Tea disease and pest classification system based on space-frequency mixed node graph neural network

Abstract

The invention discloses a tea plant disease and insect pest classification system based on a space-frequency mixed node graph neural network, and the system comprises the following steps: obtaining a to-be-extracted field tea plant disease and insect pest image, and transmitting the collected image to a processor for processing, adding an optimal model obtained by training under the catalog of the system and loading the processed tea disease and insect pest image; a final classification result is obtained by utilizing a classification function in the system; and printing a final classification result in a display window. Wherein the image is sent into the system to extract RGB image features and gray level image features, RGB frequency domain features and gray level frequency domain features construct space-frequency mixed nodes, and tea disease and insect pest image features are extracted through a ResNet12 network. The invention relates to the field of image processing, the space-frequency hybrid nodes improve the effectiveness in the aspect of feature representation, the hybrid edge features improve the accuracy of distance measurement of the adjacent space-frequency hybrid nodes, and efficient disease and pest diagnosis and classification are realized.

Description

Technical Field

[0001] The invention relates to the technical field of image processing, and in particular to a tea pest and disease classification system based on a space-frequency hybrid node graph neural network. Background Art

[0002] Tea is an important agricultural cash crop with high nutritional and economic value. However, tea leaves are very susceptible to diseases and pests such as anthracnose, red leaf spot, black mold, white spot disease, and tea geometrid, which seriously reduce tea yield and quality. Therefore, efficient and accurate classification of tea pests and diseases is the key to preventing and controlling tea pests and diseases. Traditional manual methods for identifying plant pests and diseases have problems such as low efficiency, low recognition accuracy, and limited processing range, and have failed to effectively curb the spread of pests and diseases. With the advancement of science and technology, automatic classification systems based on deep learning and computer vision have gradually become a research hotspot for tea pest and disease classification. These systems can not only greatly improve classification efficiency and accuracy, but also realize pest and disease monitoring in large-scale tea gardens, and provide timely prevention and control suggestions for tea farmers, thereby effectively reducing pest and disease losses and improving tea yield and quality. To address this problem, a tea pest and disease classification system based on space-frequency hybrid node graph neural network was developed. The system aims to use the graph neural network algorithm to enable the system to more accurately classify the types of tea pests and diseases, provide tea farmers with real-time pest and disease warnings and accurate prevention and control recommendations, and thus improve the quality and yield of tea production. Summary of the invention

[0003] 1. Technical issues to be resolved

[0004] In view of the deficiencies in the prior art, the present invention provides a tea pest and disease classification system based on a space-frequency hybrid node graph neural network.

[0005] (II) Technical solution

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: A tea pest and disease classification system based on a space-frequency hybrid node graph neural network, comprising the following steps:

[0007] S1: Obtain on-site tea pest and disease images to be classified;

[0008] S2: Send the collected image to the CPU processor for image processing;

[0009] S3: Add the optimal model obtained through training and load the processed tea pest and disease images in the directory of the system;

[0010] S4: Use the classification function in the system to obtain the final classification result;

[0011] S5: Print the final classification result in the display window;

[0012] The training method used for the optimal model in step S3 includes the following steps:

[0013] The first step is image cropping and processing;

[0014] The second step is optimal network training: the obtained dataset of at least 350 tea pests and diseases is divided into training set, validation set and test set; the training set is augmented before training, and all tea pests and diseases images are filled with square shapes, and then scaled to 128×128; the augmented and preprocessed training set is used to train the graph neural network, and the divided validation set and test set are used to verify and test the network to evaluate the performance of the network;

[0015] Among them, the specific steps of the second step of graph neural network training are as follows:

[0016] (1) Construct a backbone network for extracting tea pests and diseases features: Tea pests and diseases image features are extracted through the ResNet12 network. Resnet12 contains 4 residual blocks, each with 3 convolutional layers. "12" means there are 12 convolutional layers in total; since the characteristics of different tea pests and diseases studied have a large degree of distinction and are not very complex, a deeper network may lead to excessive attention to certain features, resulting in the neglect of other important features, and the use of ResNet12 can avoid this situation. The backbone network is composed of four independent convolutional neural networks ResNet12, which are used to extract the following features: extract RGB tea pests and diseases features from RGB images, extract grayscale tea pests and diseases features from GRAY images, extract RGB frequency domain tea pests and diseases features from the frequency domain images of RGB images, and extract grayscale frequency domain tea pests and diseases features from the frequency domain images of grayscale images; perform feature splicing on images in four different domains, and initialize the space-frequency mixed features of pests and diseases, so that the graph neural network can aggregate more space-frequency mixed node information when updating the space-frequency mixed nodes; use the relative metric information between space-frequency mixed nodes to form an edge between nodes, called the relative metric edge e ij1 , represents the first dimension edge feature between node i and node j, and the similarity edge e between nodes ij2 , dissimilarity edge e ij3 The mixed edges of the splicing graph are constructed, where e ij2 and e ij3 Respectively represent the intra-class and inter-class relationships of nodes;

[0017] (2) Constructing a space-frequency mixing node update module: Each space-frequency mixing node in the l-1th layer updates the space-frequency mixing node features by mixing edges to aggregate the features of adjacent space-frequency mixing nodes. The updated space-frequency mixing node features are used as the space-frequency mixing node features in the lth layer. Indicates the update of the space-frequency mixed node features constituting the RGB sub-node, grayscale sub-node, RGB frequency domain sub-node and grayscale frequency domain sub-node, It represents the space-frequency mixed node features of the l-1th layer, which consists of RGB sub-nodes, grayscale sub-nodes, RGB frequency domain sub-nodes, and grayscale frequency domain sub-nodes. It represents the space-frequency mixed node features of the l-th layer, which consists of RGB sub-nodes, gray sub-nodes, RGB frequency domain sub-nodes, and gray frequency domain sub-nodes. Indicates that the space-frequency mixed node aggregates the features of adjacent space-frequency mixed nodes connected by the mixed edge features; and Represents the mixed edge features between space-frequency mixed nodes at the l-1th layer. Since the mixed edge features can represent the similarity between space-frequency mixed nodes, by aggregating the features of adjacent space-frequency mixed nodes, the similarity of nodes within a class and the difference of nodes between classes can be fully utilized to improve the representation ability of node features.

[0018] (3) Constructing the mixed edge update module: The mixed edges of the graph structure are composed of relative metric edges and similarity edges. For the relative metric edges, the space-frequency mixed nodes Subtract the offset to get the relative metric node The i-th similarity node minus the i+1-th similarity node gets the relative metric node. When i=N×(K+q), the last similarity node minus the first similarity node gets the relative metric node. Represents a relative metric node, which is used to update the relative metric edge according to Get relative metric node After that, the distance metric encoding is performed on the relative metric nodes to obtain high-order relative metric nodes. Finally, the distance encoding is mapped into the features representing the high-order relative metric between space-frequency hybrid nodes through the MLP module, thereby realizing the relative metric edge Feature update, where N and K are N-wayK-shot, and q is the number of query set samples; represents the updated similarity edge, The updated space-frequency hybrid node features The distance measurement feature mapping is obtained, and the mixed edge update process aggregates the different domain information between space-frequency mixed nodes. After feature mapping, the relative measurement edge and the similarity edge aggregate the multi-dimensional information between space-frequency mixed nodes, and contain the distance measurement information of RGB sub-nodes, grayscale sub-nodes, RGB frequency domain sub-nodes and grayscale frequency domain sub-nodes as well as the edge information of the previous layer of graph, which improves the measurement ability of the mixed edge feature for the same type of space-frequency mixed nodes and the distinction ability between different types of space-frequency mixed nodes.

[0019] As an optimization, the data preprocessing in step S1 includes two steps: cropping of pest and disease images and data augmentation: since the resolution of the collected tea pest and disease images is too large, the speed of sending them to the network preprocessing is slow, and the images contain multiple leaves, which makes it impossible to accurately detect the desired pest and disease areas. Therefore, the tea pest and disease images need to be preprocessed; in order to obtain the required images that can be input into the network, the captured images are cropped into a single pest and disease leaf as an image. Since the number of collected pest and disease images is small, the classification results of the model can be improved after data expansion.

[0020] As an optimization, all image shapes are filled with squares during preprocessing in step S2, and their sizes are proportionally scaled to 128×128 to achieve the best detection effect.

[0021] As an optimization, step S3 first crops the multi-leaf images in the first step to remove irrelevant leaves; secondly, for the single leaf image captured by the camera, if the background noise is large, the background is cropped; finally, different types of tea pests and diseases are identified and classified; since the number of pest and disease images collected is small, one or several data enhancement methods including adding salt and pepper noise, Retinex data enhancement, and random rotation are used for data enhancement, and the classification results of the model can be improved after data expansion.

[0022] As an optimization, step S3 divides the data into a training set, a validation set, and a test set in a ratio of 6:2:2 in the second step, wherein all samples in the training set are trained to obtain a model, and the augmented images are added to the training set. The purpose of this is to expand the diversity of the data so that the effect of the final model can be optimized; the role of the training set is to test the current model to judge the quality of the model, and to achieve the optimal network model through continuous testing.

[0023] As an optimization, the final classification result of step S4 is based on the graph structure containing multiple categories of pests and diseases. The true edge label of the graph structure is the probability that the i-th node and the j-th node belong to the same category. The edge label of the same node is 1, and the edge label of the different node is 0. During the test, the query set and the support set nodes constitute the predicted edge label. Since the graph structure contains multiple categories, the probability that the query set node belongs to a certain category can be judged by comparing the closeness of the predicted edge label with the true edge label. Specifically, the edge label will be calculated between the query set node and the node of each category, and compared with the edge label of the category. For a certain category, the larger the edge label value, the greater the probability that the query set node belongs to the category. Therefore, the category with the maximum predicted probability will eventually be selected as the prediction result of the query set node. The category of pests and diseases is finally obtained by finding the maximum probability of the corresponding category.

[0024] (III) Beneficial effects

[0025] The present invention provides a tea pest and disease classification system based on a space-frequency hybrid node graph neural network, which has the following beneficial effects:

[0026] By combining the spatial and frequency domain features, the system can fully capture the multi-dimensional information of tea pests and diseases, and significantly improve the classification accuracy and robustness. At the same time, the spatial-frequency hybrid node graph neural network enhances the feature expression ability, can deeply extract spatial information and frequency features, and comprehensively classify pest and disease patterns. The system adopts the adaptive learning and optimization strategy of the graph neural network to automatically optimize the network structure and parameters, reduce manual intervention, and achieve efficient diagnosis and classification of pests and diseases. In addition, the system supports real-time early warning and accurate diagnosis to help tea farmers respond to pests and diseases in a timely manner and reduce the impact. Its application is not limited to tea, but can also be extended to the classification of pests and diseases of other crops. It has strong versatility and scalability, thereby promoting the intelligence and precision of agricultural production, and effectively improving the efficiency of crop production, reducing the use of pesticides, reducing production costs, and improving tea quality and yield. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flow chart of the tea pest classification system method based on space-frequency hybrid node graph neural network of the present invention;

[0028] Figure 2 It is a framework diagram of the network optimization and tea plant disease and insect pest algorithm of the present invention;

[0029] Figure 3 This is a framework diagram of a space-frequency hybrid node update module of the present invention;

[0030] Figure 4 This is a framework diagram of the hybrid edge update module of the present invention;

[0031] Figure 5 This is an end-to-end graph of the tea pest and disease classification system based on the space-frequency hybrid node graph neural network of the present invention. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] See also Figure 1-5 , the present invention provides a technical solution:

[0034] like Figure 2 As shown, model training and testing are completed in the following steps:

[0035] The first step is to crop the multi-leaf images and remove irrelevant leaves. Secondly, for the single leaf images captured by the camera, if the background noise is large, the background is cropped. Finally, the categories of different types of tea pests and diseases are identified and classified. Since the number of pest and disease images collected is small, one or more data enhancement methods such as adding salt and pepper noise, Retinex data enhancement, and random rotation are used for data enhancement. After data expansion, the classification results of the model can be improved.

[0036] The second step is optimal network training: divide the obtained 350 tea pest and disease data sets into training set, validation set and test set; augment the training set before training, fill all tea pest and disease image shapes with squares, and then scale their size to 128×128; train the graph neural network with the augmented and preprocessed training set, and use the divided validation set and test set to verify and test the network to evaluate the network performance.

[0037] Among them, the specific steps of the second step of graph neural network training are as follows:

[0038] (1) Construct a backbone network for extracting tea pests and diseases features: Tea pests and diseases image features are extracted through the ResNet12 network. Resnet12 contains 4 residual blocks, each with 3 convolutional layers. "12" means that there are 12 convolutional layers in total; since the different tea pests and diseases features studied have a large degree of distinction and are not very complex, a deeper network may lead to excessive attention to certain features, resulting in the neglect of other important features, and the use of ResNet12 can avoid this situation. The embedding network consists of four independent convolutional neural networks ResNet12, which are used to extract the following features: RGB tea pests and diseases features from RGB images, grayscale tea pests and diseases features from GRAY images, RGB frequency domain tea pests and diseases features from the frequency domain images of RGB images, and grayscale frequency domain tea pests and diseases features from the frequency domain images of grayscale images. Feature splicing is performed on images in four different domains, and the pest and disease space-frequency mixed features are initialized. In order to enable the graph neural network to aggregate more space-frequency mixed node information when updating the space-frequency mixed node, the relative metric information between the space-frequency mixed nodes is used to form an edge between the nodes, which is called the relative metric edge e ij1 , represents the first dimension edge feature between node i and node j, and the similarity edge e between nodes ij2 , dissimilarity edge e ij3 The mixed edges of the splicing graph are constructed, where e ij2 and e ij3 They represent the intra-class and inter-class relationships of nodes respectively.

[0039] (2) Constructing the space-frequency mixing node update module: Each space-frequency mixing node in the 1-1 layer updates the space-frequency mixing node features by mixing the edge aggregation of the adjacent space-frequency mixing node features, and the updated space-frequency mixing node features are used as the space-frequency mixing node features of the 1st layer. Indicates the update of the space-frequency mixed node features constituting the RGB sub-node, the grayscale sub-node, the RGB frequency domain sub-node and the grayscale frequency domain sub-node. Indicates the spatial-frequency mixed node features of the 1-1 layer, which consists of RGB sub-nodes, grayscale sub-nodes, RGB frequency domain sub-nodes, and grayscale frequency domain sub-nodes. The first layer consists of space-frequency mixed node features of RGB sub-nodes, grayscale sub-nodes, RGB frequency domain sub-nodes and grayscale frequency domain sub-nodes. It represents that the space-frequency mixing node aggregates the adjacent space-frequency mixing node features connected by the mixing edge features. and It represents the mixed edge features between space-frequency mixed nodes in the l-1th layer. Since the mixed edge features can represent the similarity between space-frequency mixed nodes, by aggregating the features of adjacent space-frequency mixed nodes, the similarity of nodes within a class and the difference of nodes between classes can be fully utilized to improve the representation ability of node features.

[0040] (3) Constructing the mixed edge update module: The mixed edges of the graph structure are composed of relative metric edges and similarity edges. For the relative metric edges, the space-frequency mixed nodes Subtract the offset to get the relative metric node The relative metric node is obtained by subtracting the i+1 similarity node from the i-th similarity node. When i=N×(K+q), the relative metric node is obtained by subtracting the 1st similarity node from the last similarity node. Represents a relative metric node, which is used to update the relative metric edge. Get relative metric node After that, the distance metric encoding is performed on the relative metric nodes to obtain high-order relative metric nodes. Finally, the distance encoding is mapped into the features representing the high-order relative metric between space-frequency hybrid nodes through the MLP module, thereby realizing the relative metric edge Feature update, where N and K are N-wayK-shot, and q is the number of query set samples. represents the updated similarity edge, The updated space-frequency hybrid node features The mixed edge update process aggregates the information of different domains between space-frequency mixed nodes. After feature mapping, the relative metric edge and similarity edge aggregate the multi-dimensional information between space-frequency mixed nodes, and contain the distance metric information of RGB sub-nodes, grayscale sub-nodes, RGB frequency domain sub-nodes and grayscale frequency domain sub-nodes as well as the edge information of the previous graph, which improves the measurement ability of the mixed edge feature for the same type of space-frequency mixed nodes and the ability to distinguish between different types of space-frequency mixed nodes.

[0041] The third step is the tea pest and disease classification function: after the optimal model is obtained through training, the tea pest and disease image classification function becomes the core of the system. When an unknown tea pest and disease image is input into the trained small sample learning network model, the system can quickly and accurately classify the types of pests and diseases through the classification method of the present invention. Specifically, the system inputs the image to be classified into the trained network model, and uses the distance metric function of the model to calculate the similarity between the image and the known pest and disease images. Finally, the system will select the category with the highest similarity to the image to be classified as the classification result to ensure the efficiency and accuracy of pest and disease classification.

[0042] Based on the above steps, Figure 1As shown in the figure, a tea pest classification system based on space-frequency hybrid node graph neural network is shown in the figure. The specific operation steps are as follows:

[0043] S1: Obtain on-site tea pest and disease images to be classified;

[0044] S2: Send the collected image to the CPU processor for image processing;

[0045] S3: Add the optimal model obtained through training and load the processed tea pest and disease images in the directory of the system;

[0046] S4: Use the classification function in the system to obtain the final classification result;

[0047] S5: Print the final classification result in the display window.

[0048] In summary, the implementation of the present invention mainly combines tea pest image information from different domains through graph neural networks, designs space-frequency hybrid node update modules, hybrid edge update modules and multi-level feature aggregation mechanisms to extract features of each channel in different domain images (such as RGB, grayscale, frequency domain, etc.). By designing similarity edges and dissimilarity edges, the system can effectively perform hierarchical learning of nodes within and between classes, and improve the classification ability of pest images. Through this technology, the system can more accurately calculate the similarity between nodes, thereby improving the accuracy of edge similarity calculation for node similarity measurement, and finally classify nodes and optimize classification results. The implementation of the present invention can significantly reduce the manpower and material resources required for artificial pest classification, and provide an intelligent and automated pest classification scheme. During the training process, by performing data enhancement and feature extraction on tea pest images, the model can efficiently learn on less sample data, thereby improving the classification efficiency and accuracy of the system in practical applications. Therefore, the implementation of the present invention not only improves the classification accuracy of tea pests and diseases, but also effectively reduces the workload of manual diagnosis through intelligent classification technology, promotes intelligent pest and disease control in tea planting, and has broad application prospects.

[0049] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0050] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A tea pest and disease classification system based on space-frequency hybrid node graph neural network, characterized by: The following steps are involved: S1: Obtain on-site tea pest and disease images to be classified; S2: Send the collected image to the CPU processor for image processing; S3: Add the optimal model obtained through training and load the processed tea pest and disease images in the directory of the system; S4: Use the classification function in the system to obtain the final classification result; S5: Print the final classification result in the display window; The training method used for the optimal model in step S3 includes the following steps: The first step is image cropping and processing; The second step is optimal network training: the obtained dataset of at least 350 tea pests and diseases is divided into training set, validation set and test set; the training set is augmented before training, and all tea pests and diseases images are filled with square shapes, and then scaled to 128×128; the augmented and preprocessed training set is used to train the graph neural network, and the divided validation set and test set are used to verify and test the network to evaluate the performance of the network; Among them, the specific steps of the second step of graph neural network training are as follows: (1) Construct a backbone network for extracting tea pest and disease features: The backbone network consists of four independent convolutional neural networks ResNet12, which are used to extract the following features: extract RGB tea pest and disease features from RGB images, extract grayscale tea pest and disease features from GRAY images, extract RGB frequency domain tea pest and disease features from frequency domain images of RGB images, and extract grayscale frequency domain tea pest and disease features from frequency domain images of grayscale images; perform feature concatenation on images in four different domains, and initialize the pest and disease space-frequency mixed features so that the graph neural network can aggregate more space-frequency mixed node information when updating the space-frequency mixed node; use the relative metric information between space-frequency mixed nodes to form an edge between nodes, called the relative metric edge e ij1 , represents the first dimension edge feature between node i and node j, and the similarity edge e between nodes ij2 , dissimilarity edge e ij3 The mixed edges of the splicing graph are constructed, where e ij2 and e ij3 Respectively represent the intra-class and inter-class relationships of nodes; (2) Constructing a space-frequency mixing node update module: Each space-frequency mixing node in the l-1th layer updates the space-frequency mixing node features by mixing edges to aggregate the features of adjacent space-frequency mixing nodes. The updated space-frequency mixing node features are used as the space-frequency mixing node features in the lth layer. Indicates the update of the space-frequency mixed node features constituting the RGB sub-node, grayscale sub-node, RGB frequency domain sub-node and grayscale frequency domain sub-node, It represents the space-frequency mixed node features of the l-1th layer, which consists of RGB sub-nodes, grayscale sub-nodes, RGB frequency domain sub-nodes, and grayscale frequency domain sub-nodes. It represents the space-frequency mixed node features of the l-th layer, which consists of RGB sub-nodes, gray sub-nodes, RGB frequency domain sub-nodes, and gray frequency domain sub-nodes. Indicates that the space-frequency mixed node aggregates the features of adjacent space-frequency mixed nodes connected by the mixed edge features; and Represents the mixed edge features between space-frequency mixed nodes at the l-1th layer. Since the mixed edge features can represent the similarity between space-frequency mixed nodes, by aggregating the features of adjacent space-frequency mixed nodes, the similarity of nodes within a class and the difference of nodes between classes can be fully utilized to improve the representation ability of node features. (3) Constructing the mixed edge update module: The mixed edges of the graph structure are composed of relative metric edges and similarity edges. For the relative metric edges, the space-frequency mixed nodes Subtract the offset to get the relative metric node The i-th similarity node minus the i+1-th similarity node gets the relative metric node. When i=N×(K+q), the last similarity node minus the first similarity node gets the relative metric node. Represents a relative metric node, which is used to update the relative metric edge according to Get relative metric node After that, the distance metric encoding is performed on the relative metric nodes to obtain high-order relative metric nodes. Finally, the distance encoding is mapped into the features representing the high-order relative metric between space-frequency hybrid nodes through the MLP module, thereby realizing the relative metric edge Feature update, where N and K are N-wayK-shot, and q is the number of query set samples; represents the updated similarity edge, The updated space-frequency hybrid node features The distance metric feature mapping is obtained, and the mixed edge update process aggregates the different domain information between the space-frequency mixed nodes. After feature mapping, the relative metric edge and the similarity edge aggregate the multi-dimensional information between the space-frequency mixed nodes, and contain the distance metric information of the RGB sub-nodes, grayscale sub-nodes, RGB frequency domain sub-nodes and grayscale frequency domain sub-nodes as well as the edge information of the previous layer.

2. A tea pest classification system based on space-frequency hybrid node graph neural network according to claim 1, characterized in that: The data preprocessing in step S1 includes two steps: cropping of pest and disease images and data augmentation: since the resolution of the collected tea pest and disease images is too large, the speed of sending them to the network for preprocessing is slow, and the images contain multiple leaves, which makes it impossible to accurately detect the desired pest and disease areas. Therefore, the tea pest and disease images need to be preprocessed; in order to obtain the required images that can be input into the network, the captured images are cropped into a single pest and disease leaf as an image.

3. The tea pest classification system based on space-frequency hybrid node graph neural network according to claim 1 is characterized by: During the preprocessing in step S2, all image shapes are filled into squares and then proportionally scaled to 128×128 to achieve the best detection effect.

4. The tea pest classification system based on space-frequency hybrid node graph neural network according to claim 1 is characterized by: In the first step, step S3 first crops the multi-leaf images to remove irrelevant leaves; secondly, for the single leaf image captured by the camera, if the background noise is large, background cropping is performed; finally, different types of tea pests and diseases are identified and classified; data enhancement is performed by adding salt and pepper noise, Retinex data enhancement, and random rotation. One or more data enhancement methods are used.

5. The tea pest classification system based on space-frequency hybrid node graph neural network according to claim 1 is characterized by: In the second step, step S3 divides the data into a training set, a validation set, and a test set in a ratio of 6:2:2, wherein all samples in the training set are trained to obtain a model, and the augmented images are added to the training set.

6. The tea pest classification system based on space-frequency hybrid node graph neural network according to claim 1 is characterized by: At the end of step S4, the classification result is based on the nodes of multiple categories of pests and diseases on the graph structure. The true edge label of the graph structure is the probability that the i-th node and the j-th node belong to the same category. The edge label of the same node is 1, and the edge label of the node of different categories is 0. During the test, the query set and the support set nodes constitute the predicted edge labels. Since the graph structure contains multiple categories, the probability that the query set node belongs to a certain category can be determined by comparing the closeness between the predicted edge labels and the true edge labels. The category with the largest predicted probability is selected as the prediction result of the query set node, and the category of pests and diseases is finally obtained by finding the corresponding category with the largest probability.

Images