Specific image classification method based on sequence relation modeling
By using a specific image classification method based on sequence relationship modeling and leveraging global sample aggregation and dual-view constraint modules, this method addresses the issues of low robustness and accuracy in traditional image classification methods, achieving highly robust and interpretable image stage recognition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NINGBO UNIV
- Filing Date
- 2026-02-06
- Publication Date
- 2026-05-19
AI Technical Summary
Traditional image classification methods ignore the natural sequential relationships between stages when dealing with image sequences that change over time or in structural states, resulting in low robustness, accuracy, and interpretability.
A specific image classification method based on sequence relationship modeling is adopted. Through a global sample aggregation module and a dual-view sequence relationship constraint module, KNN module and confidence prediction module are used for sample classification. The distance-based sequence relationship constraint and probability unimodal constraint sub-modules ensure that the sample distance follows the natural stage progression order. Cross-entropy, sequence relationship and unimodal loss are calculated to update the model parameters.
It significantly improves the robustness, accuracy, and interpretability of recognition at specific image stages, ensuring that the predicted probability distribution presents a single peak at the real category location, and the probability decreases with increasing distance from the real label, exhibiting obvious step-like characteristics.
Smart Images
Figure CN122067015A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of specific image processing and intelligent recognition, and in particular to a method for specific image classification based on sequence relationship modeling. Background Technology
[0002] In specific images, such as medical images (CT images, MRI images, ultrasound images), material tissue images, and biological tissue images, there are often image sequences that exhibit gradual changes over time, structural state, or tissue changes. For example, a certain tissue region may show a series of stages from mild to severe changes (e.g., mild → moderate → severe). However, traditional image classification methods make independent predictions based on single images, ignoring the natural sequential relationships between stages and the characteristic neighborhood structure between global samples. Furthermore, the difference in magnitude between adjacent stages and across stages varies, and the output probability should reflect a unimodal pattern of decreasing or increasing stages. Thus, the results of such image-specific stage recognition methods are robust, accurate, and have low interpretability. Summary of the Invention
[0003] The technical problem to be solved by the present invention is to provide a specific image classification method based on sequence relationship modeling that is robust, accurate and interpretable, in light of the above-mentioned existing technology.
[0004] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a specific image classification method based on sequence relationship modeling, characterized by comprising the following steps:
[0005] Step 1: Obtain specific image data as a sample set, label the sample set according to the natural stage order, and then divide the sample set into a training set, a validation set, and a test set;
[0006] Step 2: Establish a specific image classification model, which includes a global sample aggregation module and a dual-view sequence relationship constraint module; map the input sample i into embedded features through encoder E. The global sample aggregation module includes a classifier, a KNN module, and a confidence prediction module based on global neighbors, which embeds features. The stage prediction probability of sample i is obtained through the classifier. Simultaneously embedding features The input is the KNN module, which obtains the K nearest neighbors of sample i; the confidence prediction module based on global neighbors predicts the stage probability of sample i. The group probability is obtained by aggregating the stage prediction probabilities of the K nearest neighbors of sample i. Then, based on the confidence threshold, sample i is divided into easy sample, hard sample, or noisy sample, and the cross-entropy classification loss for easy and hard samples is calculated. The dual-view sequence relationship constraint module includes a distance-based sequence relationship constraint submodule and a probability unimodal constraint submodule. The distance-based sequence relationship constraint submodule applies constraints to easy samples and calculates the distance-based sequence relationship loss. This ensures that the distances between easy samples follow the natural progression order; the probability unimodal constraint submodule applies constraints to both easy and difficult samples and calculates the unimodal loss. To ensure the highest probability of the true class, and that the probability monotonically decreases along the sequence direction; then, the cross-entropy classification loss is used. Distance-based sequence relation loss and single-peak loss Add them together to get the final total loss. Finally, using the total final loss Update the parameters of the network model to obtain a specific image classification model after one training cycle;
[0007] Step 3: Use the training set and validation set to train and validate the specific image classification model to obtain the trained specific image classification model;
[0008] Step 4: Input the specific image to be tested from the test set into the trained specific image classification model to obtain the image classification result of the specific image to be tested.
[0009] Preferably, the set of the K nearest neighbor samples of sample i obtained by the KNN module is denoted as ,satisfy:
[0010] And j≠i
[0011] Where t represents the current training period, t-1 represents the previous training period, and sample j represents samples other than i and j. Any sample other than Let i be the embedding features of sample i in the previous training cycle. For the sample In the embedding features of the previous training cycle, ... For the sample The embedded features from the previous training cycle, For sample j, the embedding features in the previous training cycle, Let K be the set of K nearest neighbor samples corresponding to sample i in the feature space, where K is the number of nearest neighbors; This represents the Euclidean distance operation for extracting features.
[0012] Furthermore, the step of the confidence prediction module based on global neighbors to obtain group probabilities is as follows:
[0013] Step a, the weight adjustment formula for obtaining the K nearest neighbors of sample i is:
[0014]
[0015] in, Let k be the weight of sample k with respect to sample i in the current training period, where k = 1, 2, ..., K. For sample i, the embedding features in the current training period For the embedding features of sample j in the current training period, For sample k, the embedding features in the current training period This is a temperature coefficient used to adjust the smoothness of the weight distribution;
[0016] Step b: Obtain group probabilities :
[0017]
[0018] in, This represents the group probability of sample i in the current training period. This represents the predicted probability of sample i at the current stage of the training cycle. This represents the group probability of sample k in the previous training period.
[0019] Furthermore, the confidence threshold includes a first sample threshold th1 and a second sample threshold th2, and the confidence prediction module based on global neighbors determines... The relationship with th1 and th2, if That is, corresponding to its actual stage The probability is greater than If sample i is a common sample, then it is determined that sample i is a common sample; if the previous condition is not met, then... That is, except for the true phase, the maximum probability for all other phases does not exceed If sample i is a difficult sample, then it is considered a difficult sample; otherwise, it is considered a noisy sample.
[0020] Preferably, in step 1, the sample set is labeled according to the natural stage order as follows:
[0021]
[0022] in, , , These correspond to the mild, moderate, and severe stages of structural changes in a specific image, respectively.
[0023] In step 2, sample i is mapped to embedded features. The formula is:
[0024]
[0025] Where E represents the encoder;
[0026] In the global sample aggregation module, the classifier obtains the stage prediction probability of sample i. Obtained in the following way:
[0027]
[0028] in, For sample i to belong to The predicted probability of the stage For sample i to belong to The predicted probability of the stage For sample i to belong to The predicted probability of the stage ; This is the stage prediction probability vector.
[0029] Furthermore, in the dual-view sequence relationship constraint module, the above-mentioned sequences are selected from the training set. One sample from each of the three stages, and data augmentation is performed on each sample to obtain six specific images: Stage Including original samples and enhanced samples ;stage Including original samples and enhanced samples ;stage Including original samples and enhanced samples Input the above six samples into encoder E(⋅) to obtain feature vectors, and define the feature distance as Euclidean distance:
[0030]
[0031] in, For the sample and Euclidean distance in feature space; m=1, 2, 3, 4, 5, 6; n=1, 2, 3, 4, 5, 6;
[0032] In the distance-based sequence relation constraint submodule, three distance subsets are constructed. , , For the distance subset Construction:
[0033]
[0034] For the distance subset Construction:
[0035]
[0036] For the distance subset Construction:
[0037]
[0038] Define the average distance of the three distance subsets above:
[0039]
[0040] in, Represents a distance subset The average distance, denoted as ;
[0041] Construct the average distance sorting vector :
[0042]
[0043] right Constructing the average sorting loss :
[0044]
[0045] Define the maximum-minimum distance relationship for the three distance subsets mentioned above:
[0046]
[0047]
[0048]
[0049]
[0050] in, The distance between two samples in the same period is denoted as . ; The closest distance between two samples in adjacent stages is denoted as . ; The distance between two samples in adjacent stages is denoted as . ; The closest distance between two samples across stages is denoted as . ;
[0051] Construct the maximum-minimum distance sorting vector :
[0052]
[0053] right Constructing the maximum-minimum sorting loss :
[0054]
[0055] Final distance order relation loss for:
[0056]
[0057] Preferably, in the probabilistic unimodal constraint submodule, a unimodal loss is constructed using the difference between adjacent stages:
[0058]
[0059] in:
[0060]
[0061] For sample i, the loss is unimodal; These represent the model's relationship to adjacent stages. The predicted probability; This indicates the true stage label of the sample; and This indicates the sequence relationship between stages. Representation phase Lighter than stage , Representation phase Equal to or more serious than stage .
[0062] Furthermore, the final total loss The calculation formula is:
[0063]
[0064] in, This is a hyperparameter used to balance the weights of different loss terms.
[0065] Preferably, in step 4, the formula for obtaining the classification result of the specific image to be tested is:
[0066]
[0067] in, For the specific image to be tested, The classification result for the specific image to be tested. Indicates will Input encoder, Indicates will Input classifier.
[0068] Compared with existing technologies, the advantages of this invention are as follows: In this specific image classification method based on sequence relationship modeling, the distance-based sequence relationship constraint submodule imposes constraints on easy samples and calculates distance-based sequence relationship loss, ensuring that the distance between easy samples follows the natural stage progression order; the probability unimodal constraint submodule imposes constraints on easy and difficult samples respectively and calculates unimodal loss, ensuring that the probability of the true class is the highest and monotonically decays along the sequence direction; such unimodal loss ensures that the predicted probability distribution presents a single peak, and this peak appears at the position of the true class label. As the distance from the true label increases, the probability of that class gradually decreases, thereby ensuring the stepwise characteristics of stage progression and significantly improving the robustness, accuracy, and interpretability of specific image stage recognition. Attached Figure Description
[0069] Figure 1 This is an architecture diagram of a specific image classification method based on sequence relationship modeling in an embodiment of the present invention. Detailed Implementation
[0070] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.
[0071] Figure 1 shows a preferred embodiment of the present invention. This embodiment provides a specific image classification method based on sequence relationship modeling, characterized by the following steps:
[0072] Step 1: Obtain specific image data as a sample set, label the sample set according to the natural stage order, and then divide the sample set into training set, validation set and test set;
[0073] Step 2: Establish a specific image classification model, which includes a global sample aggregation module and a dual-view sequence relationship constraint module; map the input sample i into embedded features through encoder E. The global sample aggregation module includes a classifier, a KNN module, and a confidence prediction module based on global neighbors, which embeds features. The stage prediction probability of sample i is obtained through the classifier. Simultaneously embedding features The input is the KNN module, which obtains the K nearest neighbors of sample i; the global neighbor-based confidence prediction module predicts the stage probability of sample i. The group probability is obtained by aggregating the stage prediction probabilities of the K nearest neighbors of sample i. Then, based on the confidence threshold, sample i is divided into easy sample, hard sample, or noisy sample, and the cross-entropy classification loss for easy and hard samples is calculated. The method for calculating the cross-entropy classification loss is existing technology and will not be described further. The dual-view sequence relationship constraint module includes a distance-based sequence relationship constraint submodule and a probability unimodal constraint submodule. The distance-based sequence relationship constraint submodule imposes constraints on easy samples and calculates the distance-based sequence relationship loss. This ensures that the distances between easy samples follow the natural progression order; the probabilistic unimodal constraint submodule applies constraints to easy and difficult samples respectively, and calculates the unimodal loss. To ensure the highest probability of the true class, and that the probability monotonically decreases along the sequence direction; then, the cross-entropy classification loss is used. Distance-based sequence relation loss and single-peak loss Add them together to get the final total loss. Finally, using the total final loss Update the parameters of the network model to obtain a specific image classification model after one training cycle;
[0074] Step 3: Use the training set and validation set to train and validate the specific image classification model to obtain the trained specific image classification model;
[0075] Step 4: Input the specific image to be tested from the test set into the trained specific image classification model to obtain the image classification result of the specific image to be tested.
[0076] In step 1, the sample set is labeled according to the natural stage order as follows:
[0077]
[0078] in, , , These correspond to mild, moderate, and severe stages of structural changes in specific images, for example, in medical images, NC (Normal control) < MCI (Mild Cognitive Impairment) < AD (Alzheimer's disease).
[0079] In step 2, sample i is mapped to embedded features. The formula is:
[0080]
[0081] Where E represents the encoder;
[0082] Subsequently, the classifier in the global sample aggregation module obtains the stage prediction probability of sample i. Obtained in the following way:
[0083]
[0084] in, For sample i to belong to The predicted probability of the stage For sample i to belong to The predicted probability of the stage For sample i to belong to The predicted probability of the stage ; This is the stage prediction probability vector.
[0085] Let the set of the K nearest neighbors of sample i obtained by the KNN module be denoted as . ,satisfy:
[0086] And j≠i
[0087] Where t represents the current training period, t-1 represents the previous training period, and sample j represents samples other than i and j. Any sample other than Let i be the embedding features of sample i in the previous training cycle. For the sample In the embedding features of the previous training cycle, ... For the sample The embedded features from the previous training cycle, For sample j, the embedding features in the previous training cycle, Let K be the set of K nearest neighbor samples corresponding to sample i in the feature space, where K is the number of nearest neighbors; This represents the Euclidean distance operation for extracting features.
[0088] In this embodiment, the steps for the confidence prediction module based on global neighbors to obtain group probabilities are as follows:
[0089] Step a, the weight adjustment formula for obtaining the K nearest neighbors of sample i is:
[0090]
[0091] in, Let k be the weight of sample k with respect to sample i in the current training period, where k = 1, 2, ..., K. For sample i, the embedding features in the current training period For the embedding features of sample j in the current training period, For sample k, the embedding features in the current training period This is a temperature coefficient used to adjust the smoothness of the weight distribution;
[0092] Step b: Obtain group probabilities :
[0093]
[0094] in, This represents the group probability of sample i in the current training period. This represents the predicted probability of sample i at the current stage of the training cycle. This represents the group probability of sample k in the previous training period.
[0095] The confidence thresholds include a first sample threshold th1 and a second sample threshold th2, which are determined by the global neighbor-based confidence prediction module. The relationship with th1 and th2, if That is, corresponding to its actual stage The probability is greater than If sample i is a common sample, then it is determined that sample i is a common sample; if the previous condition is not met, then... That is, except for the true phase, the maximum probability for all other phases does not exceed If sample i is a difficult sample, then it is considered a difficult sample; otherwise, it is considered a noisy sample.
[0096] In the dual-view sequence relation constraint module, the distance-based sequence relation constraint submodule selects the above from the training set. One sample from each of the three stages, and data augmentation is performed on each sample to obtain six specific images: Stage Including original samples and enhanced samples ;stage Including original samples and enhanced samples ;stage Including original samples and enhanced samples Input the above six samples into encoder E(⋅) to obtain feature vectors, and define the feature distance as Euclidean distance:
[0097]
[0098] in, For the sample and Euclidean distance in feature space; m=1, 2, 3, 4, 5, 6; n=1, 2, 3, 4, 5, 6;
[0099] Therefore, we obtain a 6×6 symmetric distance matrix:
[0100]
[0101] The diagonal elements of this matrix are 0 (distance within the same image is 0), and the upper and lower triangles are symmetrical.
[0102] Construct three distance subsets , , For distance subset The construction is based on two samples from the same stage:
[0103]
[0104] For distance subset The structure consists of two samples from adjacent stages:
[0105]
[0106] For distance subset The construction consists of two samples from non-adjacent stages:
[0107]
[0108] Define the average distance of the three distance subsets above:
[0109]
[0110] in, Represents a distance subset The average distance, denoted as ;
[0111] The ideal relationship is:
[0112]
[0113] Construct the average distance sorting vector :
[0114]
[0115] Define the maximum-minimum distance relationship for the three distance subsets mentioned above:
[0116]
[0117]
[0118]
[0119]
[0120] in, This represents the furthest distance between two samples within the same phase, and cannot be farther than any pair of adjacent phases; it is denoted as . ; The closest distance between two samples in adjacent stages is denoted as . ; The distance between two samples in adjacent stages is denoted as . ; This indicates that the closest distance between two samples across stages is also farther than any pair of adjacent stages, denoted as . ;
[0121] The ideal order relation is:
[0122]
[0123] Construct the maximum-minimum distance sorting vector :
[0124]
[0125] To simultaneously constrain the maximum-minimum and average relationships, this embodiment employs the Plackett-Luce (PL) sorting model, transforming the distance relationships above into an "ordered list" for modeling. For any sorted vector... The probability that it is considered by the model to be "arranged in ascending order" is defined as:
[0126]
[0127] in, The probability that the sorting is considered by the model to be "arranged in ascending order";
[0128] Then for Constructing the average sorting loss :
[0129]
[0130] get:
[0131]
[0132] right Constructing the maximum-minimum sorting loss :
[0133]
[0134] get:
[0135] Final distance order relation loss for:
[0136]
[0137] This is used to maintain the sequential structure of the sequence, thus ensuring that the worst-case scenario does not go out of bounds while also ensuring that the overall trend is correct.
[0138] In the probabilistic unimodal constraint submodule, the unimodal loss is constructed using the difference between adjacent stages:
[0139]
[0140] in:
[0141]
[0142] For sample i, the loss is unimodal; These represent the model's relationship to adjacent stages. The predicted probability; This indicates the true stage label of the sample; and This indicates the sequence relationship between stages. Representation phase Lighter than stage , Representation phase Equal to or more serious than stage .
[0143] Ideally, for The probability should increase as the stage approaches the true stage; for The probability should decrease as the stage moves further away from the true stage.
[0144] This is used to transform predicted probabilities into a unimodal distribution centered on the true class label, penalizing any predicted probability distribution that violates the unimodal structure. When the true label is an intermediate class, the true class probability is the highest; when the true label is a boundary class, the predicted probability changes monotonically along the stage order, ensuring that the probability of adjacent classes is no higher than that of the true class, and the probability of non-adjacent classes is no higher than that of adjacent classes.
[0145] Final total loss The calculation formula is:
[0146]
[0147] in, This is a hyperparameter used to balance the weights of different loss terms.
[0148] In step 4, the formula for obtaining the classification result of the specific image to be tested is:
[0149]
[0150] in, For the specific image to be tested, The classification result for the specific image to be tested. Indicates will Input encoder, Indicates will Input classifier.
[0151] This testing phase only requires one forward inference operation, without the need for neighborhood search, group probability update, or ranking loss. It has high inference efficiency and is suitable for deployment.
[0152] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs a specific image classification method based on sequence relationship modeling as described above to detect which stage a specific image belongs to.
Claims
1. A specific image classification method based on sequence relationship modeling, characterized in that, Includes the following steps: Step 1: Obtain specific image data as a sample set, label the sample set according to the natural stage order, and then divide the sample set into a training set, a validation set, and a test set; Step 2: Establish a specific image classification model, which includes a global sample aggregation module and a dual-view sequence relationship constraint module; map the input sample i into embedded features through encoder E. ; The global sample aggregation module includes a classifier, a KNN module, and a confidence prediction module based on global neighbors, which embeds features. The stage prediction probability of sample i is obtained through the classifier. Simultaneously embedding features The input is the KNN module, which obtains the K nearest neighbors of sample i; the confidence prediction module based on global neighbors predicts the stage probability of sample i. The group probability is obtained by aggregating the stage prediction probabilities of the K nearest neighbors of sample i. Then, based on the confidence threshold, sample i is divided into easy sample, hard sample, or noisy sample, and the cross-entropy classification loss for easy and hard samples is calculated. The dual-view sequence relationship constraint module includes a distance-based sequence relationship constraint submodule and a probability unimodal constraint submodule. The distance-based sequence relationship constraint submodule applies constraints to easy samples and calculates the distance-based sequence relationship loss. This ensures that the distances between easy samples follow the natural progression order; the probability unimodal constraint submodule applies constraints to both easy and difficult samples and calculates the unimodal loss. This ensures that the probability of the true class is the highest and decreases monotonically along the sequence direction; Then, the cross-entropy classification loss is used. Distance-based sequence relation loss and single-peak loss Add them together to get the final total loss. Finally, using the total final loss Update the parameters of the network model to obtain a specific image classification model after one training cycle; Step 3: Use the training set and validation set to train and validate the specific image classification model to obtain the trained specific image classification model; Step 4: Input the specific image to be tested from the test set into the trained specific image classification model to obtain the image classification result of the specific image to be tested.
2. The specific image classification method based on sequence relationship modeling according to claim 1, characterized in that: Let the set of the K nearest neighbors of sample i obtained by the KNN module be denoted as . ,satisfy: And j≠i Where t represents the current training period, t-1 represents the previous training period, and sample j represents samples other than i and j. Any sample other than Let i be the embedding features of sample i in the previous training cycle. For the sample In the embedding features of the previous training cycle, ... For the sample The embedded features from the previous training cycle, For sample j, the embedding features in the previous training cycle, Let K be the set of K nearest neighbor samples corresponding to sample i in the feature space, where K is the number of nearest neighbors; This represents the Euclidean distance operation for extracting features.
3. The specific image classification method based on sequence relationship modeling according to claim 2, characterized in that: The steps for the confidence prediction module based on global neighbors to obtain group probabilities are as follows: Step a, the weight adjustment formula for obtaining the K nearest neighbors of sample i is: in, Let k be the weight of sample k with respect to sample i in the current training period, where k = 1, 2, ..., K. For sample i, the embedding features in the current training period For the embedding features of sample j in the current training period, For sample k, the embedding features in the current training period This is a temperature coefficient used to adjust the smoothness of the weight distribution; Step b: Obtain group probabilities : in, This represents the group probability of sample i in the current training period. This represents the predicted probability of sample i at the current stage of the training cycle. This represents the group probability of sample k in the previous training period.
4. The specific image classification method based on sequence relationship modeling according to claim 3, characterized in that: The confidence threshold includes a first sample threshold th1 and a second sample threshold th2, which are determined by the global neighbor-based confidence prediction module. The relationship with th1 and th2, if That is, corresponding to its actual stage The probability is greater than If sample i is a common sample, then it is determined that sample i is a common sample; if the previous condition is not met, then... That is, except for the true phase, the maximum probability for all other phases does not exceed If sample i is a difficult sample, then it is considered a difficult sample; otherwise, it is considered a noisy sample.
5. The specific image classification method based on sequence relationship modeling according to claim 1, characterized in that: In step 1, the sample set is labeled according to the natural stage order as follows: in, , , These correspond to the mild, moderate, and severe stages of structural changes in a specific image, respectively. In step 2, sample i is mapped to embedded features. The formula is: Where E represents the encoder; In the global sample aggregation module, the classifier obtains the stage prediction probability of sample i. Obtained in the following way: in, For sample i to belong to The predicted probability of the stage For sample i to belong to The predicted probability of the stage For sample i to belong to The predicted probability of the stage ; This is the stage prediction probability vector.
6. The specific image classification method based on sequence relationship modeling according to claim 5, characterized in that: In the dual-view sequence relationship constraint module, the above-mentioned parameters are selected from the training set. One sample from each of the three stages, and data augmentation is performed on each sample to obtain six specific images: Stage Including original samples and enhanced samples ;stage Including original samples and enhanced samples ; stage Including original samples and enhanced samples Input the above six samples into encoder E(⋅) to obtain feature vectors, and define the feature distance as Euclidean distance: in, For the sample and Euclidean distance in feature space; m=1, 2, 3, 4, 5, 6; n=1, 2, 3, 4, 5, 6; In the distance-based sequence relation constraint submodule, three distance subsets are constructed. , , For the distance subset Construction: For the distance subset Construction: For the distance subset Construction: Define the average distance of the three distance subsets above: in, Represents a distance subset The average distance, denoted as ; Construct the average distance sorting vector : right Constructing the average sorting loss : Define the maximum-minimum distance relationship for the three distance subsets mentioned above: in, The distance between two samples in the same period is denoted as . ; The closest distance between two samples in adjacent stages is denoted as . ; The distance between two samples in adjacent stages is denoted as . ; The closest distance between two samples across stages is denoted as . ; Construct the maximum-minimum distance sorting vector : right Constructing the maximum-minimum sorting loss : Final distance order relation loss for: 。 7. The specific image classification method based on sequence relationship modeling according to claim 6, characterized in that: In the probabilistic unimodal constraint submodule, a unimodal loss is constructed using the difference between adjacent stages: in: For sample i, the loss is unimodal; These represent the model's relationship to adjacent stages. The predicted probability; This indicates the true stage label of the sample; and This indicates the sequence relationship between stages. Representation phase Lighter than stage , Representation phase Equal to or more serious than stage .
8. The specific image classification method based on sequence relationship modeling according to claim 7, characterized in that: The final total loss The calculation formula is: in, This is a hyperparameter used to balance the weights of different loss terms.
9. The specific image classification method based on sequence relationship modeling according to claim 1, characterized in that: In step 4, the formula for obtaining the classification result of the specific image to be tested is: in, For the specific image to be tested, The classification result for the specific image to be tested. Indicates will Input encoder, Indicates will Input classifier.