calculation is as follows: In the above formula, Represents the set of unknown labels under category c in an image collection; Represents the combination of known labels under category c in an image set; s u Indicates that the label belongs to The corresponding predicted score; s p Indicates that the label belongs to The corresponding predicted score; P represents the average predicted probability of each category in an image set; the parameter γ represents the balance exponent, which controls the penalty for different proportions of known labels; γ This represents the rebalancing factor, used to adjust the contribution of unknown labels to the loss function.

4. The method for constructing a multi-label classification model according to claim 3, characterized in that, The process of constructing the overall loss function in step S7 is as follows: Step 1: Adjust the adaptive temperature coefficient τ in step S3. c Substituting into formula (15), for the classification loss term L... var Optimize: Where, τ c This represents the adaptive temperature coefficient for the c-th category; Step 2: Add L to formulas (6) and (15) reg and L var Substituting into the equation yields the overall loss function L. pu-mlc That is, we get: THE pu-mlc =L var +λL reg (17) In formula (17), the parameter λ is used to adjust the weight of the regularization term.

5. The method for constructing a multi-label classification model according to claim 4, characterized in that, The adaptive temperature coefficient τ is calculated as follows: In formula (7), Represents the temperature coefficients for category C; s represents the average predicted score for category c in N images; n,c This represents the predicted score of the c-th category in the n-th image of S; the parameter α represents the penalty coefficient, which controls the adjustment capability of the temperature coefficient.

6. A method for constructing a multi-label classification model according to any one of claims 1-5, characterized in that, In step S1, deep neural networks are used to extract image features from the training image set.

7. A multi-label classification method, characterized in that: The sample to be tested is input into the model constructed according to any one of claims 1-6 to obtain the predicted value of the sample, and then the presence of a labeled category in the sample is determined according to the classification threshold.

8. A multi-marker classification device, characterized in that, Includes a multi-label classification model building unit for constructing multi-label classification models; The system also includes a label prediction unit, used to input the sample to be tested into the constructed multi-label classification model, obtain the predicted value of the sample, and then determine whether there is a label category in the sample based on a classification threshold; wherein, the multi-label classification model construction unit includes: The feature extraction module is used to extract image features from the training image set; The semantic decoupling module is used to obtain a category-specific semantic representation vector by integrating the semantic information of each category in the image features extracted by S1; The score prediction module is used to feed the obtained category-specific semantic representation vector into the linear classifier to calculate the predicted score for each label; The image-level blending module performs image-level blending on the input model to obtain a blended image. Steps S1-S3 are repeated to calculate the prediction score of the blended image. Simultaneously, the prediction scores of the original images are blended to obtain a mixed prediction score. The prediction scores of the blended image and the mixed prediction score are input into the loss function to obtain the regularization term L of the overall loss. reg ; The adaptive temperature coefficient module is used to calculate the standard deviation of the predicted score obtained in S3, and to calculate the adaptive temperature coefficient τ accordingly. The classification loss term calculation module is used to input the predicted score of the original image and the label under the PU setting into the loss function to calculate the classification loss term; The model optimization module is used to construct the overall loss function based on the calculated regularization term and classification loss term, thereby optimizing the model.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the classification method described in claim 7.

10. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the classification method of claim 7.

Images