Fusion weakly supervised image classification method based on skew similarity and unmarked samples

Abstract

The application discloses a fusion weak supervision image classification method based on skew similarity and unmarked samples, belongs to the weak supervision image processing field, and comprises the following steps: constructing a data set containing sample pairs with skew similar confidence and unmarked sample points; according to the size and complexity of the data set, a classification neural network model is built, and the weight parameters and bias parameters of the classification neural network model are initialized; a skew similar confidence parameterization model is designed, and an empirical risk loss function is designed as a training standard of the classification neural network model; a gradient-based optimization algorithm is selected to minimize the empirical risk loss function; the classification neural network model is trained to obtain an optimal solution of the classification neural network model; and the trained classification neural network model is used to classify images to be labeled. The application overcomes the problems of inaccurate labeling, low information resource utilization efficiency and cost limitation, and can cope with complex data forms and solve diversified image classification tasks in the real world.

Description

Technical Field

[0001] This invention relates to the field of weakly supervised image processing, and in particular to a weakly supervised image classification method based on skew similarity and unlabeled samples. Background Technology

[0002] With the rapid advancement of technology, the volume of image data across various fields is increasing rapidly, making image management and classification more complex. Faced with complex data formats, effective information organization and retrieval are crucial. Against this backdrop, image classification technology has emerged as a solution, providing effective support for information organization and retrieval.

[0003] Image classification technology improves information organization and retrieval by labeling and categorizing unclassified images. This technology can effectively learn image features, enhancing work efficiency, accelerating decision-making processes, and improving decision quality across various industries. This positively impacts the overall competitiveness of industries and plays a vital role in driving technological and industrial development.

[0004] Traditional image classification methods primarily rely on supervised learning, employing machine learning techniques to learn and train on accurately labeled images. However, this approach is heavily dependent on accurate labeling information, leading to high time and labor costs. To overcome this difficulty of traditional supervised learning methods, researchers have proposed various weakly supervised models, such as "A Weakly Supervised Target Classification and Localization Method Based on Divergence Learning" (Chinese Patent Publication No. CN110689081A) and "A Weakly Supervised Whole-Slide Histopathological Image Classification Method Based on Contrast Learning" (Chinese Patent Publication No. CN113628197A). These models do not require accurate labeling of samples, thus effectively reducing time and labor costs.

[0005] In recent years, researchers have proposed many innovative weakly supervised learning architectures to address the increasingly complex data formats in the field of weak supervision. For example, Shinoda K, Kaji H, and Sugiyama M proposed a method for training a binary classifier based on skewed confidence of positive samples in their paper "Binary classification from positive data with skewed confidence" [J]. arXiv preprint arXiv:2001.10642,2020. On the other hand, Cao Y, Feng L, Xu Y, et al. introduced a method for training a binary classifier based on similarity confidence between paired samples in their paper "Learning from similarity-confidence data" [C] / / International Conference on Machine Learning.PMLR,2021:1272-1282. However, these methods have not solved the problem of resource waste caused by the presence of unlabeled sample points. Furthermore, the model proposed in the paper "Learning from similarity-confidence data" fails to account for situations where the estimated similarity confidence distribution differs significantly from the true confidence distribution, highlighting some challenges and limitations in the field of weakly supervised learning.

[0006] Image classification is a common task in many real-world applications. Take medical pathology as an example: hospitals and research institutions typically possess databases containing a large number of tissue slide images, including samples of normal tissue, benign lesions, and malignant lesions. Accurately classifying these tissue slide images is a critical task. In this context, paired tissue slides are considered two samples from the same patient. However, due to patient privacy concerns, some patients are unwilling to provide diagnostic information, making accurate annotation impossible for all samples. Nevertheless, annotators can still provide a portion of the similarity confidence between paired samples. Due to varying levels of expertise among annotators, there may be significant deviations between the similarity confidence and the true confidence distribution.

[0007] In image sample classification, two common scenarios are typically encountered: First, when estimating the similarity between two unlabeled samples, the lack of expertise among annotators leads to a significant deviation between the provided similarity confidence and the true confidence, resulting in a severe distortion of the estimated similarity confidence distribution compared to the true similarity confidence distribution. Second, the high time and manual annotation costs result in a large number of unlabeled samples in the data, and these valuable information resources are not fully utilized, leading to resource waste.

[0008] Therefore, it is necessary to develop a weakly supervised image classification method based on skew similarity and unlabeled samples to solve the above problems. Summary of the Invention

[0009] The technical problem to be solved by this invention is to provide a weakly supervised image classification method based on skewed similarity and unlabeled samples, which not only overcomes the problems of inaccurate labeling and cost limitations, but also provides a more accurate and efficient solution for medical pathology image classification.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0011] A weakly supervised image classification method based on the fusion of skew similarity and unlabeled samples, specifically including the following steps:

[0012] S1, Construct a dataset containing sample pairs with skewed similarity confidence and unlabeled sample points;

[0013] S2. Select an appropriate classification neural network model structure based on the size and complexity of the dataset, build the classification neural network model, and initialize the weight parameters and bias parameters of the classification neural network model.

[0014] S3. Design a skewed similarity confidence parameterized model, and design an empirical risk loss function based on the skewed similarity confidence parameterized model as the training standard for the classification neural network model;

[0015] S4, Select a gradient-based optimization algorithm to minimize the empirical risk loss function;

[0016] S5. Train the classification neural network model. Batch process the image dataset in step S1. Input each batch of data into the classification neural network model. Iteratively update the weight parameters and bias parameters of the classification neural network model through the optimization algorithm so that the empirical risk loss function converges to a local minimum, thereby obtaining the optimal solution of the classification neural network model.

[0017] S6: Obtain the trained classification neural network model and classify the image to be labeled.

[0018] A further improvement to the technical solution of this invention is that: in S1, the sample pairs only carry skewed similarity confidence information and do not have any category label information; the independent sample points also do not carry any category label information; the dataset of paired samples in the training sample set is defined as... Where n s The number of combinations of paired samples; the dataset of unlabeled independent point samples in the training sample set is n uThis represents the number of unlabeled point samples.

[0019] A further improvement of the technical solution of the present invention is that, in S2, the classification neural network model is any network model that has been trained with a large amount of image data and has been fully converged.

[0020] A further improvement to the technical solution of this invention lies in the following: In S3, the loss function in the empirical risk loss function model can be any loss function used for binary classification tasks; specifically, it includes the following steps:

[0021] S31 defines a skewed similarity confidence parameterized model designed to describe the skewed similarity confidence between pairs of samples.

[0022] The skew similarity confidence parameterization model can adopt various mathematical forms. When using an exponential model, the specific implementation method is as follows:

[0023] Define s(x,x′) as the exact similarity distribution between samples x and x′. The skewed similarity confidence distribution is defined; the skewed confidence parameterization model is defined as follows:

[0024]

[0025] Where 0 < k < ∞ are the hyperparameters of the skewed similarity confidence parameterization model;

[0026] S32 defines the false negative rate as prior knowledge and determines the optimal hyperparameter k in the skewed confidence parameterization model by minimizing the squared loss between the empirical classification error rate and the false negative rate. * ;

[0027] Specifically, the false negative rate Φ is defined as follows: In a binary classification problem, the proportion of samples that are actually positive but are incorrectly predicted as negative by the classification neural network model; the expression for the false negative rate is as follows:

[0028] Φ=∫{x:g(x)<0}p(x|y=+1)dx (9)

[0029] Where x represents an image sample; y∈{+1,-1} represents the set of sample labels; p(x|y=+1) represents the distribution of positive class samples with a class label of +1; g(x) represents the class label result of the output of the classification neural network model;

[0030] Define the empirical classification error rate:

[0031]

[0032] Among them, l 01(z)=(1-sign(z)) / 2 is a 0-1 loss function used to quantify the error between the empirical misclassification error rate and the actual false negative rate;

[0033] Given that the false negative rate is known prior knowledge, the selection criteria for the optimal hyperparameters in the skewed similarity confidence parameterization model, based on the false negative rate and empirical classification error rate, are as follows:

[0034]

[0035] Equation (4) represents the mean square error between the empirical classification error rate and the false negative rate Φ; therefore, the optimal hyperparameter k of the skew similarity confidence parameterization model is determined by minimizing equation (4). * ;

[0036] S33, Based on the defined skewed similarity confidence parameterized model, design an empirical risk loss function;

[0037] The empirical risk loss function based on sample pairs with skewed similarity confidence and unlabeled point samples is defined as follows:

[0038]

[0039] in, l(·) represents the loss function, used to quantify the error between the predicted output of the classification neural network model and the actual label during training; s k (x i ,x′ i )∈(0,1] and τ p τ represents the prior probability of a positive class sample; n n represents the prior probability of a negative class sample; c n represents the number of combinations of paired samples; u Represents the number of unlabeled sample points; It can select any loss function for binary classification tasks.

[0040] A further improvement to the technical solution of the present invention is that S5 specifically includes the following steps:

[0041] S51, Set the hyperparameters of the classification neural network model;

[0042] S52, batch processing is performed on the image dataset from step S1. A batch of images is passed to the neural network, and the weight parameters and bias parameters of the classification neural network model are continuously updated through optimization algorithms. The hyperparameter k and empirical risk loss function in the skew similarity confidence parameter model are also calculated.

[0043] S53: Optimize continuously through iteration, and determine whether the empirical risk loss function in step S33 has converged. If it has converged, that is, the empirical risk loss function no longer decreases significantly, stop training. Otherwise, if it has not converged, return to S51 and continue the training process by modifying the hyperparameters in the classification neural network model to achieve better training results.

[0044] A further improvement of the technical solution of the present invention is that the hyperparameters include at least the learning rate, batch size and number of iterations.

[0045] A further improvement to the technical solution of the present invention is that S52 specifically includes the following steps:

[0046] S521, each batch of samples is input into the classification neural network model, and the category result predicted by the current classification neural network model is calculated through forward propagation;

[0047] The input image data is linearly calculated using the weight and bias parameters of the classification neural network model to obtain intermediate results without activation function processing. For the input and hidden layers, the calculation formulas are as follows:

[0048] z = xW + b (13)

[0049] Where x represents an image sample, W represents a weight parameter, b represents a bias parameter, and z represents an intermediate result; the intermediate result obtained by linear calculation is transformed nonlinearly through an activation function, and the output layer of the classification neural network model maps the output value to the range [0,1] through the activation function, representing the probability that the sample belongs to the positive class; a threshold of 0.5 is set, if the output value of the classification neural network model is greater than or equal to 0.5, the sample is classified as positive and the class label is +1; if it is less than 0.5, the sample is classified as negative and the class label is -1;

[0050] S522, using the category label results output by the classification neural network model, calculate the current optimal hyperparameter k in the skew similarity confidence parameter model by minimizing equation (4) in step S32. * ;

[0051] S523, the current optimal hyperparameter k * Substituting into equation (5) in step S33, the value of the empirical risk loss function in equation (5) is calculated using the category label results output by the classification neural network model; the formula is as follows:

[0052]

[0053] S524, the optimization algorithm calculates the gradient information of the loss function and backpropagates along the negative gradient direction of the loss function to update the weight parameters and bias parameters of the classification neural network model, thereby gradually reducing the empirical risk loss function.

[0054] The technological advancements achieved by this invention due to the adoption of the above technical solutions are as follows:

[0055] 1. Considering the skewed distribution of similarity confidence between paired samples and the inability to effectively utilize unlabeled samples in practice, this invention introduces a parameterized model to describe the skewed similarity confidence by comprehensively utilizing the skewed similarity confidence information and the data distribution characteristics of unlabeled samples, and defines an empirical risk loss function to adjust the model. By optimizing the empirical risk loss function, efficient and accurate discrimination of the categories of the image samples to be labeled is achieved, improving the model's ability to deal with complex samples.

[0056] 2. This invention defines an empirical risk loss function adjustment model by comprehensively utilizing the skewed similarity confidence distribution and the distribution of unlabeled samples. This model better adapts to the skewed similarity confidence distribution that differs significantly from the true distribution, alleviates the impact of skewed similarity confidence on training the classification model, and improves robustness to skewed similarity confidence distribution.

[0057] 3. This invention also fully utilizes the additional attribute information carried by unlabeled samples and incorporates it into the training process. By comprehensively utilizing the data distribution characteristics of unlabeled samples, the efficiency of the model in utilizing unlabeled samples is further improved. This enables the model to understand the data more comprehensively, enhances the classifier's ability to process complex samples, and allows it to better adapt to the data during the training phase, thereby achieving efficient and accurate image sample classification.

[0058] 4. This invention provides a more economical and efficient solution for solving diverse image classification tasks in the real world, and provides strong support for reducing manual annotation costs and improving training efficiency in practical applications. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This illustration shows a weakly supervised image classification method based on paired samples and unlabeled samples with accurate similarity confidence, provided by an embodiment of the present invention.

[0061] Figure 2 The illustration shows a weakly supervised image classification method based on paired samples and unlabeled samples with skewed similarity confidence, provided for an embodiment of the present invention.

[0062] Where ○ represents the positive class sample to be labeled, △ represents the negative class sample to be labeled, and the red line represents the similarity between two samples. The darker the color, the higher the similarity between the two samples.

[0063] Figure 3 This is a flowchart illustrating the weakly supervised image classification method based on skew similarity and unlabeled samples provided in an embodiment of the present invention.

[0064] Figure 4 This is a flowchart of the training process of the classification neural network model in an embodiment of the present invention. Detailed Implementation

[0065] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products or devices.

[0066] This invention aims to train a binary classifier from sample pairs carrying skewed similarity confidence information and unlabeled samples. For samples existing in the form of data pairs, in practical applications, some privacy issues arise; annotators may find it difficult to obtain accurate sample labels, but they can provide the degree of similarity between two samples. Due to differences in the professional level of annotators, the similarity confidence distribution labeled by annotators may differ somewhat from the true similarity confidence distribution. For example... Figure 1 The true similarity confidence distribution shown can accurately reflect the degree of similarity between samples, while Figure 2 The skewed similarity confidence distribution shown in the diagram has a significant error compared to the true situation. To address this issue, this invention introduces a parameterized model to describe the skewed similarity confidence and defines an empirical risk function adjustment model to mitigate the impact of skewed similarity confidence. Furthermore, this invention fully utilizes the additional attribute information carried by unlabeled samples and incorporates it into the training process, thereby increasing the model's generalization ability.

[0067] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:

[0068] like Figure 3As shown, this invention provides a weakly supervised image classification method based on skew similarity and unlabeled samples, specifically including the following steps:

[0069] S1, obtain sample pairs with skew similarity confidence and unlabeled independent sample points, and construct a dataset containing sample pairs with skew similarity confidence and unlabeled sample points;

[0070] Specifically, the dataset of paired samples in the training sample set is defined as follows: Where n s The number of combinations of paired samples; the dataset of unlabeled independent point samples in the training sample set is n u This represents the number of unlabeled point samples.

[0071] S2. Select an appropriate classification neural network model structure based on the size and complexity of the dataset, build a classification neural network model including an input layer, an output layer, and hidden layers, and initialize the weight parameters and bias parameters of the classification neural network model using a random initialization method.

[0072] Optionally, the classification neural network model can be any neural network model that has been trained on a large amount of image data and has converged sufficiently. Commonly used image processing models such as MLP, ResNet, and CNN can all be used as classification neural network models. To ensure classification accuracy, multiple models can be selected based on the actual scenario (training time cost, available computing power), and the classification results can be compared comprehensively to determine the label category.

[0073] S3. Design a skewed similarity confidence parameterized model, and design an empirical risk loss function based on the skewed similarity confidence parameterized model as the training standard for the classification neural network model. The specific implementation method of this step is as follows;

[0074] S31, Define a skewed similarity confidence parameterization model. This model aims to describe the skewed similarity confidence between pairs of samples. This skewed similarity confidence parameterization model can adopt various mathematical forms, such as exponential models, additive models, etc. The specific implementation is as follows;

[0075] In this invention, similarity confidence refers to the degree of similarity between two samples in a set of data.

[0076] Specifically, s(x,x′) represents the exact similarity distribution between samples x and x′. The similarity confidence distribution is skewed.

[0077] In a preferred embodiment of the present invention, the skew confidence parameterization model is defined as follows:

[0078]

[0079] Where 0 < k < ∞ are the hyperparameters of the skewed similarity confidence parameterization model.

[0080] It should be understood here that the exponential skew confidence parameterization model is used to ensure the flatness of the skew similarity confidence distribution. Any other model, such as the additive model max(0,min(1+k)),-1≤k≤1, can be used as an embodiment of the present invention.

[0081] S32 defines the false negative rate as prior knowledge and determines the optimal hyperparameter k in the skewed confidence parameterization model by minimizing the squared loss between the empirical classification error rate and the false negative rate. * .

[0082] Specifically, the false negative rate Φ is defined as: in a binary classification problem, the proportion of samples that are actually positive but are incorrectly predicted as negative by the classification neural network model. The expression for the false negative rate is as follows:

[0083] Φ=∫{x:g(x)<0}p(x|y=+1)dx (16)

[0084] Where x represents an image sample, y∈{+1,-1} represents the set of sample labels, p(x|y=+1) represents the distribution of positive class (category label is +1) samples, and g(x) represents the category label result of the output of the classification neural network model.

[0085] Define the empirical classification error rate:

[0086]

[0087] Among them, l 01 (z) = (1 - sign(z)) / 2 is a 0-1 loss function used to quantify the error between the empirical misclassification error rate and the actual false negative rate.

[0088] This invention requires the false negative rate as known prior knowledge. Based on the false negative rate and the empirical classification error rate, the selection criteria for the optimal hyperparameter in the skewed similarity confidence parameterization model are as follows:

[0089]

[0090] Equation (4) represents the mean square error between the empirical classification error rate and the false negative rate Φ. Therefore, the optimal hyperparameter k of the skewed similarity confidence parameterization model can be determined by minimizing equation (4). * .

[0091] S33. Based on the defined skewed similarity confidence parameterized model, design an empirical risk loss function.

[0092] Specifically, the empirical risk loss function based on sample pairs with skewed similarity confidence and unlabeled point samples is defined as follows:

[0093]

[0094] in, l(·) represents the loss function, used to quantify the error between the predicted output of the classification neural network model and the actual label during training; s k (x i ,x′ i )∈(0,1] and τ p τ represents the prior probability of a positive class sample; n n represents the prior probability of a negative class sample; c n represents the number of combinations of paired samples; u This represents the number of unlabeled sample points.

[0095] Optionally, You can choose any loss function for the binary classification task. This includes, but is not limited to, some common loss functions such as: Cross-Entropy Loss, Hinge Loss, Double Hinge Loss, Mean SquaredError, and Logistic Loss.

[0096] S4. Select a suitable gradient-based optimization algorithm to minimize the empirical risk loss function;

[0097] Alternatively, common optimization algorithms include SGD, Adam, and Adagrad.

[0098] S5, train the classification neural network model, such as Figure 4 As shown, the image dataset in step S1 is batch-processed, with each batch of data input into the classification neural network model. The weight and bias parameters of the classification neural network model are iteratively updated using an optimization algorithm until the empirical risk loss function converges to a local minimum, thus obtaining the optimal solution for the classification neural network model. The specific steps are as follows:

[0099] S51, set the hyperparameters of the classification neural network model, such as learning rate, batch size, and number of iterations;

[0100] S52, batch processing is performed on the image dataset from step S1. A batch of images is passed to the neural network, and the weight parameters and bias parameters of the classification neural network model are continuously updated through optimization algorithms. The hyperparameter k and empirical risk loss function in the skew similarity confidence parameter model are calculated. The specific implementation steps are as follows:

[0101] S521, each batch of samples is input into the classification neural network model, and the category result predicted by the current classification neural network model is calculated through forward propagation;

[0102] Specifically, the input image data is linearly calculated using the weight and bias parameters of the classification neural network model to obtain intermediate results without activation function processing. For the input layer and hidden layers, the calculation formula is as follows:

[0103] z = xW + b (20)

[0104] Where x represents an image sample, W represents a weight parameter, b represents a bias parameter, and z represents an intermediate result; the intermediate result obtained by linear calculation is transformed nonlinearly through an activation function, such as Sigmoid, Tanh, ReLU, etc.; the output layer of the classification neural network model maps the output value to the range [0,1] through the Sigmoid function, representing the probability that the sample belongs to the positive class; a threshold of 0.5 is set, if the output value of the classification neural network model is greater than or equal to 0.5, the sample is classified as positive (class label is +1), and if it is less than 0.5, the sample is classified as negative (class label is -1);

[0105] S522, using the category label results output by the classification neural network model, calculate the current optimal hyperparameter k in the skew similarity confidence parameter model by minimizing equation (4) in step S32. * ;

[0106] S523, the current optimal hyperparameter k * Substitute into equation (5) in step S33 and use the category label results output by the classification neural network model to calculate the value of the empirical risk loss function in equation (5);

[0107]

[0108] S524, the optimization algorithm calculates the gradient information of the loss function and backpropagates along the negative gradient direction of the loss function to update the weight parameters and bias parameters of the classification neural network model, thereby gradually reducing the empirical risk loss function.

[0109] In step S53, the system iteratively optimizes the algorithm and determines whether the empirical risk loss function from step S33 has converged. If it has converged (meaning the empirical risk loss function no longer decreases significantly), training stops. Otherwise, if it has not converged, the process returns to step S51, where hyperparameters such as the learning rate, batch size, and number of iterations in the classification neural network model are modified to continue training the network and achieve better training results.

[0110] S6: Obtain the trained classification neural network model and classify the image to be labeled.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A weakly supervised image classification method based on the fusion of skew similarity and unlabeled samples, characterized in that, Specifically, the following steps are included: S1, Construct a dataset containing sample pairs with skewed similarity confidence and unlabeled sample points; The sample pairs only carry skewed similarity confidence information and do not have any category label information; independent sample points also do not carry any category label information; the dataset of paired samples in the training sample set is defined as... ,in The number of combinations of paired samples; the dataset of unlabeled independent point samples in the training sample set is , Represents the number of unlabeled point samples; S2. Select an appropriate classification neural network model structure based on the size and complexity of the dataset, build the classification neural network model, and initialize the weight parameters and bias parameters of the classification neural network model. S3. Design a skewed similarity confidence parameterized model, and design an empirical risk loss function based on the skewed similarity confidence parameterized model as the training standard for the classification neural network model; The loss function in the empirical risk loss function can be any loss function used for binary classification tasks; specifically, it includes the following steps: S31 defines a skewed similarity confidence parameterized model designed to describe the skewed similarity confidence between pairs of samples. The skew similarity confidence parameterization model can adopt various mathematical forms. When using an exponential model, the specific implementation method is as follows: definition Representative sample and The accurate distribution of similarity between them The skewed similarity confidence distribution is defined; the skewed confidence parameterization model is defined as follows: in, It is a hyperparameter of the skewed similarity confidence parameterization model; S32 defines the false negative rate as prior knowledge and determines the optimal hyperparameters in the skewed confidence parameterization model by minimizing the squared loss between the empirical classification error rate and the false negative rate. ; Specifically, define the false negative rate. In binary classification problems, the false negative rate is the proportion of samples that are actually positive but are incorrectly predicted as negative by the classification neural network model; the expression for the false negative rate is as follows: in, Represents image samples; Represents the set of sample labels; Represents the distribution of positive class samples, with class labels as follows: ; The category label results representing the output of the classification neural network model; Define the empirical classification error rate: in, It is a 0-1 loss function used to quantify the error between the empirical misclassification error rate and the actual false negative rate; Given that the false negative rate is known prior knowledge, the selection criteria for the optimal hyperparameters in the skewed similarity confidence parameterization model, based on the false negative rate and empirical classification error rate, are as follows: Formula (4) represents the empirical classification error rate and the false negative rate. The mean square error between them; therefore, the optimal hyperparameters of the skew similarity confidence parameterization model are determined by minimizing equation (4). ; S33, Based on the defined skewed similarity confidence parameterized model, design an empirical risk loss function; The empirical risk loss function based on sample pairs with skewed similarity confidence and unlabeled point samples is defined as follows: in, ; This represents the loss function, used to quantify the error between the predicted output of the classification neural network model and the actual label during training; and ; Represents the prior probability of a positive class sample; Represents the prior probability of a negative class sample; This represents the number of unlabeled sample points; It can choose any loss function for binary classification tasks; S4, Select a gradient-based optimization algorithm to minimize the empirical risk loss function; S5. Train the classification neural network model. Batch process the image dataset in step S1. Input each batch of data into the classification neural network model. Iteratively update the weight parameters and bias parameters of the classification neural network model through the optimization algorithm so that the empirical risk loss function converges to a local minimum, thereby obtaining the optimal solution of the classification neural network model. S6: Obtain the trained classification neural network model and classify the image to be labeled.

2. The weakly supervised image classification method based on skew similarity and unlabeled samples according to claim 1, characterized in that, In S2, the classification neural network model is any network model that has been trained on a large amount of image data and has converged sufficiently.

3. The weakly supervised image classification method based on skew similarity and unlabeled samples according to claim 1, characterized in that, In S5, the specific steps include: S51, Set the hyperparameters of the classification neural network model; S52, batch processing is performed on the image dataset from step S1. A batch of images is passed to the neural network, and the weight parameters and bias parameters of the classification neural network model are continuously updated through optimization algorithms. The hyperparameters in the skew similarity confidence parameter model are also calculated. and empirical risk loss function, S53: Optimize continuously through iteration, and determine whether the empirical risk loss function in step S33 has converged. If it has converged, that is, the empirical risk loss function no longer decreases significantly, stop training. Otherwise, if it has not converged, return to S51, modify the hyperparameters in the classification neural network model, and continue to implement the training process to achieve better training results.

4. The weakly supervised image classification method based on skew similarity and unlabeled samples according to claim 3, characterized in that, The hyperparameters include at least the learning rate, batch size, and number of iterations.

5. The weakly supervised image classification method based on skew similarity and unlabeled samples according to claim 3, characterized in that, S52 specifically includes the following steps: S521, each batch of samples is input into the classification neural network model, and the category result predicted by the current classification neural network model is calculated through forward propagation; The input image data is linearly calculated using the weight and bias parameters of the classification neural network model to obtain intermediate results without activation function processing. For the input and hidden layers, the calculation formulas are as follows: in, Represents image samples, Represents the weight parameter. Represents the bias parameter. This represents an intermediate result; the intermediate result obtained by linear calculation is transformed nonlinearly through an activation function. The output layer of the classification neural network model maps the output value to the range [0,1] through the activation function, representing the probability that the sample belongs to the positive class; a threshold of 0.5 is set. If the output value of the classification neural network model is greater than or equal to 0.5, the sample is classified as positive and the class label is +1; if it is less than 0.5, the sample is classified as negative and the class label is -1. S522, using the category label results output by the classification neural network model, calculate the current optimal hyperparameter in the skew similarity confidence parameter model by minimizing equation (4) in step S32. ; S523, the current optimal hyperparameters Substituting into equation (5) in step S33, the value of the empirical risk loss function in equation (5) is calculated using the category label results output by the classification neural network model; the formula is as follows: S524, the optimization algorithm calculates the gradient information of the loss function and backpropagates along the negative gradient direction of the loss function to update the weight parameters and bias parameters of the classification neural network model, thereby gradually reducing the empirical risk loss function.

Images