Rice blast disease grade intelligent evaluation and disease resistance preliminary screening method, system and equipment

Abstract

The embodiment of the invention provides a rice blast disease grade intelligent evaluation and disease resistance preliminary screening method, system and equipment, and the method comprises the steps: S1, collecting image data of wild rice leaves infected with rice blast, carrying out the instance segmentation of the image data through an improved YOLOv8-Seg segmentation model, obtaining a segmentation result, and carrying out the segmentation of the image data through an improved YOLOv8-Seg segmentation model; the improvement measures comprise a GhostHierachicalNet backbone network, a CAHSFPN neck network and an LSDES Head segmentation head, and the LSDES Head segmentation head is connected with the CAHSFPN neck network and the LSDES Head segmentation head; s2, according to the segmentation result, counting the number and proportion of pixels of diseased leaves and disease areas, and obtaining a disease grade evaluation result in combination with disease resistance identification technical regulations; and S3, screening wild rice materials with strong disease resistance based on the disease grade evaluation results which are repeated for multiple times. The method can effectively improve the accuracy and efficiency of crop disease evaluation, and provides support for large-scale screening of disease resistance of wild rice, thereby promoting the development of rice blast-resistant breeding work.

Description

Technical Field

[0001] The present application relates to the field of crop breeding, and in particular to a method, system and equipment for intelligently evaluating the disease grade of rice blast and initial screening of disease resistance. Background Art

[0002] Rice blast is one of the most serious rice diseases in the world. It not only significantly reduces rice yield and quality, but also causes huge economic losses to farmers. Although traditional chemical control methods can control the disease to a certain extent, long-term use will lead to environmental pollution and increased resistance of pathogens. Therefore, screening and breeding rice varieties with strong disease resistance and genetic improvement through the use of disease-resistant genes is a more economical, environmentally friendly and sustainable strategy for rice blast prevention and control. However, mining rice blast resistance genes from wild rice is a complex and systematic task. Traditional methods are labor-intensive and highly dependent on manual judgment. They are time-consuming, labor-intensive, subjective, and have unstable evaluation results.

[0003] To this end, this application proposes a method, system and equipment for intelligent evaluation of rice blast disease grade and initial screening of disease resistance in response to the application needs of rice blast resistance breeding. This method aims to improve the accuracy and efficiency of the evaluation to support large-scale wild rice disease resistance screening and promote the development of rice blast resistance breeding. Summary of the invention

[0004] The present application provides a method, system and equipment for intelligent evaluation of rice blast disease grade and initial screening of disease resistance. The method can effectively improve the accuracy and efficiency of rice blast disease evaluation, provide support for large-scale wild rice disease resistance screening, and promote the development of rice blast resistance breeding.

[0005] In a first aspect, a method for intelligently evaluating the disease grade of rice blast and initial screening of disease resistance is provided, the method comprising: S1: Collect image data of wild rice leaves infected with rice blast, perform instance segmentation on the image data by improving the YOLOv8-Seg segmentation model, and obtain the segmentation results. The improvement measures include: GhostHierarchicalNet backbone network, CAHSFPN neck network and LSDECS Head segmentation head; S2: Based on the segmentation results, the number of pixels and their proportions of diseased leaves and diseased areas are counted, and the disease grade assessment results are obtained in combination with the technical procedures for disease resistance identification; S3: Based on the results of repeated disease grade assessments, wild rice materials with strong disease resistance were screened.

[0006] It should be understood that YOLOv8-Seg is the latest instance segmentation version of the YOLO (You Only Look Once, YOLO) series. Based on the YOLOv8 architecture, it integrates a segmentation head for generating fine segmentation masks. It combines the advantages of Anchor-Free and Anchor-Based methods, and adopts the Center-based methods of Anchor-Free. By introducing flexible anchor point and offset prediction, it effectively solves the problem of adaptability to wild rice diseased leaves and disease targets with large aspect ratios or large scale differences. Therefore, this application adopts the YOLOv8-Seg model as the benchmark segmentation model, and proposes further enhancements based on this model in response to the challenges faced.

[0007] In combination with the first aspect, in some implementations of the first aspect, the GhostHierarchicalNet backbone network includes: HierarchicalStem module,HierarchicalStem module extracts global and local features through a dual-branch structure; GHGBlock module,The GHGBlock module consists of multiple series-connected Ghost convolutional layers, each of which uses the same convolution kernel size and activation function to gradually refine and enhance the input features; Depthwise separable convolution, depthwise separable convolution includes depthwise convolution and pointwise convolution, which reduces the computational complexity.

[0008] It should be understood that the HierarchicalStem module is a lightweight dual-branch feature extraction module, which consists of five convolutional layers and a maximum pooling layer, and uses a dual-branch structure for feature extraction and downsampling. This module uses a smaller convolution kernel size and a smaller number of channels, which significantly reduces the number of parameters and computational complexity, while fully extracting global and local features.

[0009] It should be understood that the GHGBlock module utilizes the efficiency of Ghost convolution to generate features through linear operations, which significantly improves computational efficiency while maintaining model accuracy, and captures feature information at different levels by splicing multiple layers of output.

[0010] It should be understood that depthwise separable convolution achieves efficient feature extraction by decomposing standard convolution into depthwise convolution and pointwise convolution. Depthwise convolution assigns an independent convolution kernel to each channel of the input feature map, allowing each channel to perform convolution operations independently; then, pointwise convolution integrates features in the channel dimension. This design greatly reduces redundant calculations, reduces the number of parameters and computational complexity, and ensures that the number of channels of the input and output feature maps is consistent.

[0011] In combination with the first aspect, in some implementations of the first aspect, the CAHSFPN neck network includes: BiDirectionalFusion module,BiDirectionalFusion module expands the receptive field through horizontal and vertical convolution; C2f_Faster_CGLU module,The C2f_Faster_CGLU module combines the convolutional gated linear unit (CGLU) with partial convolution to achieve lightweight feature fusion.

[0012] It should be understood that the BiDirectionalFusion module adopts a strategy of combining horizontal and vertical convolutions to improve the efficiency of multi-scale feature fusion. Through this combination of horizontal and vertical convolutions, the module maintains the relative stability of computational complexity and model size while expanding the receptive field. Compared with the Neck structure of YOLOv8-Seg, the BiDirectionalFusion module shows advantages in computational cost and feature fusion efficiency by reducing redundant calculations and reducing the number of parameters.

[0013] It should be understood that the C2f module introduced by YOLOv8-Seg adds additional Bottleneck structures and cross-layer connections, thereby enhancing the gradient flow. However, these improvements also lead to more convolution operations and higher computational load. The C2f_Faster_CGLU module proposed in this paper consists of multiple Faster_Block_CGLUs. Each Faster_Block_CGLU combines convolutional gated linear units and partial convolutions to achieve efficient feature mixing and conversion.

[0014] It should be noted that the convolutional gated linear unit combines the convolution operation and the gating mechanism to transform the input features through two different convolution kernels, one of which is used to extract features, and the other generates a gating signal through the Sigmoid activation function to control the degree of information flow. Finally, the two results are multiplied element by element to achieve feature screening and enhancement. Compared with traditional activation functions, convolutional gated linear units allow more refined information regulation and improve the effectiveness of feature extraction. In addition, partial convolution is an improved convolution operation specifically designed to handle missing or incomplete data. Unlike traditional convolution, partial convolution only convolves the valid (non-zero) area during calculation and adjusts it by a normalization factor to avoid information loss.

[0015] In combination with the first aspect, in some implementations of the first aspect, the LSDECS Head segmentation header includes: Shared convolutional layer: Shared convolutional layer is used to reduce redundant calculations of feature maps with different resolutions; Detail enhancement convolution module,The detail enhancement convolution module is used to capture the detail features of lesions.

[0016] It should be understood that shared convolutional layers refer to sharing the same convolution kernel weights between multiple tasks or multiple inputs, thereby reducing the number of parameters and improving computational efficiency. It can extract similar features between different inputs and improve the generalization ability of the model.

[0017] It should also be understood that the detail enhancement convolution module is a deep learning module specifically used to capture and enhance the detail features of the diseased area. This module can effectively extract the edge information, texture features and local details of the diseased area, thereby improving the model's recognition accuracy of the diseased area and ensuring that the disease detection task is still highly robust in complex backgrounds or low-contrast conditions.

[0018] In combination with the first aspect, in certain implementations of the first aspect, the disease grade assessment method in step S2 includes: Divide diseased leaves and spots; Assign the diseased spots to the diseased leaves, and assign each diseased spot to the corresponding diseased leaf by calculating the overlap between the segmentation masks; According to the segmentation results, the number of pixels of each diseased leaf and its corresponding diseased spot is calculated; Calculate the ratio of diseased spot pixels to diseased leaf pixels using the following formula: , in, is the number of pixels of the segmented lesion, is the number of pixels of the segmented diseased leaves, and LCR represents the ratio of diseased spots to diseased leaves.

[0019] It should be understood that this method can accurately quantify the extent of the disease, improve the objectivity and consistency of the assessment, reduce human errors compared to traditional manual assessment methods, and improve the efficiency of large-scale screening of disease-resistant wild rice germplasm resources.

[0020] In combination with the first aspect, in certain implementations of the first aspect, the method for screening wild rice materials with strong disease resistance in step S3 includes: Summarize the results of multiple replicates to evaluate disease resistance: The disease grade evaluation data from multiple replicate tests of the same variety are summarized using the following formula to obtain the results of the disease resistance intelligent evaluation: , in, is the sum of the disease levels in different replicates, N The number of replicates, R Results of the intelligent disease resistance assessment of representative varieties.

[0021] It should be understood that this method can efficiently screen out wild rice materials with excellent disease resistance, provide a reliable screening basis for subsequent disease-resistant variety breeding, while reducing the manpower and time costs required in the screening process and improving screening efficiency and accuracy.

[0022] In combination with the first aspect, in some implementations of the first aspect, a hierarchical adaptive amplitude pruning method is used to optimize the improved YOLOv8-Seg segmentation model.

[0023] It should be understood that the core idea of ​​adaptively adjusting the sparsity of each layer to minimize the impact of pruning on model performance is to consider not only the absolute size of each weight, but also the global distortion of each layer to achieve a more effective pruning process.

[0024] In combination with the first aspect, in some implementations of the first aspect, a federated gradient coordination algorithm is used to train an improved YOLOv8-Seg segmentation model, including: The client calculates the gradient after local training; The server detects gradient conflicts and adjusts the gradient direction by projecting conflicting gradients. Aggregate the optimized gradients to update the global model parameters.

[0025] It should be understood that in the federated learning framework, the client model and the local model have both synergy and characteristics. The client model refers to an individual model deployed on the edge device, which is trained and inferred through local data. Its core feature is that the original data always resides locally, and only the model parameters or gradients are uploaded to the server, achieving the dual advantages of privacy protection and distributed computing. As a specific implementation form of the client model, the local model focuses on participating in the global model optimization through periodic gradient updates. The parameter changes generated by its training are aggregated by the federated gradient coordination algorithm to form a global model update, and then distributed to each client through the server to form a closed-loop federated learning process. The federated gradient coordination algorithm proposed in this application reduces the mutual cancellation between the gradients uploaded by the client through the gradient deconfliction mechanism, thereby ensuring the rapid convergence of the global model in a non-IID data environment. Unlike traditional federated learning methods, the federated gradient coordination algorithm does not rely on additional hyperparameter adjustments, but achieves adaptive coordination of gradients through parameter-free remote term design, ensuring that the model update process is more efficient and smooth.

[0026] It should also be understood that non-IID data refers to data distribution that differs significantly between different clients, rather than being independent and identically distributed (IID). In a federated learning environment, non-IID data is usually manifested as imbalanced data categories across different clients (e.g., some client data only contains specific categories), different feature distributions, or a large disparity in the number of samples. This data imbalance can make it difficult for the model to converge during global updates, affecting generalization performance. The Federated Gradient Reconciliation Algorithm reduces gradient conflicts, alleviates the convergence problem caused by non-IID data, and improves the stability and efficiency of federated learning.

[0027] In a second aspect, a rice blast disease grade intelligent assessment and disease resistance initial screening system is provided, the system comprising: An image acquisition module, which is used to acquire images of wild rice leaves; A segmentation and evaluation module, which can execute any implementation of the first aspect; Data Analysis Module,The data analysis module is used to query the results of the segmentation and evaluation modules.

[0028] A third aspect provides a device, comprising a processor and a memory, wherein the memory stores a computer program, and the program is executed by the processor so that the implementation method described in any one of the first aspects is executed. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A flowchart of an intelligent evaluation method for rice blast disease grade and initial screening of disease resistance provided in an embodiment of the present application.

[0030] Figure 2 A schematic diagram of a GhostHierarchicalNet backbone network structure provided in an embodiment of the present application.

[0031] Figure 3 A schematic diagram of a CAHSFPN neck network structure provided in an embodiment of the present application.

[0032] Figure 4 A schematic diagram of the LSDECS Head segmentation head structure provided in an embodiment of the present application.

[0033] Figure 5 A schematic diagram of the structure of a rice blast disease grade intelligent assessment and disease resistance initial screening system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0034] The terms used in the following embodiments are only for the purpose of describing specific embodiments and are not intended to be limiting of the present application. As used in the specification and appended claims of the present application, the singular expressions "one", "a kind of", "", "above", "the" and "this" are intended to also include expressions such as "one or more", unless there is a clear contrary indication in the context. It should also be understood that in the following embodiments of the present application, "at least one", "one or more" refer to one, two or more. The term "and / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist; for example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship.

[0035] References to "one embodiment" or "some embodiments" etc. described in this specification mean that a particular feature, structure or characteristic described in conjunction with the embodiment is included in one or more embodiments of the present application. Thus, the phrases "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear at different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0036] Rice blast is a global disease that seriously harms rice, significantly affecting yield and quality. Traditional chemical control is prone to cause environmental pollution and drug resistance, while genetic improvement through screening of disease-resistant genes is a more economical and environmentally sustainable strategy. However, the process of mining disease-resistant genes from wild rice is complicated, and traditional methods rely on manual evaluation, which has problems such as low efficiency, strong subjectivity and unstable results.

[0037] The embodiments of the present application provide a method, system and equipment for intelligent evaluation of rice blast disease grade and initial screening of disease resistance. The method can effectively improve the accuracy and efficiency of crop disease assessment, provide support for large-scale wild rice disease resistance screening, and thus promote the development of rice blast resistance breeding.

[0038] The technical solution of the embodiments of the present application will be described below with reference to the accompanying drawings.

[0039] Figure 1 A flowchart of a method for intelligently evaluating the disease grade of rice blast and initial screening of disease resistance provided in an embodiment of the present application is provided. In some examples, the method includes: S1: Collect image data of wild rice leaves infected with rice blast, perform instance segmentation on the image data by improving the YOLOv8-Seg segmentation model, and obtain the segmentation results. The improvement measures include: GhostHierarchicalNet backbone network, CAHSFPN neck network and LSDECS Head segmentation head; S2: Based on the segmentation results, the number of pixels and their proportions of diseased leaves and diseased areas are counted, and the disease grade assessment results are obtained in combination with the technical procedures for disease resistance identification; S3: Based on the results of repeated disease grade assessments, wild rice materials with strong disease resistance were screened.

[0040] Figure 2 A schematic diagram of a GhostHierarchicalNet backbone network structure provided in an embodiment of the present application. In some examples, the GhostHierarchicalNet backbone network includes: HierarchicalStem module,HierarchicalStem module extracts global and local features through a dual-branch structure; GHGBlock module,The GHGBlock module consists of multiple series-connected Ghost convolutional layers, each of which uses the same convolution kernel size and activation function to gradually refine and enhance the input features; Depthwise separable convolution, depthwise separable convolution includes depthwise convolution and pointwise convolution, which reduces the computational complexity.

[0041] In one possible implementation, the HierarchicalStem module consists of five convolutional layers and one maximum pooling layer, using a dual-branch structure for feature extraction and downsampling. First, the input is downsampled through a 3×3 convolutional layer, and the feature map is then divided into two paths: one is passed through a 2×2 maximum pooling layer to extract global features, and the other is passed through two 2×2 convolutional layers in sequence to extract detail features. Next, the outputs of the two paths are concatenated in the channel dimension to fuse global and local information. The fused feature map is further downsampled through a 3×3 convolutional layer, and finally the number of channels is adjusted through a 1×1 convolutional layer, and the output is used for feature representation in subsequent network layers. This module uses a smaller convolution kernel size and a smaller number of channels, which significantly reduces the number of parameters and computational complexity, while fully extracting global and local features.

[0042] In one possible implementation, the GHGBlock module consists of multiple series-connected Ghost convolutional layers, each of which uses the same convolution kernel size and activation function to gradually refine and enhance the input features. First, the input feature map passes through a series of Ghost convolutional layers to extract rich feature information; then, the outputs of all intermediate layers are concatenated with the initial input in the channel dimension to form a feature map containing multi-level features. Next, a 1×1 convolutional layer is used to compress the number of channels to reduce computational overhead, and then another 1×1 convolutional layer is used to adjust the number of channels to enhance nonlinear expression capabilities. The GHGBlock module uses the efficiency of Ghost convolution to generate features through linear operations, maintaining model accuracy while significantly improving computational efficiency, and capturing feature information at different levels by concatenating multiple layers of output.

[0043] Optionally, the GhostHierarchicalNet backbone network also includes spatial pyramid pooling at the end, which can fuse local details with global context information and enhance the model's perception of objects of different sizes.

[0044] Figure 3 A schematic diagram of a CAHSFPN neck network structure is provided for an embodiment of the present application. In some examples, the CAHSFPN neck network includes: BiDirectionalFusion module,BiDirectionalFusion module expands the receptive field through horizontal and vertical convolution; C2f_Faster_CGLU module,The C2f_Faster_CGLU module combines the convolutional gated linear unit (CGLU) with partial convolution to achieve lightweight feature fusion.

[0045] In one possible implementation, first, the BiDirectionalFusion module downsamples the input features through a mean pooling layer, significantly reducing the burden of subsequent calculations; then, the features are preliminarily processed through an initial convolution layer to balance the depth and width of the feature map; next, the horizontal and vertical convolution layers are introduced in parallel to expand the receptive field of the features in the horizontal and vertical directions respectively. This two-way expansion strategy effectively captures the complex spatial relationships in the feature map; finally, the features are further refined through another convolution layer, and the nonlinear expression ability of the model is enhanced through the activation function Sigmoid. Through this combination of horizontal and vertical convolutions, the module maintains the relative stability of computational complexity and model size while expanding the receptive field.

[0046] In one possible implementation, the convolutional gated linear unit processes input features through a parallel standard convolution path and a gated path, and combines the outputs of the two paths through an element-wise multiplication operation to generate a more discriminative feature representation. Grouped convolution further integrates spatial information, effectively expands the receptive field, and reduces redundancy between feature channels.

[0047] Figure 4 A schematic diagram of the structure of an LSDECS Head segmentation head provided in an embodiment of the present application. In some examples, the LSDECS Head segmentation head includes: Shared convolutional layer: Shared convolutional layer is used to reduce redundant calculations of feature maps with different resolutions; Detail enhancement convolution module,The detail enhancement convolution module is used to capture the detail features of lesions.

[0048] In a possible implementation, an embodiment of the present application proposes a lightweight segmentation head module, LSDECS Head. The module introduces a shared convolution layer, so that feature maps of different resolutions can share parameters, significantly reducing redundant calculations, and reducing the number of parameters and memory usage of the segmentation head. At the same time, a detail enhancement convolution module is adopted to integrate multiple feature extraction methods to improve the robustness of the model. The detail enhancement convolution module is a core component of the LSDECS Head, which aims to enhance the spatial feature extraction capability of the convolution operation. It integrates multiple specialized convolution layers, including: center difference convolution (Conv2d_cd), horizontal difference convolution (Conv2d_hd), vertical difference convolution (Conv2d_vd) and angle difference convolution (Conv2d_ad). Each convolution branch is optimized for different spatial feature patterns to collaboratively capture rich spatial information. This design strengthens the extraction of detail features of rice blast diseased leaves and diseased areas, ensuring that high segmentation accuracy can be maintained under limited computing resources.

[0049] In some examples, the disease level assessment method in step S2 includes: Divide diseased leaves and spots; Assign the diseased spots to the diseased leaves, and assign each diseased spot to the corresponding diseased leaf by calculating the overlap between the segmentation masks; According to the segmentation results, the number of pixels of each diseased leaf and its corresponding diseased spot is calculated; Calculate the ratio of diseased spot pixels to diseased leaf pixels using the following formula: , in, is the number of pixels of the segmented lesion, is the number of pixels of the segmented diseased leaves, and LCR represents the ratio of diseased spots to diseased leaves.

[0050] In some examples, the method for screening wild rice materials with strong disease resistance in step S3 includes: Summarize the results of multiple replicates to evaluate disease resistance: The disease grade evaluation data from multiple replicate tests of the same variety are summarized using the following formula to obtain the results of the disease resistance intelligent evaluation: , in, is the sum of the disease levels in different replicates, N The number of replicates, R Results of the intelligent disease resistance assessment of representative varieties.

[0051] In some examples, a hierarchical adaptive amplitude pruning method is used to optimize the improved YOLOv8-Seg segmentation model.

[0052] In one possible implementation, the LAMP pruning process includes the following steps: 1) Weight sorting and scoring calculation: First, the weights of each layer are sorted by their magnitude and the cumulative value of their square sum is calculated. Then, each weight is assigned a score according to the following pruning scoring formula: , in, is the value of the u-th weight, is the sum of the squares of all unpruned weights in the current layer. This formula is used to sort the weights. When , this formula assigns scores to weights based on the ratio of the square of each weight to the sum of the squares of the remaining weights. Larger weights have higher scores and are preferred to be retained; smaller weights have lower scores and are preferred to be pruned.

[0053] 2) Global sorting and pruning priority determination: Globally sort the weight scores of all layers and determine the pruning priority. In order of scores from low to high, gradually prune the weights with low scores until the preset sparsity target is reached.

[0054] 3) After pruning, fine-tune to restore model performance. This process is repeated multiple times: Re-evaluate the model after fine-tuning; Recalculate the LAMP score and continue to remove the weights with low scores; Until the pruning and fine-tuning process meets the preset conditions, the model finally reaches the optimized state.

[0055] 4) Minimize L2 norm distortion: In order to ensure the stability of the model performance after pruning, LAMP adopts the principle of minimizing the L2 norm distortion. During the pruning process, LAMP uses the following formula to find the optimal pruning mask M : , in, represents the weight matrix after pruning, is the input vector. The optimization problem can be further simplified to the Frobenius norm minimization problem: , By minimizing this norm, LAMP can effectively ensure that the output difference between the pruned model and the original model is minimized.

[0056] In some examples, the federated gradient coordination algorithm is used to train an improved YOLOv8-Seg segmentation model, including: The client calculates the gradient after local training; The server detects gradient conflicts and adjusts the gradient direction by projecting conflicting gradients. Aggregate the optimized gradients to update the global model parameters.

[0057] In one possible implementation, in each iteration, the server first randomly selects a subset of clients , and the global model Distributed to all selected clients. Each client performs a round of stochastic gradient descent (SGD) training on its local data to generate locally updated model parameters , and returns it to the server. The server receives these updates and calculates the gradient of each client , and adjust these gradients through gradient conflict removal operations. If the inner product of the gradient of a client and the gradient of other clients is found to be negative, the server will adjust the gradient to ensure that the update direction of each client is consistent. Then, the server performs weighted aggregation on the adjusted gradients to generate new global model parameters and sync it back to all clients.

[0058] Figure 5 A schematic diagram of the structure of a rice blast disease grade intelligent assessment and disease resistance initial screening system provided in an embodiment of the present application. In some examples, the system includes: An image acquisition module, which is used to acquire images of wild rice leaves; A segmentation and evaluation module, which can execute any implementation of the first aspect; Data Analysis Module,The data analysis module is used to query the results of the segmentation and evaluation modules.

[0059] In a possible implementation, augmented reality glasses are used to collect images of wild rice leaves, and the collected images are input into the segmentation and evaluation module. The results queried by the data analysis module can be displayed on the augmented reality glasses.

[0060] Optionally, a smartphone APP is developed, and the model weights are integrated and deployed to the smartphone APP by converting the PyTorch format (.pt) to the TensorFlow Lite format (.tflite). In one possible implementation, the smartphone APP includes disease assessment, data analysis, system management and settings modules. The intelligent assessment module is mainly responsible for the segmentation of diseased leaves and diseases and the assessment of disease levels. In this module, users can pre-fill in information such as the wild rice material number for assessment, so that they can select through the AR glasses APP during the investigation process, and display the results of disease assessment and disease resistance screening in real time. In the data analysis module, users can enter search information to query the results of disease assessment and disease resistance screening. The system management and settings module includes functions such as personal account, version information, and system preference settings.

[0061] Optionally, an augmented reality glasses APP is developed, which includes two modules: information collection and result display. The information collection module can select the currently evaluated wild rice material number and other survey information through voice commands or smart rings, and control the AR glasses to collect images. The collected images will be transmitted to the smartphone via Bluetooth or hotspot. The result display module is used to receive the results of disease assessment and disease resistance screening returned by the smartphone, and display them synchronously and in real time on the AR glasses.

[0062] Optionally, the server side is built based on the Micronaut framework, including cloud synchronization and federated training modules. The cloud synchronization module provides an API interface for real-time synchronization of disease assessment and disease resistance screening results. The federated training module is responsible for the aggregation and management of model weights. In order to meet the needs of some users for low-performance mobile devices, users can select the server side for cloud processing in the system preferences of the smartphone APP. At this time, the images collected by the AR glasses will be automatically uploaded to the server side for processing via the HTTPS protocol, and the processing results will be returned to the smartphone.

[0063] An embodiment of the present application further provides a device, including a processor and a memory, wherein the memory stores a computer program, and the program is executed by the processor so that the method described in any of the above examples is executed.

[0064] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. Any equivalent modifications or changes made by ordinary technicians in this field based on the contents disclosed by the present invention should be included in the protection scope recorded in the claims.

Claims

1. A method for intelligently evaluating the disease grade of rice blast and initial screening of disease resistance, characterized in that: The method comprises: S1: Collect image data of wild rice leaves infected with rice blast, perform instance segmentation on the image data by improving the YOLOv8-Seg segmentation model, and obtain segmentation results. The improvement measures include: GhostHierarchicalNet backbone network, CAHSFPN neck network and LSDECS Head segmentation head; S2: According to the segmentation results, the number of pixels of diseased leaves and diseased areas and their proportions are counted, and the disease grade assessment results are obtained in combination with the disease resistance identification technical procedures; S3: Based on the repeated disease grade evaluation results, wild rice materials with strong disease resistance are screened.

2. The method according to claim 1, characterized in that The GhostHierarchicalNet backbone network includes: A HierarchicalStem module, wherein the HierarchicalStem module extracts global and local features through a dual-branch structure; GHGBlock module, which consists of multiple series-connected Ghost convolutional layers, each of which uses the same size of convolution kernel and activation function to gradually refine and enhance the input features; Depthwise separable convolution, where the depthwise separable convolution includes depthwise convolution and pointwise convolution, and the depthwise separable convolution is used to reduce computational complexity.

3. The method according to claim 1, characterized in that The CAHSFPN neck network includes: A BiDirectionalFusion module, which expands the receptive field through horizontal and vertical convolution; C2f_Faster_CGLU module, the C2f_Faster_CGLU module combines convolutional gated linear units with partial convolutions to achieve lightweight feature fusion.

4. The method according to claim 1, characterized in that: The LSDECS Head segmentation head includes: A shared convolution layer, wherein the shared convolution layer is used to reduce redundant calculations of feature maps of different resolutions; A detail enhancement convolution module is used to capture the detail features of the lesions.

5. The method according to claim 1, characterized in that The disease grade assessment method in step S2 includes: Divide diseased leaves and spots; Assigning the diseased spots to the diseased leaves, and assigning each diseased spot to the corresponding diseased leaf by calculating the overlapping parts between the segmentation masks; According to the segmentation result, the number of pixels of each diseased leaf and the diseased spot to which it belongs is calculated; Calculate the ratio of diseased spot pixels to diseased leaf pixels using the following formula: ， in, is the number of pixels of the segmented lesion, is the number of pixels of the segmented diseased leaves, and LCR represents the ratio of diseased spots to diseased leaves.

6. The method according to claim 1, characterized in that The method for screening wild rice materials with strong disease resistance in step S3 comprises: Summarize the results of multiple replicates to evaluate disease resistance: The disease grade evaluation data from multiple replicate tests of the same variety are summarized using the following formula to obtain the results of the disease resistance intelligent evaluation: ， in, is the sum of the disease levels in different replicates, N The number of replicates, R Results of the intelligent disease resistance assessment of representative varieties.

7. The method according to claim 1, characterized in that The improved YOLOv8-Seg segmentation model is optimized using a hierarchical adaptive amplitude pruning method.

8. The method according to claim 1, characterized in that The improved YOLOv8-Seg segmentation model is trained using the federated gradient coordination algorithm, including: The client calculates the gradient after local training; The server detects gradient conflicts and adjusts the gradient direction by projecting conflicting gradients. Aggregate the optimized gradients to update the global model parameters.

9. A rice blast disease grade intelligent assessment and disease resistance screening system, characterized in that: The system comprises: An image acquisition module, wherein the image acquisition module is used to acquire images of wild rice leaves; A segmentation and evaluation module, wherein the segmentation and evaluation module can execute the method according to any one of claims 1 to 8; A data analysis module is used to query the results of the segmentation and evaluation module.

10. A device, characterized in that: The method comprises a processor and a memory, wherein the memory stores a computer program, and the program is executed by the processor so that the method according to any one of claims 1 to 8 is executed.

Images