Crop disease detection method, terminal equipment and storage medium

Abstract

The invention relates to a crop disease detection method, terminal equipment and a storage medium. The method comprises the following steps: carrying out the following improvement on the basis of traditional YOLOv11 for crop disease detection; replacing a standard convolution module with a multi-scale parallel hybrid convolution module; the C3k2 module of the neck part is replaced by a C3k2-ED module; the C3k2 module of the main part is replaced by a C3k2-SG module; a bidirectional feature pyramid network BiFPN is introduced into the neck part to replace an original PANet network. The method can effectively extract the characteristics of the diseased leaves of the crops, improves the accuracy and practicability of disease recognition, and achieves the efficient diagnosis of the diseases of the crops.

Description

Technical Field

[0001] The present invention relates to the field of crop management, and in particular to a crop disease detection method, terminal equipment, and storage medium. Background Art

[0002] Traditional methods for detecting crop pests and diseases rely primarily on manual experience, which is not only time-consuming and labor-intensive, but also difficult to achieve rapid, accurate, and widespread diagnosis. Machine learning extracts representative patterns from large amounts of data to predict and classify them. However, machine learning identification of crop pests and diseases is susceptible to environmental interference, and high-precision models require extensive data support, placing high demands on the model. In recent years, deep learning methods have been applied to disease detection. Compared to traditional machine learning, they can automatically extract and classify image features, significantly reducing manual intervention, making them a research hotspot in crop disease identification.

[0003] However, current deep learning-based crop disease detection technology still has some shortcomings. First, the diverse scales of targets in complex field environments lead to insufficient multi-scale target capture capabilities. Second, tiny disease spot features are easily submerged in complex backgrounds, affecting detection accuracy. Finally, the high deployment cost of high-performance models makes them difficult to widely apply on resource-constrained edge devices. Summary of the Invention

[0004] In order to solve the above problems, the present invention proposes a crop disease detection method, terminal equipment and storage medium.

[0005] The specific plan is as follows:

[0006] A method for detecting crop diseases comprises the following steps:

[0007] S1: Collect crop disease images to form a training set;

[0008] S2: Build a disease detection model based on the improved YOLOv11 and train it using the training set;

[0009] The improved YOLOv11 makes the following improvements based on the traditional YOLOv11:

[0010] Replace the standard convolution module with a multi-scale parallel hybrid convolution module;

[0011] The C3k2 module in the neck part is replaced by the C3k2-ED module; the C3k2-ED module adds an EMA attention mechanism module after the first convolutional layer of the original C3k2 module, and replaces the standard bottleneck module Bottleneck of each C3k block contained in the original C3k2 module with a dilatable residual module DWR;

[0012] The C3k2 module in the backbone is replaced by the C3k2-SG module. The C3k2-SG module replaces the Bottleneck layer in the original C3k2 module with the star-shaped operation module StarBlock, and replaces the standard convolution operation inside the original C3k2 module with the GhostConv structure.

[0013] In the neck part, a bidirectional feature pyramid network BiFPN is introduced to replace the original PANet network;

[0014] S3: Detect crop diseases using the trained model.

[0015] Furthermore, in the multi-scale parallel hybrid convolution module, the input is first passed through three convolutional layers in parallel, and then the outputs of the three convolutional layers are fused, and then passed through the normalization layer, activation function, and 1×1 standard convolution in sequence before output.

[0016] Furthermore, the three convolutional layers include one layer of partial convolution and two layers of depthwise separable convolution with different kernel sizes.

[0017] Furthermore, the convolution kernel of the partial convolution is 3×3, and the convolution kernels of the two layers of depthwise separable convolution are 3×3 and 5×5 respectively.

[0018] Furthermore, the activation function adopts the SiLU activation function.

[0019] A crop disease detection terminal device includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the method described above are implemented.

[0020] Furthermore, it also includes an image uploading module, a result visualization module and a data management module, wherein the image uploading module is used to upload the crop images to be detected, the result visualization module is used to display the results of crop disease detection, and the data management module is used to store the results of crop disease detection.

[0021] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method described above in an embodiment of the present invention.

[0022] The present invention adopts the above technical solution, which can effectively extract the characteristics of diseased leaves of crops, improve the accuracy and practicality of disease identification, and achieve efficient diagnosis of crop diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Shown is a flow chart of a method according to a first embodiment of the present invention.

[0024] Figure 2 Fig. 1 shows a network structure diagram of the model in this embodiment.

[0025] Figure 3 Fig. 2 shows a network structure diagram of the multi-scale parallel mixed convolution module in this embodiment.

[0026] Figure 4 Fig. 3 shows a network structure diagram of the improved C3k2-ED module in this embodiment.

[0027] Figure 5 Fig. 4 shows a network structure diagram of the improved C3k2-SG module in this embodiment. DETAILED DESCRIPTION

[0028] To further illustrate the embodiments, the present application provides accompanying drawings. These drawings are part of the disclosure of the present application, mainly used to illustrate the embodiments, and can be used to explain the operating principle of the embodiments in conjunction with the related description of the specification. Those of ordinary skill in the art should understand other possible implementations and advantages of the present application in conjunction with these.

[0029] The present application will be further described in conjunction with the accompanying drawings and specific embodiments.

[0030] Embodiment one:

[0031] The embodiment of the present application provides a crop disease detection method, as shown in the figure, the method comprises the following steps: Figure 1

[0032] S1: Collecting crop disease images to build a training set.

[0033] In this embodiment, the crop disease images are collected and screened through multiple channels, and each collected image is subjected to data enhancement to build a data set containing 10300 original disease images, covering 27 categories of diseases and healthy leaves of 13 plant species. The disease image categories in the data set are more, including images with different backgrounds, images with different brightness conditions, small-scale images covering dense diseases and corrosive diseases, etc. LabelImg is used as a labeling tool to frame the target objects in the images, and the labeling work of the data set is completed.

[0034] S2: Building a disease detection model based on improved YOLOv11, and training through the training set.

[0035] In view of the problem of insufficient multi-scale target capturing capability in complex field environment, the improved YOLOv11 network is used to build a disease detection model in this embodiment, and the network structure is as shown in the figure. Figure 2 ​​

[0036] The basic convolution module conv in the original YOLOv11 model is replaced with the multi-scale parallel hybrid convolution module (PHyConv) proposed in this embodiment to improve the detection capability of the model for multi-scale targets. The original C3k2 module in the neck part is replaced with a C3k2-ED module to improve the detection capability of the model for small lesions and irregular diseases. The C3k2 module in the backbone part is replaced with a C3k2-SG module. The C3k2-SG module replaces the bottleneck layer in the original C3k2 module with a lightweight module, i.e., a star operation module StarBlock, which has higher parameter efficiency, to reduce the computational complexity, and replaces the standard convolution operation conv in the module with a GhostConv structure to further compress the parameter redundancy. A BiFPN bidirectional feature pyramid network is introduced in the neck part to replace the original PANet network used in the neck part of the original YOLOv11, which effectively reduces the structural complexity while improving the feature fusion quality. The above improvements are described in detail below.

[0037] 1. Multi-scale parallel hybrid convolution module

[0038] As shown in Figure 3 , the multi-scale parallel hybrid convolution (PHyConv) module is designed to solve the problem of insufficient multi-scale target capture capability in crop disease detection. The core idea of this module is to fuse different types and sizes of convolution operations through a parallel structure to maximize the model's extraction capability for multi-scale features while maintaining computational efficiency. The PHyConv module adopts a parallel-fusion form and contains three independent convolution branches, with the following specific structure.

[0039] (1) Parallel convolution layer:

[0040] 3×3 partial convolution (PConv): Partial convolution (PConv) uses a 3×3 standard convolution kernel, but in the implementation process, it groups the input feature map channels and only calculates part of the channels, while the remaining channels retain the original values. The unconvoluted channels maintain the original spatial information, which alleviates gradient diffusion. Through partial convolution, the unnecessary amount of calculation in the convolution operation is reduced, improving the efficiency of the overall model, which not only retains the ability of traditional convolution to extract spatial features, but also effectively reduces redundant calculations.

[0041] 3×3 Depthwise Separable Convolution (DWConv): Utilizing 3×3 depthwise separable convolution, the convolution operation is decomposed into channel-by-channel depthwise convolution and cross-channel point-by-point convolution, focusing on extracting local detail features. The smaller kernel size of the 3×3 convolution is suitable for capturing fine edge and texture information in an image, which is particularly important for small-scale lesion detection. Depthwise separable convolution significantly reduces the number of parameters and computational overhead, helping to improve model execution speed.

[0042] 5×5 Depthwise Separable Convolution (DWConv): This also employs the depthwise separable convolution strategy, but uses a 5×5 convolution kernel to expand the receptive field and capture a wider range of contextual information. The 5×5 convolution kernel integrates features across a larger area, extracting global, context-rich information from the image, helping to understand the positional relationship of lesions within the entire crop image. This complements the 3×3 convolution kernel, expanding the receptive field and ensuring that subtle details are captured while also not neglecting large-scale global structure.

[0043] (2) Feature fusion and normalization processing:

[0044] Channel concatenation (Concat): This concatenates the feature information obtained from 3×3PConv, 3×3DWConv, and 5×5DWConv, preserving the independence of features at each scale and fully integrating the feature information from different convolution kernels. The concatenated feature map contains both detailed and global information, and the information extracted by convolution at different scales complements each other, improving the richness and robustness of feature representation.

[0045] Batch Normalization (BN) normalization: After feature fusion, a batch normalization (BN) layer is used to normalize the fused features. This normalization eliminates the scale differences between features generated by different convolutions, ensuring that all features are at the same scale. BN normalization can alleviate the vanishing gradient problem, accelerate model training convergence, and improve overall performance.

[0046] SiLU activation function: The SiLU (Sigmoid Linear Unit) activation function is used to introduce nonlinear transformations to the normalized features. The SiLU activation function can smoothly activate the input signal, making the model more robust and flexible in feature expression.

[0047] (3) 1×1 standard convolution dimensionality reduction: After the features are multi-scale fused, normalized, and activated, a 1×1 standard convolution is used to adjust the number of channels to achieve dimensionality reduction. 1×1 convolution can integrate information from different channels, reducing the number of channels in the feature map, thereby reducing the amount of computation while maintaining important features. Dimensionality reduction not only reduces the computational complexity of subsequent layers but also prevents overfitting to a certain extent, improving the generalization ability of the model.

[0048] The PHyConv module, through the synergistic effect of the aforementioned components, successfully constructs a multi-scale convolutional architecture capable of capturing diverse patterns in input features. Partial convolution (PConv) retains the advantages of traditional convolution while optimizing computational efficiency. 3×3 and 5×5 depthwise separable convolutions extract features for detail and global information, respectively. Feature fusion and reconstruction are then achieved through channel concatenation, batch normalization, and SiLU activation. Finally, 1×1 convolutions achieve dimensionality reduction, ultimately improving the model's ability to recognize defects of varying scales.

[0049] According to the structural order of the PHyConv module, let the input feature map be X∈R H×W×C , where H and W are the height and width of the feature map respectively, and C is the number of input channels. The mathematical expression of the PHyConv module is as follows:

[0050] F(X)=Conv 1×1 (SiLU(BN(Concat(F PConv (X),F DWConv3 (X),F DWConv5 (X)))))

[0051] Where F(X) is the output result of the PHyConv module, F PConv is the partial convolution, F DWConv3 is a 3×3 depth-wise separable convolution, F DWConv5 It is a 5×5 depthwise separable convolution.

[0052] 2.C3k2-ED module

[0053] like Figure 4 As shown in the figure, to improve the feature processing capabilities of the C3k2 module, an EMA attention mechanism module is added after the first convolutional layer of the original C3k2 module. This module dynamically assigns weights to input features and determines the focus based on the actual content of the input. This improves the model's accuracy and generalization while adding a small number of parameters, and enhances its ability to handle minor diseases. In the original C3k2 module, the standard bottleneck module of the C3k block is replaced with a dilation-wise residual (DWR) module to refine features and further enhance plant disease detection capabilities. The DWR module introduces dilated convolution technology to expand the receptive field of the convolution kernel, allowing the network to capture global contextual information while maintaining sensitivity to local details. At the same time, the use of a residual connection structure effectively alleviates the problems of vanishing gradients and information loss in deep networks.

[0054] In the improved C3k2-ED module, the EMA module can adaptively weight features of different scales and different regions, so that the model pays more attention to the lesion area and key details, reduces the interference of redundant information and background noise, enhances the key information capture ability, and solves the problem that the micro lesion features are easily submerged. The DWR module expands the receptive field of the convolution kernel by using dilated convolution, and effectively alleviates the problems of gradient disappearance and information loss in deep network through residual connection, so as to capture the global context while maintaining the sensitivity to local details, expand the receptive field, and retain the detail information. After the combination of the two, the overall module can not only fully integrate multi-scale features, but also realize more efficient expression of disease features in micro lesions and complex backgrounds through the mechanism of adaptive weighting and expanded receptive field, improve the overall detection accuracy and robustness, and realize multi-scale feature integration and optimization.

[0055] 3. C3k2-SG module

[0056] As shown in Figure 5 , in the C3k2 module, the standard bottleneck module Bottleneck in C3k is replaced by the star operation module StarBlock, reducing the parameter amount of the model; secondly, the original standard convolution module is replaced by the GhostConv module, reducing the calculation amount of the model. The StarBlock lightweight module uses the characteristics of star operation (element-wise multiplication) to map the input to an implicit high-dimensional and nonlinear feature space without significantly increasing the number of channels or computational overhead. At the same time, the reason for replacing the standard convolution with GhostConv is mainly to use GhostConv to expand the feature representation by mining redundant information through inexpensive linear operations, thereby reducing the computational cost while maintaining network performance.

[0057] The improved C3k2-SG module integrates lightweight design and efficient feature fusion mechanism, realizes efficient extraction and aggregation of multi-scale and high-order nonlinear features while greatly reducing computational complexity and parameter amount, thereby improving the robustness and precision of the overall model in real-time target detection tasks. The StarBlock lightweight module is used to replace the traditional Bottleneck layer to utilize the mapping capability of star operation in the implicit high-dimensional space, and the internal standard convolution is replaced by GhostConv to generate redundant features through low-cost linear operations and efficient fusion, thereby optimizing information transmission and feature expression, so that the module not only has faster inference speed and lower memory occupancy in resource-constrained environments such as embedded and mobile terminals, but also maintains high detection accuracy and stability in complex scenarios.

[0058] 4. BiFPN

[0059] BiFPN(Bidirectional Feature Pyramid Network) is a kind of efficient multi-scale feature fusion network, and its core idea is efficient bidirectional cross-scale connection and weighted feature fusion. The BiFPN bidirectional feature pyramid network is introduced in the neck part of the model to replace the original PANet network, which can significantly enhance the model's representation ability for multi-scale targets in complex scenes through dynamic bidirectional interaction and weight adaptive fusion mechanism of multi-scale features. BiFPN constructs a dense feature interaction path between layers through bidirectional cross-layer connection architecture from top to bottom and from bottom to top, realizes efficient complementary fusion of high-level semantic information and low-level detail features, and effectively alleviates the problem of loss of shallow details and deviation of deep positioning caused by single information transmission path in traditional single-direction feature pyramid network. The core innovation lies in the introduction of learnable normalization weight coefficients to dynamically balance the contribution of different resolution feature maps to the target detection task through soft attention mechanism, thereby suppressing redundant noise features and strengthening key region response. In addition, BiFPN realizes deep mining of high-order semantic features while reducing computational complexity by pruning redundant nodes and reusing bidirectional fusion units. Therefore, the introduction of BiFPN not only enhances the multi-scale feature fusion effect of the model, but also achieves a better balance between precision and efficiency, thereby improving the positioning and classification accuracy of various targets in the target detection task.

[0060] S3: detecting crop diseases through the trained model.

[0061] The crop image to be detected is input into the trained model, and according to the model output result, it can be identified whether the crop has diseases.

[0062] Experimental verification

[0063] In order to verify the effectiveness of the improved model, in this embodiment, through comparative experiments, the performance of the model before and after lightening is compared and analyzed from four aspects of average precision (mAP@0.5 and mAP@0.5:0.95), inference speed (FPS), model parameter quantity (Params) and GFLOPs, and the experimental results are shown in Table 1:

[0064] Table 1

[0065]

[0066] From the results of Table 1, it can be seen that compared with the original YOLOv11 model, after adding the PHyConv module, the average precision mAP@0.5 of the YOLOv11-PHyConv model is improved by 1.7%, and after improving the C3k2-ED module, the mAP@0.5 of the YOLOv11-PHyConv+C3k2-ED model is improved by 4.5%, but the model parameter amount and FLOPs have a small amount of increase, and the FPS decreases.

[0067] After replacing the C3k-SG module in the backbone network, the YOLOv11-PHyConv+C3k2-ED+C3k2-SG model, compared with the YOLOv11-PHyConv+C3k2-ED model, the average precision mAP@0.5 is improved by 0.16%, and the average precision mAP@0.5:0.95 is reduced by 1.56%, which has a small decrease in detection accuracy, but is still higher than the original YOLOv11 model; and the model parameter amount (Params) and GFLOPs are reduced by 4.97% and 7.04% respectively, and the inference speed (FPS) is increased by 14.29%.

[0068] Compared with the YOLOv11-PHyConv+C3k2-ED model, after introducing the BiFPN bidirectional feature pyramid network in the Neck part, the YOLOv11-PHyConv+C3k2-ED+C3k2-SG+BiFPN model, the average precision mAP@0.5 is improved by 3.2%, and the average precision mAP@0.5:0.95 is improved by 2.23%; the model parameter amount (Params) and GFLOPs are reduced by 29.8% and 22.54% respectively, and the inference speed (FPS) is increased by 39.97%.

[0069] Finally, compared with the original YOLOv11 model, after performance improvement and lightweight improvement, the YOLOv11-PHyConv+C3k2-ED+C3k2-SG+BiFPN model, the average precision mAP@0.5 is improved by 7.86%, and the average precision mAP@0.5:0.95 is improved by 12.25%; the model parameter amount (Params) and GFLOPs are reduced by 18.46% and 12.7% respectively, and the inference speed (FPS) is increased by 15.27%.

[0070] In summary, the improved YOLOv11 model has a large overall performance improvement, which proves the effectiveness of the improvement strategy proposed in this embodiment, and provides a technical path with high efficiency and robustness for the landing of lightweight target detection models in mobile and edge computing scenarios.

[0071] The embodiment of the application improves the average precision of the model and reduces the parameter quantity of the model to realize lightweight design by sequentially introducing a PHyConv multi-scale parallel mixed convolution module, a C3k2-ED attention enhancement module, a C3k2-SG lightweight module and a BiFPN bidirectional feature pyramid in the YOLOv11 network; and the detection robustness in a complex environment is significantly enhanced in combination with a diversified data set after data enhancement.

[0072] Embodiment two:

[0073] The application further provides a crop disease detection terminal device, including a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the steps in the above method embodiments of the embodiment one of the application when executing the computer program.

[0074] Further, as an executable solution, the crop disease detection terminal device can be a desktop computer, a notebook computer, a palm computer and a cloud server and the like computing device. The crop disease detection terminal device can include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the above composition structure of the crop disease detection terminal device is only an example of the crop disease detection terminal device, and does not constitute a limitation on the crop disease detection terminal device, and can include more or fewer components than the above, or combine certain components, or different components, for example, the crop disease detection terminal device can also include an input and output device, a network access device, a bus and the like, and the embodiment of the application does not limit this.

[0075] Further, as an executable solution, the processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc., and the processor is the control center of the crop disease detection terminal device, and utilizes various interfaces and lines to connect each part of the entire crop disease detection terminal device.

[0076] The memory can be used to store the computer programs and / or modules, and the processor implements the various functions of the crop disease detection terminal device by running or executing the computer programs and / or modules stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system and at least one application required for a function; the data storage area can store data created based on the use of the mobile phone, etc. In addition, the memory can include a high-speed random access memory and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0077] Furthermore, the terminal device also includes an image uploading module, a result visualization module and a data management module, wherein the image uploading module is used to upload images of crops to be detected, the result visualization module is used to display the results of crop disease detection, and the data management module is used to store the results of crop disease detection.

[0078] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method in the embodiment of the present invention are implemented.

[0079] If the module / unit integrated in the crop disease detection terminal device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the process in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above-mentioned various method embodiments. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device that can carry the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM) and software distribution medium, etc.

[0080] Although the present invention has been particularly shown and described in conjunction with preferred embodiments, it will be understood by those skilled in the art that various changes in form and details may be made to the present invention without departing from the spirit and scope of the invention as defined in the appended claims, and all such changes are within the scope of protection of the present invention.

Claims

1. A method for detecting crop diseases, characterized in that: The following steps are involved: S1: Collect crop disease images to form a training set; S2: Build a disease detection model based on the improved YOLOv11 and train it using the training set; The improved YOLOv11 makes the following improvements based on the traditional YOLOv11: Replace the standard convolution module with a multi-scale parallel hybrid convolution module; The C3k2 module in the neck part is replaced by the C3k2-ED module; the C3k2-ED module adds an EMA attention mechanism module after the first convolutional layer of the original C3k2 module, and replaces the standard bottleneck module Bottleneck of each C3k block contained in the original C3k2 module with a dilatable residual module DWR; The C3k2 module in the backbone is replaced by the C3k2-SG module. The C3k2-SG module replaces the Bottleneck layer in the original C3k2 module with the star-shaped operation module StarBlock, and replaces the standard convolution operation inside the original C3k2 module with the GhostConv structure. In the neck part, a bidirectional feature pyramid network BiFPN is introduced to replace the original PANet network; S3: Detect crop diseases using the trained model.

2. The crop disease detection method according to claim 1, wherein: In the multi-scale parallel hybrid convolution module, the input is first passed through three convolutional layers in parallel, and then the outputs of the three convolutional layers are fused, and then passed through the normalization layer, activation function, and 1×1 standard convolution in sequence before output.

3. The crop disease detection method according to claim 2, wherein: The three convolutional layers include one layer of partial convolution and two layers of depthwise separable convolution with different kernel sizes.

4. The crop disease detection method according to claim 3, wherein: The convolution kernel of the partial convolution is 3×3, and the convolution kernels of the two layers of depth-wise separable convolution are 3×3 and 5×5 respectively.

5. The crop disease detection method according to claim 2, wherein: The activation function uses the SiLU activation function.

6. A crop disease detection terminal device, characterized by: The method comprises a processor, a memory, and a computer program stored in the memory and running on the processor, wherein the processor implements the steps of the method according to any one of claims 1 to 5 when executing the computer program.

7. The crop disease detection terminal device according to claim 6, characterized in that: It also includes an image uploading module, a result visualization module and a data management module. Among them, the image uploading module is used to upload the crop images to be detected, the result visualization module is used to display the results of crop disease detection, and the data management module is used to store the results of crop disease detection.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images