A small sample disease classification method based on association learning

Abstract

The present application relates to a kind of small sample disease classification methods based on association learning, comprising: first, build a small sample asphalt disease data set containing multiple asphalt disease types;Second, using the principle of association learning, develop a set of deep learning model and its training strategy suitable for small sample disease classification;Then, by setting "precision-speed" comprehensive evaluation index and a variety of damage simulation mode, evaluate the comprehensive performance and robustness of the model;Finally, the model is further adjusted and optimized, and deployed to road disease inspection equipment, realize the accurate classification of road disease.Compared with the prior art, the present application aims to solve the problem that traditional deep learning method excessively depends on training data size, alleviate the deficiency of existing method under small sample condition, improve the classification accuracy and system robustness, and provide a new solution for road disease detection and classification.

Description

Technical Field

[0001] This invention relates to the field of road defect detection technology, and in particular to a small-sample defect classification method based on association learning. Background Technology

[0002] Road defect classification is crucial for daily road condition management and maintenance. It can prevent serious road damage caused by the spread of defects and reduce the incidence of traffic accidents. In recent years, with the development of automation technologies such as computer vision, automated road defect classification has significantly reduced the workload and time required for manual defect inspection, improved the efficiency of road maintenance and management, and reduced management costs.

[0003] Currently, road defect classification methods based on computer vision technology can be mainly divided into two categories: traditional image processing-based methods and deep learning-based methods. Traditional methods often employ directional gradient histograms, local binary patterns, and wavelet transforms for image feature extraction, followed by classifiers such as backpropagation neural networks (BP neural networks) and support vector machines (SVMs) to classify road defects. While this approach overcomes the limitations of manual identification to some extent, the features in these algorithms rely on expert knowledge and lack universality. Furthermore, the performance of these methods is still limited by complex structures, diverse shapes, complex backgrounds, and strong interference from various noises (such as oil spots, gravel, and zebra crossings), failing to meet the needs of practical applications.

[0004] In recent years, with the increased availability of large-scale datasets (such as ImageNet) and the development of high-performance computing units, deep learning-based methods have attracted significant attention in various visual tasks. These methods use convolutional neural networks (CNNs) to obtain multi-level features from input data to complete the representation learning of input images, thus effectively supporting the solution of asphalt road distress classification problems. However, most of these supervised models require many labeled samples to fit the parameters of the deep CNN. To meet the requirements of deep learning in road distress detection tasks, some datasets on road cracks have been developed, such as the CFD dataset and the CRACK500 dataset, but the amount of data in these datasets is still quite limited compared to some large datasets. Secondly, road distress not only includes road cracks, but also rare categories such as ruts and wave bulges. More importantly, with the updating of road materials, the forms of road distress may undergo new changes. Currently, publicly available datasets containing various distress categories are still lacking. Among currently available road defect classification datasets, the RDD2020 dataset, while the largest and most comprehensive dataset in the field, still lacks defect types such as ruts, loosening, wave embracing, and repair. Furthermore, it lags significantly behind widely used datasets, such as the ImageNet dataset containing 1.2 million images and 1000 categories. More importantly, this data scarcity is prevalent not only in road defect classification but also in medical and industrial applications, where researchers struggle to collect enough labeled images to train deep learning models. Therefore, supervised neural network-based methods are difficult to apply to tasks with limited labeled samples, and the overfitting problem during training is difficult to solve. Even more importantly, while increasing the scale of data endows models with powerful capabilities, it also means a surge in development costs. For example, Chatgpt's training data exceeds 40TB, with a single training session costing as much as $1.5 million. Therefore, under the combined influence of these factors, researching methods based on limited sample data has significant practical implications. In recent years, few-shot learning has also attracted attention in computer vision tasks, especially image classification. The goal of few-shot learning is to enable machines to learn new things as easily as humans and perform related tasks well when there is very little available data. This research emerged to address the pain point of deep learning requiring a large number of labeled samples.

[0005] Therefore, in view of the problems that traditional road defect classification algorithms have poor adaptability and existing deep learning road defect classification algorithms are subject to large data limitations and have few classification categories, there is an urgent need to propose an asphalt pavement defect classification method based on small sample learning. Summary of the Invention

[0006] To address the shortcomings of traditional road defect classification algorithms, such as poor adaptability and the limitations of existing deep learning-based road defect classification algorithms due to data constraints and a limited number of classification categories, this invention proposes a few-shot learning-based asphalt pavement defect classification method. Specifically, by incorporating the association learning concept from few-shot learning, a deep learning model and training strategy suitable for few-shot defect classification tasks are constructed, enabling rapid adaptation to a small number of labeled road defect types.

[0007] When conceiving this invention, the applicant considered that a similar published patent, CN115965590A, addressed a road defect detection method based on limited cross-domain data learning. However, this invention primarily targets road defect detection tasks, not the road defect classification task involved in this invention. Methodologically, this invention also differs significantly from the methods described in the aforementioned published patent. The aforementioned patent mainly introduces a defect-weighted branch based on the two-stage deep detection model Faster R-CNN, making it suitable for few-shot detection tasks. This invention, on the other hand, is based on a lightweight road defect network fused with deep convolution and transformer technologies. Based on this, an association learning training strategy is constructed to achieve few-shot road defect classification.

[0008] The objective of this invention can be achieved through the following technical solutions:

[0009] This invention provides a small-sample disease classification method based on association learning, comprising the following steps:

[0010] S1. Construct a small-sample asphalt disease dataset suitable for asphalt road scenarios containing multiple types of asphalt disease;

[0011] S2. Design a lightweight pavement distress network based on the fusion of deep convolution and transformer;

[0012] S3. Based on the small sample asphalt disease dataset constructed in S1 and the lightweight pavement disease network constructed in S2, a small sample disease classification training strategy based on association learning is further constructed.

[0013] S4. Set a comprehensive evaluation index for the "accuracy-speed" of the evaluation model and test the overall performance of the model;

[0014] S5. Based on the comprehensive evaluation index constructed in S4, optimize and update the model;

[0015] S6. Deploy the optimized and updated model on the computing equipment of the road inspection vehicle to classify common road defects through the road inspection vehicle.

[0016] Furthermore, the specific processes in S1 include:

[0017] S11. List common asphalt pavement defects;

[0018] S12. Use vehicle-mounted camera equipment to collect image samples of common asphalt pavement distress types listed in S11, and use the collected image sample data to build a dataset.

[0019] S13. Label the collected data, including the types of diseases contained in the images, and divide the dataset into training and test sets with a ratio of 10% and 90% respectively according to a random partitioning method.

[0020] S14. By introducing data augmentation methods, the images in the constructed training and test sets are processed to expand the size of the dataset.

[0021] Furthermore, in S11, the common asphalt pavement defects include: transverse cracks, longitudinal cracks, block cracks, alligator cracks, potholes, ruts, and spalling.

[0022] Furthermore, in S12, n image samples are collected for each type, where n is not less than 500, and the samples cover different lighting conditions, different weather conditions, and different road types.

[0023] The different lighting conditions include sunny days, cloudy days, and night; the different weather conditions include rain, snow, fog, and clear skies; and the different road types include highways, urban roads, and rural roads.

[0024] In S14, the data augmentation methods include one or more of the following: contrast enhancement, color correction, rotation, scaling, flipping, cropping, and color transformation.

[0025] Furthermore, the specific processes in S2 include:

[0026] S21. Construct a lightweight feature extraction network F based on depthwise separable convolution. The feature extraction network F is used to extract the input image. Convert to image features

[0027]

[0028] In the formula, H and W are the length and width of the image, and C is the number of image feature channels;

[0029] S22. Flatten the image feature z into a feature embedding. in Indicates the length of the feature embedding. Each block is a D-dimensional vector; then, in A "CLS" token block is added to aggregate global information across the entire image; finally, a position code p is added to each block to preserve its relative position information within the image.

[0030] S23. Construct a feature interaction network G based on two Transformer layers to realize the interaction of image features at various spatial locations. Each coding layer contains a multi-head self-attention mechanism and a feedforward neural network.

[0031] S24. Using the feature output corresponding to the "CLS" token position of the last Transformer layer as input, construct a fully connected layer and a softmax layer, and output the road disease classification probability.

[0032] S25. Use the large open-source image dataset ImageNet to pre-train the feature extraction network F, the feature interaction network G, and the fully connected layer. The training objective is set as the cross-entropy loss function.

[0033] Furthermore, the specific process in S21 includes:

[0034] First, based on a 3×3 convolution with a stride of 2 f 3×3 BN layer f BN h-swish activation function g h-swish Construct the initial convolutional layer F1;

[0035] Then based on depthwise convolution f deep BN layer f BN ReLU activation function g ReLU Construct an 11-layer inverse residual module, denoted as F2 to F... 12 Among them, the inverse residual modules F3, F5, F7 and F 11 The step size is set to 2, and the step size of the other inverse residual modules is set to 1.

[0036] Finally, connect F1 to F in sequence. 12 This forms a feature extraction network F.

[0037] Furthermore, the specific processes in S3 include:

[0038] S31. Divide the large open-source image dataset ImageNet into base class datasets.

[0039] S32. Using the pre-trained feature extraction network F and feature interaction network G, construct an association learning strategy. This strategy utilizes the base class dataset partitioned in S31. Train the network F+G;

[0040] S34. For each sample in the query set, average its association score with the K samples of each class in the support set to obtain a score vector T with N elements.

[0041] S35. Construct a cross-entropy loss function to train the network F+G. The input of this loss function is the score vector T and the true label value. When the query feature and the support feature belong to the same category, the true label is set to 1, otherwise it is set to 0.

[0042] S36. After training on the base class dataset, the disease dataset collected and labeled in S13 is used to fine-tune the training of the association module R. During the fine-tuning process, the disease dataset is also divided into the same support set and query set.

[0043] Furthermore, in S31, the base class dataset includes M "N-way K-type" few-shot classification tasks B i Where M represents the number of few-sample classification tasks, K represents the number of image samples, N represents the number of disease categories to be classified, and i represents the index number of the classification task.

[0044] B i =S i ∪Q i

[0045] S i ={(C i1 ,I i1,1 ,I i1 ,2,...,I i1,K ),...,(C iN ,I iN,1 ,I iN ,2,...,I iN,K )}

[0046] Q i ={(C ik ,I ik,1 (4)

[0047] In the formula, S i For the support set, there are N classes, each with K image samples, Q i Let S be the query set, which contains one image sample whose class k belongs to the support set S. i Category set, C ij The j-th image selected from ImageNet is further analyzed in the S32 process, which includes:

[0048] First, the image data of the support set and query set are mapped to support features and query features through the network F+G;

[0049] Next, the query features and each type of support features are stacked by channel and fed into the association module R, which consists of two convolutional modules and two fully connected layers, to calculate the association score between the query set samples and each sample in the support set.

[0050] Furthermore, the specific processes in S4 include:

[0051] S41. Input the disease dataset collected and labeled in S13 as the query set into the fine-tuned network composed of the training network F+G and the association module R, and output the corresponding score vector T. The predicted category of the query sample is the class with the highest score in the vector T.

[0052] S42. Introduce classification precision, accuracy, recall, F1 score, and AUC as quantitative indicators to measure model performance. Introduce inference speed, model computational complexity, and model size under the same hardware environment to measure model inference efficiency.

[0053] Based on S41 and S42, a comprehensive evaluation index of "accuracy-speed" is set to evaluate the model and test the overall performance of the model.

[0054] Compared with the prior art, the present invention has the following technical advantages:

[0055] (1) This invention designs a lightweight neural network architecture to efficiently learn and classify asphalt pavement defects in few-shot learning tasks. Specifically, this invention utilizes depthwise separable convolutions and residual modules in the network structure design to reduce model parameters and computational complexity. Furthermore, after constructing a feature extraction network using depthwise separable convolutions, a Transformer-based feature interaction network is built to achieve global association and interaction of regional features, thereby further improving classification accuracy.

[0056] (2) This invention proposes a small-sample disease classification training strategy based on association learning. By calculating the association score between samples in the query set and samples in the support set, the similarity and difference between samples can be measured more accurately. This strategy not only utilizes pre-trained feature extraction and interaction networks, but also introduces a disease classification-specific association module, thereby providing more stable and robust classification results under small-sample conditions. Attached Figure Description

[0057] Figure 1 This is a flowchart of the small-sample disease classification method based on association learning in this invention. Detailed Implementation

[0058] To address the shortcomings of traditional road defect classification algorithms, such as poor adaptability and the limitations of existing deep learning-based road defect classification algorithms due to data constraints and a limited number of classification categories, this invention proposes a few-shot learning-based asphalt pavement defect classification method, effectively alleviating the deficiencies of existing methods. Specifically, this invention first constructs a few-shot asphalt defect dataset suitable for asphalt road scenarios containing various types of asphalt defects. Based on this, and combining the idea of ​​association learning in few-shot learning, a deep learning model and training strategy suitable for few-shot defect classification tasks are constructed. Finally, a comprehensive evaluation index of "accuracy-speed" is set to assess the model, and various image corruptions are applied to the original input image, such as randomly adding noise and adjusting contrast, to test the model's overall performance and robustness to noise interference, thereby enabling further adjustment and optimization of the model.

[0059] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Component models, material names, connection structures, control methods, algorithms, and other features not explicitly described in this technical solution are considered common technical features disclosed in the prior art.

[0060] Example 1

[0061] like Figure 1 As shown, the specific steps of the small-sample disease classification method based on association learning in this embodiment are as follows:

[0062] S1. Construct a small-sample asphalt disease dataset suitable for asphalt road scenarios containing various types of asphalt defects. Specific implementation includes:

[0063] S11 lists various common asphalt pavement defects, including transverse cracks, longitudinal cracks, block cracks, alligator cracks, potholes, ruts, and spalling.

[0064] S12. Use vehicle-mounted camera equipment to collect the common asphalt pavement distress types listed in S11. Collect n image samples for each type, with n not less than 500, covering different lighting conditions (sunny day, cloudy day, night), weather conditions (rain, snow, fog, and sunshine), and road types (highway, urban road, rural road, etc.).

[0065] S13. Label the collected data, including the types of diseases contained in the images, and divide them into training and test sets with a ratio of 10% and 90% according to a random partitioning method.

[0066] S14. Introduce data augmentation techniques such as contrast enhancement, color correction, rotation, scaling, flipping, cropping, and color transformation to expand the size of the dataset.

[0067] S2. Design a lightweight pavement distress network based on the fusion of depthwise convolution and transformer, specifically including:

[0068] S21. Construct a lightweight feature extraction network F based on depthwise separable convolutions, specifically including: first, based on 3×3 convolutions f 3×3 (Step size is 2), BN layer f BN and h-swish activation function g h-swish Construct an initial convolutional layer F1; then, based on a depthwise convolution f deep BN layer f BN and ReLU activation function g ReLU Construct an 11-layer inverse residual module (denoted as F2 to F). 12 ), where the inverse residual modules F3, F5, F7 and F 11 The step size is set to 2, while the step size of the remaining inverse residual modules is set to 1; finally, connect F1 to F in sequence. 12 This forms a feature extraction network F; this network is used to extract features from the input image. Convert to image features

[0069]

[0070] In the formula, H and W are the length and width of the image, and C is the number of image feature channels;

[0071] S22. Flatten the image feature z into a feature embedding. in Indicates the length of the feature embedding. Each block is a D-dimensional vector; then, in A "CLS" token block is added to aggregate global information across the entire image; finally, a position code p is added to each block to preserve its relative position information within the image.

[0072] S23. Construct a feature interaction network G based on two Transformer layers to realize the interaction of image features at various spatial locations. Each coding layer contains a multi-head self-attention mechanism and a feedforward neural network.

[0073] S24. Using the feature output corresponding to the "CLS" token position of the last Transformer layer as input, construct a fully connected layer and a softmax layer, and output the road disease classification probability.

[0074] S25. Utilize the large open-source image dataset ImageNet (covering categories C1, C2, ..., C...). n The feature extraction network F, the feature interaction network G, and the fully connected layer are pre-trained, and the training objective is set to the cross-entropy loss function.

[0075] S3. Construct a small-sample disease classification training strategy based on association learning;

[0076] S31. Divide the large open-source image dataset ImageNet into base class datasets. This base class dataset contains M "N-way K-method" few-shot classification tasks B. i Where M represents the number of few-sample classification tasks, K represents the number of image samples, N represents the number of disease categories to be classified, and i represents the index number of the classification task.

[0077] B i =S i ∪Q i

[0078] S i ={(C i1 ,I i1,1 ,I i1,2 ,...,I i1,K ),...,(C iN ,I iN,1 ,I iN,2 ,...,I iN,K )}

[0079] Q i ={(C ik ,I ik,1 (6)

[0080] In the formula, S i For the support set (containing N classes, K image samples per class), Q i The query set contains one image sample whose category k belongs to the support set S. i (Category set), C ij I is the j-th category selected from ImageNet. ij,m From category C ij The m-th image selected;

[0081] S32. Using the pre-trained feature extraction network F and feature interaction network G, construct an association learning strategy. This strategy utilizes the base class dataset partitioned in S31. The network F+G is trained. Specifically, the image data of the support set and query set are first mapped to support features and query features through the network F+G. Then, the query features and each type of support features are stacked by channel and fed into the association module R, which consists of two convolutional modules and two fully connected layers. The association score between the query set samples and each sample in the support set is calculated.

[0082] S34. For each sample in the query set, average its association score with the K samples of each class in the support set to obtain a score vector T with N elements.

[0083] S35. Construct a cross-entropy loss function to train the network F+G. The input to this loss function is the score vector T and the true label value. When the query feature and the support feature belong to the same category, the true label is set to 1, otherwise it is set to 0.

[0084] S36. After training on the base class dataset, the disease dataset (training set) collected and labeled in S13 is used to fine-tune the training of the association module R. During the fine-tuning process, the disease dataset (training set) is also divided into the same support set and query set.

[0085] S4. Set a comprehensive evaluation index for the "accuracy-speed" of the evaluation model and test the overall performance of the model;

[0086] S41. Input the disease dataset (test set) collected and labeled in S13 as the query set into the fine-tuned network composed of the training network F+G and the association module R, and output the corresponding score vector T; the predicted category of the query sample is the class with the highest score in the vector T.

[0087] S42. Introduce quantitative indicators such as classification precision, accuracy, recall, F1 score, and AUC value to measure model performance, and introduce indicators such as inference speed, model computational complexity, and model size under the same hardware environment to measure model inference efficiency.

[0088] S5. Based on the comprehensive evaluation indicators, optimize, update, and deploy the model.

[0089] Table 1 presents the classification performance of the proposed method for seven common road defects (transverse cracks, longitudinal cracks, block cracks, alligator cracks, potholes, ruts, and spalling) in the form of accuracy, recall, F1 score, and AUC. Only 30 images were provided as the support set for each type of defect during training. As shown in Table 1, the proposed method performs relatively well in classifying potholes and transverse and longitudinal cracks among these seven categories. This is likely because potholes and transverse and longitudinal cracks have higher visual distinctiveness than other defect types, making them easier to distinguish. In contrast, the model has poor classification accuracy for alligator cracks, block cracks, and spalling, which typically possess similar appearance features or spatial relationships. For example, both block cracks and alligator cracks exhibit irregular crack distributions, while pavement spalling is usually accompanied by block cracks or alligator cracks.

[0090] Table 2 compares the proposed lightweight pavement disease network, which is trained using 30 samples per class, with the small-sample disease classification method based on association learning used in this invention.

[0091] Table 1. Average accuracy (AP) and average accuracy (mAP) of the present invention for five categories of road defects.

[0092]

[0093] Table 2 Comparison Results

[0094]

[0095] The above description of the embodiments is provided to enable those skilled in the art to understand and use the invention. It will be apparent to those skilled in the art that various modifications can be made to these embodiments, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the invention should be within the protection scope of the present invention.

Claims

1. A small-sample disease classification method based on association learning, characterized in that, Includes the following steps: S1. Construct a small-sample asphalt disease dataset suitable for asphalt road scenarios containing multiple types of asphalt disease; S2. Design a lightweight pavement distress network based on the fusion of deep convolution and transformer; The specific process in S2 includes: S21. Construct a lightweight feature extraction network based on depthwise separable convolutions. The feature extraction network Used to input image Convert to image features : (1) In the formula, , For the length and width of the image, and The number of image feature channels; S22, Image features Flattened into feature embedding ,in Indicates the length of the feature embedding. Each block is one dimensional vector; then, in Previously, a "CLS" token block was added to aggregate global information across the entire image; finally, positional encoding was added to each block. This is to preserve its relative position information in the image; S23. Construct a feature interaction network based on two Transformer layers. This enables the interaction of image features at various spatial locations, where each coding layer contains a multi-head self-attention mechanism and a feedforward neural network. S24. Using the feature output corresponding to the "CLS" token position of the last Transformer layer as input, construct a fully connected layer and a softmax layer, and output the road disease classification probability. S25. Utilizing the large open-source image dataset ImageNet to improve the feature extraction network. The feature interaction network G and the fully connected layer are pre-trained, and the training objective is set to the cross-entropy loss function. The specific process in S21 includes: First, a 3×3 convolution with a stride of 2 is applied. BN layer h-swish activation function Construct the initial convolutional layer ; Then based on depthwise convolution BN layer ReLU activation function Construct an 11-layer inverse residual module, denoted as arrive Among them, the inverse residual module , , and The step size is set to 2, and the step size of the other inverse residual modules is set to 1. Finally, connect them in order. arrive To form a feature extraction network ; S3. Based on the small sample asphalt disease dataset constructed in S1 and the lightweight pavement disease network constructed in S2, construct a small sample disease classification training strategy based on association learning. The specific processes in S3 include: S31. Divide the large open-source image dataset ImageNet into base class datasets. ; S32. Utilize the pre-trained feature extraction network A feature interaction network G is used to construct an association learning strategy, which utilizes the base class dataset partitioned in S31. Training Network ; S34. For each sample in the query set, average its association score with the K samples of each class in the support set to obtain a score vector of N elements. ; S35. Constructing the cross-entropy loss function for the network The training process takes a score vector as input to the loss function. The true label value is set to 1 when the query feature and the supporting feature belong to the same category, and 0 otherwise. S36. After training on the base class dataset, the disease dataset collected and labeled in S13 is used to train the association module. Fine-tuning training was conducted, and during the fine-tuning process, the disease dataset was also divided into the same support set and query set format. S4. Set a comprehensive evaluation index for the "accuracy-speed" of the evaluation model to test the overall performance of the model; S5. Based on the comprehensive evaluation index constructed in S4, optimize and update the model; S6. Deploy the optimized and updated model on the computing equipment of the road inspection vehicle to classify common road defects through the road inspection vehicle.

2. The method for classifying small-sample diseases based on association learning according to claim 1, characterized in that, The specific process in S1 includes: S11. List common asphalt pavement defects; S12. Use vehicle-mounted camera equipment to collect image samples of common asphalt pavement distress types listed in S11, and use the collected image sample data to build a dataset. S13. Label the collected data, including the types of diseases contained in the images, and divide the dataset into training and test sets with a ratio of 10% and 90% respectively according to a random partitioning method. S14. By introducing data augmentation methods, the images in the constructed training and test sets are processed to expand the size of the dataset.

3. The method for classifying small-sample diseases based on association learning according to claim 2, characterized in that, In S11, the common asphalt pavement defects include: transverse cracks, longitudinal cracks, block cracks, alligator cracks, potholes, ruts, and spalling.

4. The method for classifying small-sample diseases based on association learning according to claim 2, characterized in that, In S12, n image samples are collected for each type, with n not less than 500, covering different lighting conditions, different weather conditions, and different road types; The different lighting conditions include sunny days, cloudy days, and night; the different weather conditions include rain, snow, fog, and clear skies; and the different road types include highways, urban roads, and rural roads. In S14, the data augmentation methods include one or more of the following: contrast enhancement, color correction, rotation, scaling, flipping, cropping, and color transformation.

5. The method for classifying small-sample diseases based on association learning according to claim 1, characterized in that, In S31, the base class dataset includes M "N-way K-type" few-shot classification tasks. Where M represents the number of few-sample classification tasks, K represents the number of image samples, N represents the number of disease categories to be classified, and i represents the index number of the classification task. (2) In the formula, The support set contains N classes, with K image samples in each class. The query set contains one image sample, whose category is... k Belongs to support set Category collection, The first one selected from ImageNet j Categories From category The m-th image selected.

6. The method for classifying small-sample diseases based on association learning according to claim 1, characterized in that, The specific process of S32 includes: Firstly via the network Map the image data of the support set and query set to support features and query features; Next, the query features and each type of support features are stacked channel-wise and fed into the correlation module, which consists of two convolutional modules and two fully connected layers. In this process, the association score between the query set sample and each sample in the support set is calculated.

7. The method for classifying small-sample diseases based on association learning according to claim 2, characterized in that, The specific processes in S4 include: S41. Input the disease dataset collected and labeled in S13 as the query set into the fine-tuned training network. Related modules The network consists of a network that outputs the corresponding score vector. The predicted category of the query sample is the vector. The class with the highest score; S42. Classification precision, accuracy, recall, F1 score, and AUC value are introduced as quantitative indicators to measure model performance. Inference speed, model computational complexity, and model size are introduced under the same hardware environment to measure model inference efficiency. Based on S41 and S42, a comprehensive evaluation index of "accuracy-speed" is set to evaluate the model, thereby testing the overall performance of the model.

Images