A tea disease detection system based on non-complete update hypergraph

Abstract

This invention discloses a tea disease detection system based on a partially updated hypergraph, comprising the following steps: acquiring and preprocessing images of diseased tea leaves; extracting features from the images and constructing a node set through a channel slicing module; using a detection network based on a partially updated hypergraph to detect tea diseases, constructing two scale hypergraphs containing high-order relationships of features, selecting the top K groups of nodes and hyperedges with the strongest correlation to the node to be updated for node update, and updating the nodes by aggregating and connecting the original node information through hyperedges; fusing tea disease features from the multi-scale hypergraphs, and finally displaying the detection results through an interactive interface. This invention relates to the field of image target detection. This tea disease detection system based on a partially updated hypergraph overcomes the limitation of existing methods in establishing high-order relationships between channel features, thus improving the accuracy of tea disease identification.

Description

Technical Field

[0001] This invention relates to the field of image processing, specifically to a tea disease detection system based on a partially updated supergraph. Background Technology

[0002] As one of the world's most popular beverages, tea's yield and quality directly impact the economic benefits of tea farmers and the health of consumers. Detecting and identifying tea diseases helps in taking timely preventative measures and reducing losses caused by these diseases.

[0003] Traditional deep learning methods for image classification focus on spatial features, but fall short in processing deeper, higher-order semantic information, making it difficult to establish higher-order relationships between feature channels. Leveraging the advantages of hypergraphs in handling higher-order relationships, this paper aggregates multiple semantically related disease feature nodes (such as texture and contour) through hyperedges, characterizing the complex higher-order dependencies between feature channels and spatial regions. Furthermore, constructing hypergraphs at different scales to capture multi-level feature representations is crucial for detecting tea disease features. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing tea disease detection methods, this invention provides a tea disease detection system based on a partially updated hypergraph, which solves the problem that existing methods struggle to establish high-order relationships of leaf characteristics and improves the accuracy of tea disease detection.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the present invention provides the following technical solution: a tea disease detection system based on a partially updated hypergraph, comprising the following steps:

[0008] S1: Use the domestically produced embedded development board Orange Pi device with an external visible light camera to collect images of tea leaves with diseases;

[0009] S2: Preprocess the original image, including cropping, resizing, and data augmentation;

[0010] S3: Input the processed image into a feature extraction network for feature extraction. The feature extraction network can be VGG, ResNet, DenseNet, etc. The feature extraction network is a ResNet50 network. ResNet50 is a deep convolutional neural network based on a residual network architecture. It can be divided into an input processing part, multiple residual block groups and an output layer, which progressively downsamples and extracts deep features from the feature map. The input layer of the ResNet50 network receives the preprocessed image and achieves fast and efficient feature extraction through 7×7 convolutional kernels. It also obtains richer feature representations by increasing the number of layers, and its residual blocks contain multiple convolutional layers. Skip connections ensure that gradients can flow during backpropagation. A channel slicing module performs equally spaced slicing operations along the channel axis on the extracted features, preserving channel repetition between adjacent slices to form multiple sub-slices. These sub-slices are then dimensionality-reduced and compressed into nodes. Hypergraph initialization is performed first, followed by the construction of 2-node and 4-node hypergraphs. A partially updated hypergraph module optimizes the hypergraph update method. This module includes a node partially updated module and a hyperedge partially updated module, which respectively implement node updates and hyperedge updates.

[0011] S4: Spatial attention is used to fuse the features obtained from the 2-node hypergraph and the 4-node hypergraph to finally output the disease category of the leaf and the accuracy. The results are displayed on the interactive interface.

[0012] Preferably, when collecting diseased tea leaves in step S1, ensure that the interface between the Orange Pi development board and the external visible light camera is compatible. The interface includes USB and MIPI, and corresponding connection cables are configured. When collecting outdoors, ensure sufficient light. Keep the camera and tea tree relatively still during shooting to ensure that the leaves are fully presented in the viewfinder without edge cropping or blurring. The collected samples cover six categories: leaf blight, red leaf spot, tea black mold, anthracnose, white spot disease, and healthy leaves. The number of samples in each category is balanced, with at least 100 images for each category.

[0013] Preferably, in step S2, when cropping the original image, the cropping frame should be square, the image of the diseased leaf should be extracted, and then the image should be scaled proportionally to a size of 224×224 pixels to achieve the best classification effect; data augmentation strategies are used to expand the dataset, including five methods: random cropping, color change, vertical flipping, Gaussian blur and random noise addition, to reduce the impact of data imbalance.

[0014] Preferably, the method also includes model training: the collected images are divided into training set, validation set and test set in a ratio of 8:1:1. The training set is expanded using the five data augmentation strategies and then used to train the tea classification network. The network performance is evaluated and the model parameters are optimized through the validation set and test set.

[0015] Preferably, in step S3, the slicing parameters of the channel slicing module satisfy S=256 and R=64, where S is the number of slice channels and R is the repetition factor, representing the number of repeated channels. During hypergraph initialization, the channel slice blocks are averaged to obtain initial nodes, and the average cosine similarity between nodes is calculated as the initial hyperedge. The node incomplete update module selects the top K node groups with the strongest spatial dependencies to participate in the update based on the hyperedge information of the previous layer, and freezes the node groups with weak correlations and does not participate in this update, where K=3. The node update process is as follows: the node groups connected by each hyperedge are fused based on their respective hyperedge information to obtain a fused node; then each fused node is spliced ​​and sent to the node update network fv to obtain the information of neighboring nodes; finally, the node itself is fused, and the original information is retained to obtain the updated node.

[0016] Preferably, in step S3, the hyperedge update process is as follows: First, the average value of the absolute distance between the corresponding nodes is calculated, and the average value is fed into the metric network to generate the initial weight value to obtain the hyperedge similarity. The metric network consists of 4 convolutional layers, 1 linear layer and a sigmoid activation function. The initial weight value is weighted with the hyperedge information of the previous layer, and the updated hyperedge information is obtained after normalizing all hyperedge information.

[0017] Preferably, after obtaining the updated hyperedge information, the updated node features are remapped to the spatial dimension, and the global features are aggregated through an attention mechanism. Then, the features obtained from the 2-node hypergraph and the 4-node hypergraph are fused using spatial attention.

[0018] Preferably, in step S4, when fusing node features of two scale hypergraphs, the dimensions of the two should be kept consistent in order to better integrate the complementary advantages of multi-scale hypergraphs and construct a multi-level feature expression system; the output results of the embedded development board NPU processor, including the types of tea diseases and detection accuracy, are statistically analyzed and displayed on the touch screen of the Orange Pi development board. Attached Figure Description

[0019] Figure 1 This is a flowchart of the disease detection process of the present invention;

[0020] Figure 2 This is a framework diagram for model training and detection in this invention;

[0021] Figure 3 This is a framework diagram of the channel slicing module of the present invention.

[0022] Figure 4 This is a framework diagram of the incomplete update module of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figure 1-4 The present invention provides a technical solution:

[0025] like Figure 2 As shown, model training and testing are completed in the following steps:

[0026] The first step is to collect training images: use an external visible light camera in a wild tea garden to photograph the leaves of diseased tea leaves, collecting images of six types of leaves, including those with cloud-patterned leaf blight, red leaf spot, tea black mold, anthracnose, white spot disease, and healthy leaves, with at least 100 images for each type.

[0027] The second step, optimal network training: The original images are cropped, retaining key information about the leaves. The images are then compressed to a uniform size of 224×224 pixels and divided into training, validation, and test sets in an 8:1:1 ratio. Before training, the training set is augmented using five data augmentation strategies, including color modification and random noise addition. The augmented training set is used to train the tea classification network. The network's performance is evaluated by validating and testing it using the defined validation and test sets. The specific steps for training the network are as follows:

[0028] (1) Tea Disease Feature Extraction Module: The feature extraction network can be VGG, ResNet, DenseNet, etc., and the feature extraction network is ResNet50. ResNet50 is a deep convolutional neural network based on a residual network architecture. It can be divided into an input processing part, multiple residual block groups, and an output layer, which progressively downsamples and extracts deep features from the feature map. The input layer is a 224×224 RGB image, which is used to achieve fast and efficient feature extraction through 7×7 convolutional kernels. At the same time, richer feature expressions are obtained by increasing the number of layers. There are multiple convolutional layers in the residual block for feature extraction and dimensionality changes. At the same time, the input of the block is directly added to the output of the block through skip connections to ensure that the gradient can flow in backpropagation. Finally, the output layer outputs the features. Where C is the number of channels, and H and W represent the height and width of the feature, respectively.

[0029] (2) Construct the channel slicing module. The output of the feature extraction network is... The feature is sliced ​​at equal intervals along the channel axis, with a certain amount of channel overlap maintained between adjacent slices. A schematic diagram of channel slicing is shown below. Figure 3 As shown. This operation divides the original channel into several... n Individual slices Each sub-slice block corresponds to different feature information, among which S This represents the number of channels in the slice. R S is the repetition factor, representing the number of repeated channels. In the optimal model, S=256 and R=64. Channel slicing cuts continuous, globally coupled feature channels into discrete fragments. Each fragment carries a local feature pattern, separating potentially independent units from the original channel vectors. The repetition slicing strategy can establish connections between nodes, allowing nodes to share and maintain certain feature regions, forming the basis for higher-order relationships between nodes.

[0030] (3) Constructing the Incompletely Updated Hypergraph Module. Channel slices are dimensionality-reduced and compressed to serve as nodes. Based on the number of nodes connected by each hyperedge, 2-node and 4-node hypergraphs are constructed respectively. First, the hypergraph is initialized by performing average pooling on channel slices to serve as initial nodes; the average cosine similarity between nodes is calculated as the initial hyperedge connecting that node group. The incompletely updated hypergraph module includes a node incomplete update module and a hyperedge incomplete update module. The update process is as follows: Figure 4 As shown. The incomplete node update module prioritizes the top K node groups with strong spatial dependencies for update based on the hyperedge information from the previous layer, while freezing node groups with weak dependencies and excluding them from the update (K=3 in the optimal model). The node update process consists of the following three steps: First, for the selected node groups and hyperedges, the node groups connected by each hyperedge are merged based on their respective hyperedge information to obtain a merged node; then, each merged node is concatenated and sent to the node update network. fv In the process, information about neighboring nodes is obtained; finally, the information of the node itself is merged, retaining the original information, to obtain the updated node; the hyperedge update is divided into the following three steps: first, the absolute distances between the corresponding nodes are summed and averaged; then, the data is fed into a system consisting of 4 convolutional layers, one linear layer, and... sigmoid Initial weights are generated in the metric network constructed from activation functions to obtain hyperedge similarity. Finally, the hyperedge similarity is weighted with the hyperedge information from the previous layer, and all hyperedge information is normalized to obtain updated hyperedge information. After the nodes and hyperedges are updated, the updated node features are remapped to the spatial dimension. Global features are aggregated through an attention mechanism, and the features of the two scale hypergraphs are fused using spatial attention. The tea disease identification results are then output through a linear layer.

[0031] The third step is the tea disease detection function: the captured images of tea leaves with diseases are input into the trained tea disease detection network, the tea disease feature information is extracted, the features are updated through multi-scale hypergraph, and finally the category of the disease is output and the detection results are displayed on the touch screen of the Orange Pi development board.

[0032] Based on the above steps, such as Figure 1 As shown, the specific operation steps of a tea disease detection system based on a partially updated hypergraph are as follows:

[0033] S1: Use the domestically produced embedded development board Orange Pi, equipped with an external visible light camera, to collect images of tea tree diseased leaves;

[0034] S2: Preprocess the original image, including cropping, resizing, and data augmentation;

[0035] S3: The image is fed into the feature extraction network for feature extraction. The obtained feature blocks are used as nodes through the channel slicing module to construct 2-node and 4-node hypergraphs respectively. An incomplete update hypergraph module is designed to optimize the hypergraph update method.

[0036] S4: Spatial attention fusion is performed on the features obtained from the two scale hypermaps to finally output the disease category of the leaf and the accuracy. The results are displayed on the interactive interface of the touch screen.

[0037] In this invention, during the entire network optimization process, the network gradually learns the multi-scale hypergraph features and global context information of tea diseased leaves. When the model finds the optimal network model through continuous optimization, the entire network can serve as a tool for detecting six types of tea diseases. At this point, the model possesses the ability to detect tea diseases.

[0038] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A tea disease detection system based on a partially updated hypergraph, characterized in that: Includes the following steps: S1: Use the domestically produced embedded development board Orange Pi device with an external visible light camera to collect images of tea leaves with diseases; S2: Preprocess the original image, including cropping, resizing, and data augmentation; S3: The processed image is input into a feature extraction network for feature extraction. The feature extraction network can be VGG, ResNet, or DenseNet, and the feature extraction network is a ResNet50 network. The input layer of the ResNet50 network receives the preprocessed image and achieves fast and efficient feature extraction through 7×7 convolutional kernels. At the same time, it obtains richer feature representations by increasing the number of layers, and its residual blocks contain multiple convolutional layers. Skip connections ensure that the gradient can flow during backpropagation. The extracted features are sliced ​​at equal intervals along the channel axis through a channel slicing module, and channel repetition is retained between adjacent slice blocks to form multiple sub-slice blocks. The sub-slice blocks are dimensionality reduced and compressed as nodes. First, a hypergraph is initialized, and then 2-node hypergraphs and 4-node hypergraphs are constructed respectively. The hypergraph update method is optimized through a non-complete update hypergraph module, which includes a node non-complete update module and a hyperedge non-complete update module to realize node update and hyperedge update respectively. S4: Spatial attention is used to fuse the features obtained from the 2-node hypergraph and the 4-node hypergraph to finally output the disease category of the leaf and the accuracy. The results are displayed on the interactive interface of the touch screen. In step S3, the slicing parameters of the channel slicing module satisfy S=256 and R=64, where S is the number of slice channels and R is the repetition factor, representing the number of repeated channels; during hypergraph initialization, the channel slice blocks are averaged to obtain the initial nodes, and the average cosine similarity between the nodes is calculated as the initial hyperedge. The incomplete node update module selects the top K node groups with the strongest spatial dependencies to participate in the update based on the hyperedge information from the previous layer, while freezing node groups with weaker dependencies and excluding them from this update, where K=3. The node update process is as follows: each node group connected by a hyperedge is merged based on its own hyperedge information to obtain a merged node; then, each merged node is concatenated and sent to the node update network f. v In the process, information about neighboring nodes is obtained; finally, the information about the node itself is merged, retaining the original information, to obtain the updated node; In step S3, the hyperedge update process is as follows: First, the average value of the absolute distance between the corresponding nodes is calculated, and the average value is fed into the metric network to generate the initial weight value to obtain the hyperedge similarity. The metric network consists of 4 convolutional layers, 1 linear layer and sigmoid activation function. The initial weight value is weighted with the hyperedge information of the previous layer, and the updated hyperedge information is obtained after normalizing all hyperedge information.

2. The tea disease detection system based on a partially updated hypergraph according to claim 1, characterized in that: When collecting diseased tea leaves in step S1, ensure that the interface between the Orange Pi development board and the external visible light camera is compatible. The interface includes USB and MIPI. Configure the corresponding connection cable. When collecting outdoors, ensure sufficient light. Keep the camera and tea tree relatively still during shooting to ensure that the leaves are fully displayed in the viewfinder without edge cropping or blurring. The collected samples cover six categories: leaf blight, red leaf spot, tea black mold, anthracnose, white spot disease, and healthy leaves. The number of samples in each category is balanced, with at least 100 images for each category.

3. The tea disease detection system based on a partially updated hypergraph according to claim 2, characterized in that: In step S2, when cropping the original image, the cropping box should be square, and the image of the diseased leaf should be extracted. Then, the image is scaled proportionally to a size of 224×224 pixels to achieve the best classification effect. Data augmentation strategies are used to expand the dataset, including five methods: random cropping, color change, vertical flipping, Gaussian blur, and random noise addition, to reduce the impact of data imbalance.

4. The tea disease detection system based on a partially updated hypergraph according to claim 3, characterized in that: It also includes model training: the collected images are divided into training set, validation set and test set in a ratio of 8:1:

1. The training set is expanded using the five data augmentation strategies mentioned above and used to train the tea classification network. The network performance is evaluated and the model parameters are optimized through the validation set and test set.

5. The tea disease detection system based on a partially updated hypergraph according to claim 1, characterized in that: After completing the node and hyperedge updates, the updated node features are remapped to the spatial dimension. After aggregating the global features through an attention mechanism, the features obtained from the 2-node hypergraph and the 4-node hypergraph are fused using spatial attention.

6. The tea disease detection system based on a partially updated hypergraph according to claim 1, characterized in that: In step S4, when fusing node features of hypergraphs at two scales, the dimensions of the two hypergraphs should be kept consistent in order to better integrate the complementary advantages of multi-scale hypergraphs and construct a multi-level feature representation system. The output results of the embedded development board's NPU processor, including the types of tea diseases and the detection accuracy, are statistically analyzed and displayed on the touchscreen of the Orange Pi development board.

Images