Knowledge-guided Few-shot Object Detection and Recognition Method

Abstract

The present invention provides a small-sample target detection and recognition method based on knowledge guidance, which obtains an image training set of the target to be classified; performs instance segmentation on the semantic level for each target image, and extracts features as real-scene primitives through a preset residual neural network to obtain feature vectors of the real-scene primitives, and constructs a feature matrix of the target image based on the feature vectors and the number of basic primitives in the primitive space; based on the prior knowledge of basic primitives and semantic information and the prior knowledge between real-scene primitives, the prior logical knowledge and semantic information are inferred in the form of a knowledge graph; the present invention constructs a knowledge-driven knowledge graph by associating image features with semantic primitives, effectively alleviating the model's dependence on data, greatly improving the accuracy in small-sample scenarios, and can solve the problem that the existing technology relies on a large amount of accurately labeled data and often suffers significant performance loss under small-sample conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a small sample target detection and recognition method based on knowledge guidance. Background Art

[0002] In recent years, computer technology has rapidly developed, and artificial intelligence has also made continuous progress. Deep learning models and large-scale datasets have powerful expressive capabilities. As a result, deep learning-based object detection and recognition technologies have demonstrated unprecedented advantages and achieved encouraging performance. As a result, object detection technology has been widely and effectively applied in production, daily life, and the military. However, in practical applications, due to the difficulty in obtaining data, the significant human and resource requirements, or due to limitations in certain fields, we cannot obtain large datasets. Therefore, the problem of weak object detection is more common in practical applications. Currently, there are two main approaches to address the problem of weak object recognition: one is based on data adjustment, which involves enhancing the data and expanding the dataset to avoid overfitting; the other is model adjustment, which involves introducing effective modules and removing useless modules from the existing model to facilitate the extraction of image information and improve recognition accuracy. Currently, weak object recognition research both domestically and internationally is based on these two approaches.

[0003] The prior art patent document "A Method for Recognizing Small Targets Based on Deep Neural Networks" (Application No.: CN201711369730.9, Application Publication No.: CN108122003A, Publication Date: 2018.06.05) proposes a new deep neural network structure that focuses on extracting small target features and increasing the depth of image data. The method mainly includes the following steps: (1) obtaining deep neural network training data and constructing a small target image sample database based on existing small target related data information; (2) building a deep neural network and using gradient descent and backpropagation algorithms to train the deep neural network model; (3) using Gaussian pyramids and sliding window search methods to identify small targets in the image. This method still has shortcomings: it requires the construction of a dataset similar to the target to be identified in order to achieve image data enhancement, which has great limitations for target recognition where public datasets are difficult to obtain.

[0004] In their paper "Weak Target Recognition Method Based on Improved YOLOv5m" (Computer Measurement and Control: 1-9 [2022-11-03], published on August 18, 2022), Yang Wentao et al. proposed a multi-target recognition and detection method based on an improved YOLOv5m network. The main steps of this method are: (1) Mosaic data enhancement is performed on the input image to enrich the background of the image detection and recognition target; (2) The most suitable anchor size is obtained based on the adaptive anchor box, the target is framed, and the processed image is placed in the focus module for slicing; (3) The feature fusion network of FPN+PAN is improved to allow the model to learn more learning features; (4) A small target detection layer is added to combine the shallow feature layer map and the deep feature layer map for detection. However, this method still has shortcomings: the improvement of the feature fusion network and the addition of the small target detection layer increase the complexity of the network, and the time spent on detection and recognition also increases accordingly.

[0005] The prior art patent document "A Method for Detecting Small Targets" (Application No.: 201910685093.9, Application Publication No.: CN110414600 A, Publication Date: 2019.11.05) proposes a small sample spatial target recognition method based on transfer learning with high recognition accuracy. The main steps of this method are: (1) establishing an auxiliary sample spatial target dataset; (2) constructing an end-to-end deep nearest neighbor network; (3) feeding the auxiliary dataset into the deep nearest neighbor network for training; (4) constructing a spatial target dataset; (5) feeding the target dataset into the deep nearest neighbor network for recognition. However, this method still has shortcomings: the deep nearest neighbor network has a large number of network layers, the number of pixels of small targets is small, and information loss occurs during the network embedding process. Summary of the Invention

[0006] In order to solve the above problems existing in the prior art, the present invention provides a small sample target detection and recognition method based on knowledge guidance. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0007] The present invention provides a small sample target detection and recognition method based on knowledge guidance, including:

[0008] Step 1: Obtain the target image to be classified;

[0009] Step 2: Based on prior knowledge, all basic primitives that can represent the high-level semantic features of the target in the target image are pre-planned to form the primitive space of the entire target;

[0010] Step 3: Using the basic primitives in the primitive space to guide instance segmentation of each target image at the semantic level, obtaining superpixel blocks corresponding to pixels on the basic primitives, and using each superpixel block as a real scene primitive of the target image;

[0011] Step 4: Extract features of the real scene primitives of the target image through a preset residual neural network, extract the feature vector of the previous layer of the residual neural network output layer, use the feature vector as the feature vector of the real scene primitive, and construct a feature matrix of the target image based on the feature vector and the number of basic primitives in the primitive space;

[0012] Step 5: Based on the prior knowledge of basic primitives and semantic information and the prior knowledge of logical associations between basic primitives, a logical association graph between semantics and basic primitives is established to obtain a knowledge graph.

[0013] Step 6: Using the knowledge graph and the feature matrix of each target image, perform category inference and prediction on the feature matrix corresponding to the target image through the trained graph convolutional network to obtain the classification result of the target image.

[0014] Beneficial effects of the present invention:

[0015] First, since the present invention utilizes instance segmentation to segment the image into super-pixel blocks that can represent high-level semantic features, it combines human brain perception information with computer information to generate semantic primitives, making the present invention have the characteristics of human-computer integration.

[0016] Second, since the present invention utilizes the feature information of semantic primitives and prior logical knowledge to establish a knowledge graph and uses knowledge to guide target recognition, it avoids the black box problem in deep learning and makes the present invention interpretable in weak target recognition.

[0017] Third, since the present invention uses a graph convolutional network and utilizes the knowledge graph to construct a graph structure, and then embeds knowledge into the network, it reduces the dependence of ordinary networks on data sets, which greatly improves the accuracy of the present invention in identifying weak targets.

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of a knowledge-guided small sample target detection and recognition method provided by the present invention;

[0020] Figure 2 It is an experimental simulation diagram of the present invention. DETAILED DESCRIPTION

[0021] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0022] Before introducing the present invention, the technical concept of the present invention is first introduced.

[0023] The purpose of the present invention is to solve the problems of high difficulty in data acquisition and data labeling, small number of samples available for training, and poor interpretability of detection and recognition tasks. The idea of achieving the purpose of the present invention is: the present invention extracts semantic primitives by drawing on the process of the brain quickly extracting target semantic primitives from the scene under the guidance of prior knowledge, and uses knowledge association to complete small sample learning, thereby solving the problem of obvious performance loss in small sample scenarios in engineering applications due to reliance on a large amount of accurately labeled data. The present invention obtains a feature matrix of the image by performing feature extraction on the primitives representing high-level semantic information obtained by segmentation, and constructs a mutual mapping relationship between deep features and semantic primitives, thereby solving the problem of the semantic gap in computer vision in engineering applications. The present invention computationally constructs a knowledge-driven knowledge graph, embeds prior knowledge into a graph convolutional network in the form of a graph structure for reasoning, solves the data dependence and "black box" problem in engineering applications, and has the characteristic of interpretability.

[0024] To achieve the purpose of the present invention, the present invention provides a small sample target detection and recognition method based on knowledge guidance, including:

[0025] Step 1: Obtain the target image to be classified;

[0026] Step 2: Based on prior knowledge, all basic primitives that can represent the high-level semantic features of the target in the target image are pre-planned to form the primitive space of the entire target;

[0027] Step 3: Using the basic primitives in the primitive space to guide instance segmentation of each target image at the semantic level, obtaining superpixel blocks corresponding to pixels on the basic primitives, and using each superpixel block as a real scene primitive of the target image;

[0028] Specifically, step 3 includes:

[0029] Step 31: Use the trained SOLO2 algorithm network to generate the instance mask of S^2 channels corresponding to the S×S grid;

[0030] Step 32: Use the basic primitives in the primitive space to guide the instance mask to perform instance segmentation on the target image at the semantic level, obtain superpixel blocks corresponding to pixels on the basic primitives, and use each superpixel block as the real scene primitive of the target image.

[0031] Each target image in the image training set is subjected to instance segmentation. The present invention uses the SOLO2 algorithm to perform instance segmentation on the input image at the semantic level, and the superpixel blocks obtained by segmentation are used as the primitives of the image.

[0032] For generating the mask in the SOLO algorithm:

[0033] First generate the S corresponding to the S×S grid 2 The instance mask of the channel, the last layer is a pyramid feature F∈R H ×W×E As input, it is finally applied to a 2 The convolution operation of the output channel is:

[0034] M i,j = i,j *

[0035] Among them, G i,j ∈R 1×1×E is the convolution kernel, M i,j ∈R H×W is the final mask that contains only one instance whose center is at position (,j);

[0036] Secondly, the loss function in the instance segmentation algorithm SOLO2 is:

[0037] L=L cate +L mask

[0038] Among them, L cate is the conventional Focal Loss for semantic classification, L mask is the Dice Loss of mask prediction.

[0039] Step 4: Extract features of the real scene primitives of the target image through a preset residual neural network, extract the feature vector of the previous layer of the residual neural network output layer, use the feature vector as the feature vector of the real scene primitive, and construct a feature matrix of the target image based on the feature vector and the number of basic primitives in the primitive space;

[0040] Specifically, step 4 includes:

[0041] Step 41: constructing a residual neural network, and training the residual neural network to obtain a trained residual neural network;

[0042] Step 42: Using the trained residual neural network to extract features from the real scene primitives of each target image, so as to represent the real scene primitives as 1×D feature vectors;

[0043] Step 43: Initialize an all-zero matrix according to the number N of basic primitives in the primitive space and the length D of the eigenvector;

[0044] Step 44: Fill the eigenvectors in the corresponding dimensions of the all-zero matrix to obtain the eigenmatrix X i .

[0045] During specific operation, step 4 can be implemented through the following process:

[0046] (3a) For the obtained primitives, a residual neural network (ResNet) is constructed to extract their features, representing each real-scene primitive as a 1×D feature vector. During the training phase, this process is modeled as a classification task for real-scene primitives. All primitives segmented in the training set are supervised by cross-entropy loss to obtain feature expressions for different primitives.

[0047] (3b) After ResNet training is complete, the primitive is input into the network for feature extraction, and the feature vector of the layer before the output layer is taken as the feature vector of the real scene primitive. The specific method is as follows:

[0048] For the trained ResNet residual neural network model, we use the penultimate fully connected layer as the output and retain a vector of length 128 as the feature vector of each primitive. The length of the output vector can be adjusted according to the complexity of the classification space. When the category similarity in the classification space is high and the classification task is complex, the number of nodes in the penultimate fully connected layer can be increased. Conversely, when the classification space is simple, the number of nodes in the penultimate fully connected layer can be reduced.

[0049] Initialize the all-zero matrix according to the number of primitives N and the length of the output feature vector D in (3b), and use the image primitives to extract the features in (3b) to fill the all-zero matrix to obtain the feature matrix X i , the specific method is as follows:

[0050] Initialize the all-zero matrix according to the number of primitives N in the constructed classification space and the length D of the output feature vector in (3b). For the primitives extracted from the i-th image in the data set, these primitives are extracted in (3b) and placed in the corresponding dimension of the initialization matrix to obtain the feature matrix X i .

[0051] Step 5: Based on the prior knowledge of basic primitives and semantic information and the prior knowledge of logical associations between basic primitives, a logical association graph between semantics and basic primitives is established to obtain a knowledge graph.

[0052] Specifically, step 5 includes:

[0053] Step 51: mapping the basic primitives and their semantic information based on the prior knowledge of the basic primitives and semantic information in the primitive space;

[0054] Step 52: Summarize the same basic primitives and form semantic association pairs corresponding to the basic primitives;

[0055] Step 53: Construct a knowledge graph based on the semantic association pairs and the topological relationships between the basic primitives.

[0056] During specific operation, step 5 can be implemented through the following process:

[0057] (4a) Prior knowledge, that is, the understanding of things and their attributes obtained through experience, is used to guide the mapping of primitives and their semantic information, summarize the same primitives and form semantic association pairs corresponding to the primitives, thereby eliminating semantic ambiguity.

[0058] (4b) Through prior knowledge, the connections between the primitives are clarified, and the graph structure enables the topological relationship between the semantic association pairs to be reflected, so that the model can realize the cognition of semantic associations not only at the primitive level but also at the target level.

[0059] Step 6: Using the knowledge graph and the feature matrix of each target image, perform category inference and prediction on the feature matrix corresponding to the target image through the trained graph convolutional network to obtain the classification result of the target image.

[0060] Among them, the process of training the graph convolutional network is to perform iterative training based on the training set, use the test set to test the trained graph convolutional network, and finally obtain the trained graph convolutional network.

[0061] Specifically, step 6 includes:

[0062] Step 61: For a graph structure data with N nodes in the knowledge graph, each of which stores D-dimensional features, a feature matrix X with a dimension of N×D can be obtained by highlighting the nodes in the knowledge graph;

[0063] Among them, the adjacency relationship between each node constitutes an adjacency matrix A with a dimension of N×N;

[0064] Step 62: Using the feature matrix X and the adjacency matrix A as inputs to the graph convolutional network, the feature matrix is updated through the graph convolutional network, and the feature matrix of each layer is updated based on the propagation formula between the layers of the graph convolutional network;

[0065] The propagation formula between graph convolutional network layers can be expressed as:

[0066]

[0067] in, I represents the identity matrix, yes The degree matrix, W () is the learnable weight of the lth layer in the graph convolutional network, H () is the feature matrix of the lth layer in the graph convolutional network. When l = 0, H () That is, the input X, σ is a nonlinear activation function.

[0068] Step 63: Based on the updated feature matrix X, the classification result is output by the trained graph convolutional network, and the classification result of the target image is predicted based on the classification result, thereby realizing knowledge-driven interpretable small-sample reasoning.

[0069] Step 63 includes:

[0070] Step 631: The first layer in the graph convolutional network performs a graph convolution operation on the input feature matrix X and outputs a feature matrix H using a ReLU activation function.

[0071] Among them, the output of the first layer is The ReLU activation function is: f(x) = max(0,x);

[0072] Step 632: The second layer performs a graph convolution operation on the feature matrix H and outputs the result using the softmax activation function;

[0073] The output of the second layer is:

[0074] Among them, the softmax activation function is:

[0075]

[0076] Step 633: Determine the inference formula of the graph convolutional network based on the output of the first layer and the output of the second layer;

[0077] Taking into account the number of samples and the efficiency of model training and inference, the present invention constructs a two-layer graph convolutional network. If the complexity of the classification space increases, the number of layers of the graph convolutional network can be increased, allowing the graph convolutional model to perform more complex inferences.

[0078] The graph convolution model inputs the primitives of the target in matrix form according to the feature matrix X generation method, thus including the activation state of the target semantics for the graph nodes. The adjacency matrix A describes the topological relationship between the semantic primitives based on prior knowledge. The inference formula of the graph convolution network can be expressed as:

[0079]

[0080] Among them, Z is the output result of the graph convolutional network, f represents the inference function of the network, ReLU and softmax represent activation functions;

[0081] Step 634: Use the inference formula to infer the classification result output by the graph convolutional network, and predict the classification result of the target image based on the classification result, thereby realizing knowledge-driven explainable small sample reasoning.

[0082] The present invention provides a small-sample target detection and recognition method based on knowledge guidance, which obtains an image training set of the target to be classified; performs instance segmentation on the semantic level for each target image, and extracts features as real-scene primitives through a preset residual neural network to obtain feature vectors of the real-scene primitives, and constructs a feature matrix of the target image based on the feature vectors and the number of basic primitives in the primitive space; based on the prior knowledge of basic primitives and semantic information and the prior knowledge between real-scene primitives, the prior logical knowledge and semantic information are inferred in the form of a knowledge graph; the present invention constructs a knowledge-driven knowledge graph by associating image features with semantic primitives, effectively alleviating the model's dependence on data, greatly improving the accuracy in small-sample scenarios, and can solve the problem that the existing technology relies on a large amount of accurately labeled data and often suffers significant performance loss under small-sample conditions.

[0083] The effects of the present invention are further described below in conjunction with simulation experiments:

[0084] 1. Simulation experiment conditions:

[0085] The hardware platform of the simulation experiment of the present invention is: the processor is Intel (R) Xeon (R) CPU E5-2680v3, the main frequency is 2.90GHz, the memory is 16GB, and the GPU is RTX 2080Ti.

[0086] The software platforms for the simulation experiment of the present invention are: Windows 10 operating system, Python 3.8, PyTorch 1.7.0 and PyCharm 2017.

[0087] This paper experimentally validates the proposed algorithm using a small-sample military target recognition task. The military target recognition dataset used in this experiment was collected from the internet and contains five target categories: tanks, armored vehicles, trucks, camouflaged targets, and jeeps. 274 images were used to train the model, and 118 images were used to verify the algorithm's performance. Each image was 300 × 300 × 3 pixels in JPEG format. The number of target samples for each category is shown in Table 1:

[0088] Table 1 Number of samples in each category of the dataset

[0089]

[0090] From Table 1, we can see that the entire dataset meets the requirements of the recognition task in a small sample scenario, and the number of samples of individual categories such as jeep is very small.

[0091] 2. Simulation content and results analysis:

[0092] The simulation experiment of this invention selected the classic classification networks ResNet-50, DenseNet-121 and ShuffleNet v2 as comparison algorithms. The experimental results are shown in Table 2:

[0093] Table 2 Number of samples in each category of the dataset

[0094]

[0095] From Table 2, we can see that the proposed method achieved a classification accuracy of 88.2%, which is 25.5% higher than the best existing classification network DenseNet-121. This intuitively demonstrates the superiority of the object recognition method based on semantic association reasoning under small sample conditions.

[0096] The following is a further description of the steps of extracting primitives in the present invention in conjunction with the simulation diagram:

[0097] Figure 2 Figure (a) shows the "turret" primitive extracted from the simulation experiment of the present invention. As can be seen from the image, the superpixel block composed of irregular pixels can effectively express the semantic information of the turret. Compared with directly segmenting the image, the superpixel block extracted by the present invention is not affected by redundant pixels and is more in line with the human brain's association and reasoning process.

[0098] Figure 2 Figure (b) shows the "car body" primitive extracted from the simulation experiment of the present invention; Figure 2 Figure (c) shows the "body" primitive extracted from the simulation experiment of the present invention. These three primitives constitute the primitives of the tank image. After inputting them into the feature extraction network respectively, the obtained feature vectors are put into the corresponding network. The picture in the diagram only activates three rows, namely the three rows of turret, body and track.

[0099] The above simulation experiments show that the method enhances the interpretability of the model by extracting semantic primitives to resolve semantic diversity, constructing a semantic association graph to embed the multi-level semantic relationship of the target, and designing a reasoning method based on graph activation, so that the machine can, under the guidance of knowledge, draw on the knowledge association mechanism of the brain to achieve interpretable small-sample target recognition. Specifically, this method obtains primitives with high-level semantic information through instance segmentation, uses the method of the present invention to generate a feature matrix, and uses the knowledge graph composed of prior knowledge to perform reasoning in the graph convolutional network. At the same time, experiments have proved that this method is superior to existing recognition models both in terms of interpretability and task performance in small sample scenarios. It is an efficient and interpretable small-sample target detection and recognition method.

[0100] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature identified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0101] Although the present application is described herein with reference to various embodiments, those skilled in the art will be able to understand and implement other variations of the disclosed embodiments in practicing the claimed application by reviewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality.

[0102] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A knowledge-guided small sample target detection and recognition method, characterized in that: include: Step 1: Obtain the target image to be classified; Step 2: Based on prior knowledge, all basic primitives that can represent the high-level semantic features of the target in the target image are pre-planned to form the primitive space of the entire target; Step 3: Using the basic primitives in the primitive space to guide instance segmentation of each target image at the semantic level, obtaining superpixel blocks corresponding to pixels on the basic primitives, and using each superpixel block as a real scene primitive of the target image; Step 4: Extract features of the real scene primitives of the target image through a preset residual neural network, extract the feature vector of the previous layer of the residual neural network output layer, use the feature vector as the feature vector of the real scene primitive, and construct a feature matrix of the target image based on the feature vector and the number of basic primitives in the primitive space; Step 5: Based on the prior knowledge of basic primitives and semantic information and the prior knowledge of logical associations between basic primitives, a logical association graph between semantics and basic primitives is established to obtain a knowledge graph. Step 6: Using the knowledge graph and the feature matrix of each target image, perform category inference and prediction on the feature matrix corresponding to the target image through the trained graph convolutional network to obtain the classification result of the target image.

2. The knowledge-guided small sample target detection and recognition method according to claim 1, characterized in that: Step 3 includes: Step 31: Use the trained SOLO2 algorithm network to generate the instance mask of S^2 channels corresponding to the S×S grid; Step 32: Use the basic primitives in the primitive space to guide the instance mask to perform instance segmentation on the target image at the semantic level, obtain superpixel blocks corresponding to pixels on the basic primitives, and use each superpixel block as the real scene primitive of the target image.

3. The knowledge-guided small sample target detection and recognition method according to claim 1, characterized in that: Step 4 includes: Step 41: constructing a residual neural network, and training the residual neural network to obtain a trained residual neural network; Step 42: Using the trained residual neural network to extract features from the real scene primitives of each target image, so as to represent the real scene primitives as 1×D feature vectors; Step 43: Initialize an all-zero matrix according to the number N of basic primitives in the primitive space and the length D of the eigenvector; Step 44: Fill the eigenvectors in the corresponding dimensions of the all-zero matrix to obtain the eigenmatrix X i .

4. The knowledge-guided small sample target detection and recognition method according to claim 1, characterized in that: Step 5 includes: Step 51: mapping the basic primitives and their semantic information based on the prior knowledge of the basic primitives and semantic information in the primitive space; Step 52: Summarize the same basic primitives and form semantic association pairs corresponding to the basic primitives; Step 53: Construct a knowledge graph based on the semantic association pairs and the topological relationships between the basic primitives.

5. The method for detecting and recognizing small sample targets based on knowledge guidance according to claim 1, characterized in that: Step 6 includes: Step 61: For a graph structure data with N nodes in the knowledge graph, each of which stores D-dimensional features, a feature matrix X with a dimension of N×D can be obtained by highlighting the nodes in the knowledge graph; Among them, the adjacency relationship between each node constitutes an adjacency matrix A with a dimension of N×N; Step 62: Using the feature matrix X and the adjacency matrix A as inputs to the trained graph convolutional network, and updating the feature matrix X based on the propagation formula between layers of the graph convolutional network; Step 63: Based on the updated feature matrix X, the classification result is output by the trained graph convolutional network, and the classification result of the target image is predicted based on the classification result, thereby realizing knowledge-driven interpretable small-sample reasoning.

6. The knowledge-guided small sample target detection and recognition method according to claim 5, characterized in that: The propagation formula between graph convolutional network layers can be expressed as: in, I represents the identity matrix, yes The degree matrix, W () is the learnable weight of the lth layer in the graph convolutional network, H () is the feature matrix of the lth layer in the graph convolutional network. When l = 0, H () That is, the input X, σ is a nonlinear activation function.

7. The knowledge-guided small sample target detection and recognition method according to claim 6, characterized in that: Step 63 includes: Step 631: The first layer in the graph convolutional network performs a graph convolution operation on the input feature matrix X and outputs a feature matrix H using a ReLU activation function. Among them, the output of the first layer is The ReLU activation function is: f(x) = max(0,x); Step 632: The second layer performs a graph convolution operation on the feature matrix H and outputs the result using the softmax activation function; The output of the second layer is: Among them, the softmax activation function is: Step 633: Determine the inference formula of the graph convolutional network based on the output of the first layer and the output of the second layer; The reasoning formula is: Among them, Z is the output result of the graph convolutional network, f represents the inference function of the network, ReLU and softmax represent activation functions; Step 634: Use the inference formula to infer the classification result output by the graph convolutional network, and predict the classification result of the target image based on the classification result, thereby realizing knowledge-driven explainable small sample reasoning.

Images