Crop Disease Segmentation Method, System, Device and Medium Based on Multi-Scale Fusion and CBAM-ResNet50
The U-Net-ResNet50-CBAM method enhances agricultural disease segmentation by improving feature extraction and background suppression, addressing inefficiencies in traditional and deep learning methods, achieving better accuracy and speed in complex crop disease imaging.
Patent Information
- Application Number
- CN202310667637.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-07
- Publication Date
- 2025-07-15
- Estimated Expiration
- 2043-06-07
AI Technical Summary
The prior art has problems of low flexibility, high cost and long-term consumption in crop disease image segmentation, especially in complex backgrounds and small object detection, and it is difficult for traditional methods to effectively extract disease characteristics.
Using a method based on multi-scale fusion and CBAM-ResNet50, a multi-scale adaptive feature fusion module is designed by combining attention mechanisms and residual networks in the U-Net basic model, and a hollow space pyramid pooling is used at the bottom of the network to enhance the target area characterization ability, suppress background areas, and improve the model convergence speed and small object detection ability.
It improves the accuracy and speed of crop disease segmentation, especially in complex backgrounds and small goals, with good segmentation effects, supporting the development of precision agriculture.
Smart Images
Figure CN116543282B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of strawberry disease segmentation, and particularly to a crop disease segmentation method, system, device and medium based on multi-scale fusion and CBAM-ResNet50. Background Technique
[0002] Crop disease image segmentation is an important step in disease detection and disease type recognition methods, and its segmentation effect directly affects the subsequent detection and recognition results. Due to the complex, diverse, irregular and changeable shapes and colors of crop disease images, image segmentation technology can restore the irregular distribution areas of diseases, providing an effective basis for subsequent disease type recognition and disease diagnosis, which is of great significance. Previously, crop disease detection was mainly traditional manual detection, which required a large amount of manpower and material resources, and was easily affected by the subjective factors of the detection personnel, with low detection efficiency. Therefore, the automation of disease detection has high research value and application prospects.
[0003] With the development of computer vision, object detection methods based on digital image processing have been applied to crop disease detection. Sun Jun et al. improved the Otsu algorithm and used the particle swarm optimization algorithm to find the optimal threshold and apply it to lettuce leaf segmentation. Zhang et al. and Wang et al. proposed using the K-means algorithm to segment images and extract PHOG features from the segmented images to identify diseases so as to achieve the segmentation effect. Ma et al. proposed a method for segmenting vegetable disease leaf spots using comprehensive color features, realizing strong recognition of disease spots and clutter backgrounds on diseased leaves. The above-mentioned traditional image processing methods require feature extraction and selection. However, due to the very complex and irregular lesion areas of crop disease images, it is difficult for traditional methods to select and extract the best features for disease recognition from images. Traditional disease detection has high costs and long time consumption, so the flexibility of the network is not high and the application scope is limited.
[0004] In recent years, due to the better generalization performance of deep learning and its good anti-noise ability, it has gradually been applied to the segmentation of diseases. To varying degrees, it has achieved the goal of multi-scale feature extraction, improved the accuracy and segmentation speed of the model, and has good universality and strong generalization ability. Wang Xiangyu et al. combined the visible spectral images of cucumber brown spot disease and used the U-Net deep learning network to construct a semantic segmentation model for cucumber brown spot disease, achieving lesion segmentation. Zhu Lixue et al. found that the traditional U-Net network in the banana fruit bunch recognition system has problems such as poor real-time performance, a large number of parameters, and loss of spatial information after downsampling. This study proposed a lightweight segmentation network based on the U-Net model. Zhang Huimin et al. proposed a method for cucumber disease leaf segmentation based on a multi-scale fusion convolutional neural network, and proposed using a multi-scale convolutional network and a linear interpolation algorithm. The accuracy of the improved network can reach 93.12%. He Zifen et al. and Gu Xingjian et al. proposed a multi-scale fusion neural network composed of an encoder-decoder, which avoids traditional artificial feature extraction and has a simpler structure than existing segmentation networks. The above studies have all achieved certain results in crop disease detection. However, the types of crop disease detection are relatively single, generally only targeting a certain crop disease. At the same time, the problems of complex background and small target in crop disease segmentation are not handled well enough. Summary of the Invention
[0005] In order to overcome the defects of the above-mentioned existing technologies, the purpose of the present invention is to provide a crop disease segmentation method, system, device and medium based on multi-scale fusion and CBAM-ResNet50. First, a U-Net basic model is built, and the attention mechanism and residual network are combined in the backbone network layer for efficient feature extraction. On the one hand, it strengthens the representation ability of the target area and suppresses the background area, so as to achieve the problem of disease area segmentation under complex backgrounds. On the other hand, the residual error is used to reduce the vanishing gradient and improve the convergence speed of the model. Then, in the feature fusion layer, a multi-scale adaptive feature fusion module based on context information is designed to make up for the information loss of the current layer's features through adjacent feature information, so as to improve the detection ability of small targets. Finally, at the bottom layer of the network, atrous spatial pyramid pooling is used, and atrous convolutions with different dilation rates are used to increase the global receptive field of the features to achieve the perception of the overall segmentation area.
[0006] In order to achieve the above purpose, the technical solution adopted by the present invention is as follows:
[0007] A crop disease segmentation method based on multi-scale fusion and CBAM-ResNet50 specifically includes the following steps:
[0008] Step 1. Data preprocessing
[0009] We selected several publicly available images of crop diseases in natural environments at different growth stages, from sunny dry environments to cloudy humid environments, as the dataset.
[0010] 1) Image annotation: Annotate the crop disease dataset and indicate the category to which it belongs;
[0011] 2) Divide the data set: Divide the labeled data set into a training set and a test set, and compress the original image to the default image size of the U-Net network;
[0012] Step 2: Improve U-Net network
[0013] 1) Replace the feature extraction network
[0014] The VGG16 backbone feature network in U-Net is replaced with the ResNet50 backbone feature network; the CBAM attention mechanism is embedded in the bottom layer of each residual module to form a backbone feature extraction network based on the attention mechanism and residual structure;
[0015] 2) Design a multi-scale adaptive feature fusion module ASFF, specifically:
[0016] Before the jump connection in the decoding stage, all the features of different scales obtained from the backbone feature network are adjusted to a uniform size through upsampling operations, and then weighted fusion is performed according to the adaptively learned weights to obtain the fused features. Finally, the fused features are jump-connected with the upsampled features in the decoding stage. At the same time, atrous spatial pyramid pooling is used at the bottom of the network, and atrous convolutions with different expansion rates are used to increase the global receptive field of the features to achieve perception of the overall segmented area.
[0017] Step 3: Train and improve the U-Net network
[0018] The crop disease image processed by data in step 2) of step 1 is used as the input image of the improved U-Net network in step 2, and the operating parameters of the network model are set. The improved U-Net network model in step 2 is trained on the experimental operation platform. The commonly used evaluation criteria in the segmentation field, including: mPA value, F_Score value, mIoU value, and Loss value, are used to evaluate the segmentation result of the improved U-Net. If the segmentation result meets the expectation, the segmentation result is output. If it does not meet the expectation and the problem of U-Net is not solved, the above steps are repeated to adjust the parameters until the segmentation result meets the expectation and the final disease segmentation result is output.
[0019] The specific method of step 2) of step 2 is as follows:
[0020] Use dilated convolutions with different dilation rates for multi-scale feature extraction; dilated convolutions can expand the convolutional kernel to the scale constrained by the dilation coefficient, and fill the unoccupied areas in the original convolutional kernel with 0, achieving that the receptive field size increases with the increase of the dilation coefficient while the number of parameters of the convolutional kernel remains unchanged and the image resolution is not reduced.
[0021] The dilated convolution described in Step 2 has an additional parameter r compared to the ordinary convolution, where r is the rate, representing the dilation coefficient of the convolutional kernel; the calculation method of the dilated convolution is:
[0022] Assume the original convolutional kernel size is f and the dilation coefficient is r, then the size of the convolutional kernel after dilation is f':
[0023] f' = r(f - 1) + 1 (1)
[0024] The receptive field size of the convolutional kernel after dilation is:
[0025] [(f + 1)×(r - 1) + f]×[(f + 1)×(r - 1) + f] (2).
[0026] The running parameters of the network model in Step 3 include the initial learning rate, learning momentum, weight decay rate, and the network parameters are updated using the Stochastic Gradient Descent (SGD) method.
[0027] In Step 3, the commonly used evaluation criteria in the segmentation field are used to evaluate the segmentation effect of the crop disease segmentation model by Pixel Accuracy (PA) and Mean Pixel Accuracy (mPA);
[0028] Pixel Accuracy (PA) is determined by Recall (R) and Precision (P), and is an intuitive standard for the segmentation performance results of a single category.
[0029]
[0030] In the formula, T P — The number of correctly segmented target objects; F P — The number of incorrectly segmented target objects; F N — The number of unsegmented target objects; P — Precision; R — Recall;
[0031] The segmentation effects of multiple categories are evaluated using the Mean Pixel Accuracy (mPA); the higher the values of PA and mPA, the better the performance of the segmenter. The F1 score takes into account both the precision and recall of the model; the mean Intersection over Union (mIoU) is used to verify the average intersection over union of the predicted values and the ground truth values.
[0032] A crop disease segmentation system based on the above method, including:
[0033] Attention Residual Feature Extraction Module: The dataset after data preprocessing is used to enhance the representation ability of the target region through training in Step 2, suppressing the background region, thereby realizing the problem of disease region segmentation under complex backgrounds. On the other hand, the residual is used to reduce gradient disappearance and improve the convergence speed of the model.
[0034] Multi-scale Fusion Module: By using Step 2 to make up for the information loss of the current layer's features with adjacent feature information, the detection ability of small targets is improved.
[0035] Dilated Convolution Module: Dilated convolutions with different dilation rates in Step 2 are used to increase the global receptive field of features to achieve the perception of the overall segmentation region.
[0036] A crop disease segmentation device based on the above method, comprising:
[0037] A memory for storing computer programs, data, and models;
[0038] A processor for implementing the crop disease segmentation method described in Steps 1 to 3 when executing the computer program.
[0039] A computer-readable storage medium for reading and storing programs and data. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it can perform crop disease segmentation based on the segmentation method described in Steps 1 to 3.
[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0041] In the present invention, by building a U-Net basic model and combining the attention mechanism and the residual network in the backbone network layer for efficient feature extraction, on the one hand, the representation ability of the target region is enhanced, the background region is suppressed, thereby realizing the problem of disease region segmentation under complex backgrounds; on the other hand, the residual is used to reduce gradient disappearance and improve the convergence speed of the model. Then, in the feature fusion layer, a multi-scale adaptive feature fusion module based on context information is designed to make up for the information loss of the current layer's features with adjacent feature information, thereby improving the segmentation ability of small targets. Finally, at the bottom layer of the network, atrous spatial pyramid pooling is used, and atrous convolutions with different dilation rates are used to increase the global receptive field of features to achieve the perception of the overall segmentation region.
[0042] In summary, based on the standard U-Net model, by changing the model feature extraction network and designing a multi-scale fusion module, compared with the existing crop disease segmentation algorithms, the present invention improves the accuracy of crop disease segmentation, enhances the segmentation speed, and also has a good segmentation effect on small target crops and complex background situations, proving that the model in this paper can achieve fast and accurate segmentation of crop diseases and can provide support for the development of precision agriculture. Description of the Drawings
[0043] Figure 1 This is a diagram of the crop disease segmentation model based on multi-scale fusion and CBAM-ResNet50 of the present invention.
[0044] Figure 2 This is the Labelme annotation diagram of the present invention.
[0045] Figure 3 This is the backbone feature network module diagram of the present invention.
[0046] Figure 4 This is the ASFF multi-scale fusion module diagram of the present invention.
[0047] Figure 5 This is the ASPP multi-scale fusion module diagram of the present invention. DETAILED DESCRIPTION
[0048] The present invention is further described in detail below with reference to the accompanying drawings and embodiments:
[0049] See also Figure 1 , taking strawberry disease segmentation as an example:
[0050] A crop disease segmentation method based on multi-scale fusion and CBAM-ResNet50 specifically includes the following steps:
[0051] Step 1: Data preprocessing: including image annotation and data set division
[0052] 1) Image annotation: Use the Labelme image annotation tool to annotate the strawberry disease dataset, segment the strawberry disease in each image, and indicate the category to which it belongs, see Figure 2 ;
[0053] 2) Divide the dataset: 80% of the expanded dataset is used as the training set and 20% as the test set. The original images are compressed into 300×300 pixels as the input of the training model.
[0054] Training U-Net network:
[0055] The image processed in step 2) of step 1 is used as the input image of the U-Net network model, and the operating parameters of the U-Net network model are set. The U-Net network model is trained on the experimental operation platform, and then the commonly used evaluation criteria in the segmentation field, including: mPA value, F_Score value, mIoU value, and Loss value, are used to evaluate the effect of the trained U-Net network model;
[0056] Step 2: Improve U-NET network
[0057] 1) Replace the feature extraction network
[0058] See Figure 3 , replace the feature extraction network of the U-Net network with the CBAM-ResNet50 network. On the one hand, it strengthens the representation ability of the target area and suppresses the background area to achieve the problem of disease area segmentation under complex backgrounds; on the other hand, it reduces the gradient disappearance through residuals and improves the convergence speed of the model;
[0059] First, replace the VGG16 backbone feature network in the U-Net with the ResNet50 backbone feature network;
[0060] Then, embed the CBAM attention mechanism into the bottom layer of each residual module to form a backbone feature extraction network based on the attention mechanism and the residual structure;
[0061] 2) Design a multi-scale fusion module
[0062] See Figure 4 、 Figure 5 , design a multi-scale adaptive feature fusion module based on context information to make up for the information loss of the current layer's features through adjacent feature information, thereby improving the segmentation ability of small targets; at the same time, use dilated spatial pyramid pooling at the bottom layer of the network to increase the global receptive field of the features through dilated convolutions with different dilation rates to achieve the perception of the overall segmentation area. The specific method is as follows:
[0063] a) First, perform convolution processing on the feature maps of 4 different scales to the same size, sample Level1, Level2, and Level3 1, 2, and 3 times respectively, and finally achieve the unification of the sizes and channel numbers of the feature maps of different scales; then multiply the unified-size feature maps by an adaptive weight parameter, and finally add the element values at the same positions of each feature map to obtain the feature ASFF-0 that combines the shallow spatial structure information and the deep fine-grained semantic information;
[0064] b) Use dilated convolutions with different dilation rates for multi-scale feature extraction; dilated convolutions can expand the convolution kernel to the scale constrained by the dilation coefficient and fill the unoccupied areas in the original 3×3 convolution kernel with 0. In this way, without changing the number of parameters of the convolution kernel and without reducing the image resolution, the size of the receptive field increases with the increase of the dilation coefficient.
[0065] The dilated convolution is a special convolution. Compared with the ordinary convolution, it adds an r parameter, where r is the rate, representing the dilation coefficient of the convolution kernel; the calculation method of the dilated convolution is:
[0066] Assume that the original convolution kernel size is f and the dilation coefficient is r, then the size of the convolution kernel after dilation is f':
[0067] f' = r(f - 1) + 1 (1)
[0068] The receptive field size of the convolutional kernel after dilation is:
[0069] [(f + 1)×(r - 1) + f]×[(f + 1)×(r - 1) + f] (2)
[0070] A 3×3 convolutional kernel is used, and dilated convolutions with dilation rates of 6, 12, and 18 are used to sample the input image to obtain rich image context information. The outputs of the two branches are feature fused to obtain a new feature map. The size of the new feature map remains unchanged, and the number of channels changes from the original 512 to 1024.
[0071] Step 3: Train the improved U-Netw network:
[0072] The strawberry images processed in Step 1 are used as the input images of the improved U-Net network model in Step 2, and the operating parameters of the network model are set. The improved U-Net model is trained on the experimental operation platform. The evaluation criteria commonly used in the field of object segmentation are adopted, including: mPA value, F_Score value, mIoU value, Loss value, to evaluate the segmentation results of the improved U-Net, including: mPA value, F_Score value, mIoU value, Loss value,; The method proposed in the present invention improves the mPA value by 8.39% compared with the standard U-Net model, the F_Score value by 0.0822, the mIoU value by 8.53%, and the Loss is reduced by 0.3932, reducing the missed detection and misdetection situations of the model.
[0073] The operating parameters of the network model in Step 3 are: the initial learning rate is 0.001, the stochastic gradient descent method SGD is used to update the network parameters, the learning momentum is 0.9, and the weight decay rate is 0.0005.
[0074] The experimental operation platform in Step 3 is the Ubuntu 16.04 system, the PyTorch deep learning framework is adopted, the CPU model is Intel(R) Xeon(R) E5-2678 v3 @ 2.5GHz, the model of the graphics card (GPU) is NVIDIA GeForce RTX2080Ti, the graphics card memory is 11GB, and the programming language is Python.
[0075] In Step 3, the evaluation criteria commonly used in the segmentation field adopt Pixel Accuracy (PA) and mean Pixel Accuracy (mPA) to evaluate the segmentation effect of the strawberry disease segmentation model;
[0076] The Pixel Accuracy (PA) is determined by the Recall (R) and Precision (P), and is an intuitive criterion for the segmentation performance results of a single category.
[0077]
[0078] Where T P — The number of correctly segmented target objects; F P — The number of incorrectly segmented target objects; F N — The number of target objects not segmented; P — Precision (%) ; R — Recall (%) ;
[0079] For the segmentation effects of multiple categories, the mean Pixel Accuracy (mPA) is used for evaluation. The higher the values of PA and mPA, the better the performance of the segmenter. The F1 score takes into account both the precision and recall of the model. The mean Intersection over Union (mIoU) is used to verify the average intersection over union of the predicted values and the ground truth values.
[0080] Referring to Table 1, compared with the standard U-Net model, the crop disease segmentation method based on multi-scale fusion and CBAM-ResNet50 proposed in the present invention has an 8.39% increase in the mPA value, an increase of 0.0822 in the F_Score value, an increase of 8.53% in the mIoU value, and a reduction of 0.3932 in Loss, reducing the missed detection and misdetection situations of the model. Compared with other segmentation networks such as FCN, SegNet, and VGG-Unet, under the condition of the same dataset, the mPA value of the algorithm of the present invention is higher than that of the FCN, SegNet, and VGG-Unet networks respectively. Therefore, the network proposed in the present invention also has good segmentation effects on small target diseases and complex background situations, proving that the model of the present invention can achieve accurate segmentation of crop diseases and can provide support for the development of precision agriculture. The experimental results are shown in the following table. It can be seen that the segmentation effect of the improved network model of the present invention on crop diseases is better than that of other existing models.
[0081]
[0082]
[0083] A crop disease segmentation system based on the above method, comprising:
[0084] An attention residual feature extraction module, which preprocesses the dataset of the data, strengthens the representation ability of the target area through the training in step two, suppresses the background area, so as to realize the problem of disease area segmentation under complex backgrounds; on the other hand, reduces the gradient disappearance through the residual, and improves the convergence speed of the model;
[0085] The multi-scale fusion module compensates for the information loss of the current layer features through the adjacent feature information in Step 2, thereby improving the detection ability of small targets;
[0086] The dilated convolution module increases the global receptive field of the features through the dilated convolutions with different dilation rates in Step 2 to achieve the perception of the overall segmentation area.
[0087] A crop disease segmentation device based on the above method, comprising:
[0088] A memory for storing computer programs, data, and models;
[0089] A processor for implementing the crop disease segmentation method described in Steps 1 to 3 when executing the computer program.
[0090] A computer-readable storage medium for reading and storing programs and data. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it can perform crop disease segmentation based on the segmentation method described in Steps 1 to 3.
Claims
1. A crop disease segmentation method based on multi-scale fusion and CBAM-ResNet50, characterized in that The specific steps include: Step 1: Data preprocessing We selected several publicly available images of crop diseases in natural environments at different growth stages, from sunny dry environments to cloudy humid environments, as the dataset. 1) Image annotation: Annotate the crop disease dataset and indicate the category to which it belongs; 2) Divide the data set: Divide the labeled data set into a training set and a test set, and compress the original image to the default image size of the U-Net network; Step 2: Improve U-Net network 1) Replace the feature extraction network The VGG16 backbone feature network in U-Net is replaced with the ResNet50 backbone feature network; the CBAM attention mechanism is embedded in the bottom layer of each residual module to form a backbone feature extraction network based on the attention mechanism and residual structure; 2) Design a multi-scale adaptive feature fusion module ASFF, specifically: Before the jump connection in the decoding stage, all the features of different scales obtained from the backbone feature network are adjusted to a uniform size through upsampling operations, and then weighted fusion is performed according to the adaptively learned weights to obtain the fused features. Finally, the fused features are jump-connected with the upsampled features in the decoding stage. At the same time, atrous spatial pyramid pooling is used at the bottom of the network, and atrous convolutions with different expansion rates are used to increase the global receptive field of the features to achieve perception of the overall segmented area. Step 3: Train and improve the U-Net network The crop disease image processed by data in step 2) of step 1 is used as the input image of the improved U-Net network in step 2, and the operating parameters of the network model are set. The improved U-Net network model in step 2 is trained on the experimental operation platform. The commonly used evaluation criteria in the segmentation field, including: mPA value, F_Score value, mIoU value, and Loss value, are used to evaluate the segmentation result of the improved U-Net. If the segmentation result meets the expectation, the segmentation result is output. If it does not meet the expectation and the problem of U-Net is not solved, the above steps are repeated to adjust the parameters until the segmentation result meets the expectation and the final disease segmentation result is output.
2. The crop disease segmentation method based on multi-scale fusion and CBAM-ResNet50 according to claim 1, characterized in that The specific method of step 2) of step 2 is as follows: Dilated convolutions with different dilation rates are used for multi-scale feature extraction. Dilated convolution can expand the convolution kernel to the scale constrained by the dilation coefficient, and fill the unoccupied area in the original convolution kernel with 0, so that the size of the receptive field increases with the increase of the dilation coefficient while keeping the number of convolution kernel parameters unchanged and without reducing the image resolution.
3. The crop disease segmentation method based on multi-scale fusion and CBAM-ResNet50 according to claim 2, characterized in that, Compared with ordinary convolution, the dilated convolution described in step 2 adds an r parameter, where r is rate, which represents the expansion coefficient of the convolution kernel; the calculation method of the dilated convolution is: Assuming that the original convolution kernel size is f and the expansion coefficient is r, the size of the convolution kernel after expansion is f': f'=r(f-1)+1 (1) The receptive field size of the convolution kernel after expansion is: [(f+1)×(r-1)+f]×[(f+1)×(r-1)+f](2).
4. The crop disease segmentation method based on multi-scale fusion and CBAM-ResNet50 according to claim 1, characterized in that, The running parameters of the network model in step 3 include the initial learning rate, learning momentum, and weight decay rate. The network parameters are updated using the Stochastic Gradient Descent (SGD) method.
5. The crop disease segmentation method based on multi-scale fusion and CBAM-ResNet50 according to claim 1, characterized in that In step 3, the commonly used evaluation criteria in the segmentation field are used to evaluate the segmentation effect of the crop disease segmentation model by Pixel Accuracy (PA) and mean Pixel Accuracy (mPA). The Pixel Accuracy (PA) is determined by the Recall (R) and Precision (P), and it is an intuitive criterion for the segmentation performance of a single category. where T P — the number of target objects correctly segmented; F P — the number of target objects incorrectly segmented; F N — the number of target objects not segmented; P — precision; R — recall; The segmentation effect of multiple categories is evaluated using the mean Pixel Accuracy (mPA). The higher the values of PA and mPA, the better the performance of the segmenter. The F1-score takes into account both the precision and recall of the model. The mean Intersection over Union (mIoU) is used to verify the average intersection over union between the predicted value and the ground truth.
6. A crop disease segmentation system based on the crop disease segmentation method according to any one of claims 1 to 5, characterized in that, It includes: An attention residual feature extraction module. After preprocessing the dataset, it enhances the representation ability of the target area and suppresses the background area through the training in step 2, thereby realizing the segmentation of disease areas under complex backgrounds. On the other hand, it reduces the gradient vanishing through residuals and improves the convergence speed of the model. A multi-scale fusion module. It compensates for the information loss of the current layer features by neighboring feature information through step 2, thereby improving the detection ability of small targets. An atrous convolution module. It increases the global receptive field of features through atrous convolutions with different dilation rates in step 2 to achieve the perception of the overall segmentation area.
7. A crop disease segmentation device based on the crop disease segmentation method according to any one of claims 1 to 5, characterized in that, It includes: A memory for storing computer programs, data, and models. A processor for implementing the crop disease segmentation method described in steps 1 to 3 when executing the computer program.
8. A computer-readable storage medium for reading and storing programs and data. The computer-readable storage medium stores a computer program, which, when executed by a processor, can perform crop disease segmentation based on the segmentation method described in claims 1 to 5.