Wheat scab detection method and system based on unmanned aerial vehicle remote sensing and improved YOLOv8

Abstract

The invention discloses a wheat scab lightweight detection method and system based on unmanned aerial vehicle remote sensing and improved YOLOv8, and relates to the technical field of artificial intelligence computer vision detection. The method comprises the following steps: acquiring an original RGB image in a wheat filling period in a natural field environment at fixed height and fixed time by using an air route plan of an unmanned aerial vehicle; performing screening, data enhancement and labeling on the original picture data set, and establishing a wheat scab image data set; an improved YOLOv8 lightweight network model is constructed; inputting the original image data set into the improved YOLOv8 network model for training; the improved YOLOv8 network model is tested, and learning parameters of the improved YOLOv8 network model are updated; and finally, detecting wheat scab data to be detected by using the updated improved YOLOv8 network model, and evaluating the complexity and accuracy of the model. The method is suitable for detecting the wheat scab in a natural field environment, the parameter quantity and the calculated quantity of the improved model are respectively reduced by 48.9% and 50.7% compared with those of an original YOLOv8s baseline model while high recognition precision is ensured, a high-precision algorithm model which is lighter than YOLOv8 is realized, and the method is suitable for popularization and application. And reference can be provided for deployment and application of mobile terminal detection equipment such as a wheat disease detection unmanned aerial vehicle.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence computer vision detection, and in particular relates to a lightweight detection method and system for wheat fusarium wilt based on unmanned aerial vehicle remote sensing and improved YOLOv8. Background Art

[0002] Wheat fusarium head blight is one of the most widespread and destructive diseases in global wheat production caused by a variety of Fusarium. In severe cases, it can reduce production by 80% to 90%, or even cause no harvest. The secondary metabolites produced by fusarium head blight are all toxins, which seriously endanger the health of humans and animals after being eaten by mistake. Therefore, timely and accurate monitoring of the occurrence and development of fusarium head blight during wheat growth and taking effective prevention and control measures are of great significance to ensure high-quality, efficient and safe production of wheat. In addition, it is learned from GB / T15796-2011 "Technical Specifications for Monitoring and Forecasting of Wheat Fusarium Head Blight" that counting the number of wheat fusarium head blight is an important basis for obtaining the severity of fusarium head blight and an important indicator for studying wheat resistance breeding (resistance to fusarium head blight). Therefore, the detection of wheat head fusarium head blight is conducive to evaluating the impact of the disease on wheat yield and quality, and guiding farmers and relevant personnel to take corresponding control and management measures.

[0003] At present, the main methods for identifying wheat head rust disease include manual field visual surveys and hyperspectral imaging. The method based on manual visual surveys is time-consuming and labor-intensive, and is not conducive to large-scale investigation and prevention. Hyperspectral imaging technology is widely used in the field of crop disease detection due to its strong anti-interference ability and fast detection speed. Combined with spectral analysis and image processing technology, it can accurately identify wheat head rust disease grains. Although hyperspectral technology can achieve high recognition accuracy, hyperspectral imagers are relatively expensive and bulky, making it difficult to perform real-time detection in the field.

[0004] With the development of artificial intelligence, computer vision and deep learning technologies are constantly being broken through, and more and more researchers are using deep learning technology and image processing methods to detect crop diseases and pests. Image processing technology and deep learning technology are used to assist wheat growers and plant protection personnel in the prevention and control of wheat fusarium rust, which has far-reaching significance for the development of automated identification of agricultural diseases. Regarding the research on wheat disease detection methods, although existing literature has achieved relatively accurate detection goals, most of them are limited to laboratory environments, and less consideration is given to achieving wheat fusarium rust detection with smaller computational complexity, lower latency and higher detection accuracy in complex field natural environments and multi-target environments.

[0005] In order to solve the above problems, the present invention proposes a lightweight detection method and system for wheat fusarium wilt based on UAV remote sensing and improved YOLOv8. Summary of the invention

[0006] The purpose of the present invention is to provide a lightweight detection method and system for wheat fusarium scabiei based on unmanned aerial vehicle remote sensing and improved YOLOv8 to solve the problem of wheat fusarium scabiei detection in complex field natural environment and multi-target environment, which is rarely considered in the prior art proposed in the above background technology; the present invention is based on the latest YOLOv8 model in the YOLO series, which has higher accuracy, smaller parameters and model size, to improve and optimize the wheat fusarium scabiei detection method, in order to develop a wheat fusarium scabiei detection model with high accuracy and fast detection speed for mobile devices with limited computing power.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] The wheat fusarium rust detection method based on UAV remote sensing and improved YOLOv8 includes the following steps:

[0009] S1. Use drones to collect original images of wheat during the grain filling period in natural field environments;

[0010] S2, screening, data enhancement and annotation of the original image dataset to establish a wheat fusarium rust image dataset;

[0011] S3. Build a lightweight network model that improves YOLOv8;

[0012] S4, input the original image data set into the improved YOLOv8 network model for training;

[0013] S5. Test the improved YOLOv8 network model and update the learning parameters of the improved YOLOv8 network model;

[0014] S6. Use the updated improved YOLOv8 network model to test the wheat fusarium data to be tested, and evaluate the complexity and accuracy of the model.

[0015] Preferably, the S1 specifically includes the following contents:

[0016] S101, selecting a wheat field in the grain-filling period in a natural field environment that is infected with fusarium smut;

[0017] S102. Use a drone to plan an “S”-shaped flight route, take directional photos at the same altitude and intervals, and complete data collection.

[0018] Preferably, S2 specifically includes the following contents:

[0019] S201, filter out three types of image data of wheat scab, namely, small, medium and dense data, to ensure that various samples exist in the data set to meet the diversity of data;

[0020] S202, performing data enhancement by rotating images and adjusting image parameters to enrich the data set;

[0021] S203. Labelimg is used to label each wheat ear with scab in the established image dataset according to the disease condition of wheat in each image, and the labels are labeled as one category, namely Scab. When labeling the wheat ear with scab, the entire wheat ear is included in the frame, and the wheat ears with more than one-third exposed are labeled.

[0022] Preferably, S3 specifically includes the following contents:

[0023] Based on the YOLOv8 model, the backbone network is completely replaced by the StarNet network; the StarNet network adopts a 4-level hierarchical architecture, uses convolutional layers for downsampling, and uses improved modules for feature extraction; replaces layer normalization with batch normalization and places it after deep convolution, and adds a deep convolution (DW-Conv) at the end of each block; sets the channel expansion factor to 4, and doubles the network width at each stage;

[0024] In the neck network (Neck) part, a new lightweight convolution GSCOnv module is used to replace the ordinary convolution operation, and the VoV-GSCSP module is used to replace the original C2f module; the GSConv module is a hybrid convolution composed of sparse convolution (SC), depthwise separable convolution (DSC) and shuffle; the VoV-GSCSP module combines DensNet, VoVNet, and CSPNet, uses the lightweight convolution method GSConv to replace the standard convolution, and further introduces GS bottleneck on top of GSConv to enhance the learning ability of the model, reduce the computational complexity, simplify the network structure, and maintain sufficient accuracy.

[0025] Preferably, the S4 specifically includes the following contents:

[0026] S401, dividing the data set into a training set, a validation set, and a test set according to an 8:1:1 ratio;

[0027] S402, using the pre-trained network weights as initial weights, and using the transfer learning method to train the improved YOLOv8 model using the training set;

[0028] The pre-trained weights were iterated for 100 epochs. The backbone network was frozen in the first 50 epochs of training, with the learning rate set to 1e-4 and the decay rate set to 0.96. After 50 epochs, the backbone network was unfrozen, with the learning rate set to 1e-5 and the decay rate set to 0.96.

[0029] S403. Use the validation set to validate the trained improved YOLOv8 model.

[0030] Preferably, the S5 specifically includes the following contents:

[0031] Use the divided test set to test and continuously optimize the improved YOLOv8 network model, determine the learning parameters of the improved YOLOv8 network model, and record various parameters of the improved YOLOv8 model, the original model, and the comparison model (YOLOv9, YOLOv10), such as the amount of calculation, the number of parameters, and the accuracy.

[0032] Preferably, the updated improved YOLOv8 network model used in S6 to detect the wheat scab data to be detected specifically includes the following contents:

[0033] S601, click input image to select a single wheat scab image to be detected;

[0034] S602, obtaining image detection results;

[0035] S603, return to the main interface.

[0036] The wheat fusarium head blight detection system based on UAV remote sensing and improved YOLOv8 applied in the above method includes:

[0037] An image input unit, used for inputting and detecting an image to be detected;

[0038] The detection module improves the YOLOv8 model, including the input layer, the backbone network layer, the neck connection layer, the head detection head and the output layer. The StarNet model is used to replace the backbone network of the YOLOv8 network structure, the GSCOnv is used in the neck network part to replace the ordinary convolution operation, and the VoV-GSCSP module is used to replace the original C2f module for image detection;

[0039] The output unit is used to output the wheat fusarium rust detection results and display the detection results to the user using a visual interface.

[0040] A computer device is further proposed, comprising a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the instruction, program, code set or instruction set is loaded and executed by the processor to implement the above-mentioned wheat fusarium detection method based on drone remote sensing and improved YOLOv8.

[0041] A computer-readable storage medium is further proposed, in which at least one instruction, at least one program, code set or instruction set is stored, and the instruction, program, code set or instruction set is loaded and executed by a processor to implement the above-mentioned wheat fusarium detection method based on drone remote sensing and improved YOLOv8.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] (1) The present invention proposes a lightweight detection method for wheat fusarium scabiei based on UAV remote sensing and improved YOLOv8. When detecting wheat fusarium scabiei, it can not only accurately detect the characteristics of wheat fusarium scabiei in complex scenes and multi-target scenes, but also provide a reference for the deployment and application of intelligent wheat disease detection in mobile terminal detection equipment such as UAVs in terms of rapid detection applications.

[0044] (2) The present invention also proposes a lightweight detection network for wheat fusarium rust based on UAV remote sensing and improved YOLOv8. The improved YOLOv8 model introduces the StarNet network to reduce the complexity and hyperparameters of the backbone network; GSCOnv is used to replace the ordinary convolution operation in the neck network part, and the VoV-GSCSP module is used to replace the original C2f module, thereby reducing the model calculation complexity and reasoning time while maintaining its accuracy. While ensuring high recognition accuracy, the parameter amount and calculation amount of the improved model are reduced by 48.9% and 50.7% respectively compared with the original YOLOv8s baseline model, realizing a high-precision and more lightweight algorithm model than YOLOv8. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flow chart of a lightweight method for detecting wheat fusarium rust based on UAV remote sensing and improved YOLOv8 proposed in Example 1 of the present invention;

[0046] Figure 2 The device for collecting original images and the training images proposed in Embodiment 1 of the present invention;

[0047] Figure 3 The original image of the test set proposed in Example 1 of the present invention;

[0048] Figure 4 This is a network structure diagram of the improved YOLOv8 model proposed in Example 1 of the present invention;

[0049] Figure 5 The StarNet network structure diagram proposed in Embodiment 1 of the present invention;

[0050] Figure 6 GSConv module structure diagram proposed in Example 1 of the present invention;

[0051] Figure 7 This is a structural diagram of the VoV-GSCSP module proposed in Example 1 of the present invention;

[0052] Figure 8 These are example images of input and output of the improved YOLOv8 network model proposed in Example 1 of the present invention to the wheat fusarium wilt to be detected. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] Embodiment 1:

[0055] See also Figure 1 The present invention proposes a lightweight detection method for wheat scab based on UAV remote sensing and improved YOLOv8, comprising the following steps:

[0056] S1. Use drones to collect original images of wheat during the grain filling period in a natural field environment.

[0057] First, a wheat field in the grain filling period under a natural field environment with fusarium scaly was selected; using a drone, an "S"-shaped flight route was planned, and the flight area and flight altitude were determined to ensure that the image of the wheat field could be fully and clearly captured. Secondly, the camera equipment carried by the drone was configured, including setting shooting parameters, adjusting focal length and angle, etc., to ensure image quality and coverage. Then, the actual flight mission was carried out, allowing the drone to fly on a predetermined path, and at the same time, the camera was started to take directional photos at the same altitude and interval to complete data collection.

[0058] S2. Screen, enhance and annotate the original image dataset to establish a wheat fusarium rust image dataset.

[0059] We screened out three types of wheat fusarium data: less, medium, and dense, to ensure that all samples exist in the data set to meet the diversity of the data; due to the limited number of certain specific samples, image enhancement technology is often used to process the original data before deep learning network training to increase the size of the data set and improve the generalization ability of the model. We used image rotation and image parameter adjustment to enhance the data set and enrich the data set. We obtained 2,500 images. The equipment for collecting original images and training images are as follows: Figure 2 As shown, the original image of the test set is Figure 3 shown.

[0060] According to the disease condition of wheat in each image, Labelimg is used to annotate each wheat ear with scab in the established image dataset and mark them as one category, namely Scab. Since wheat ears are close to each other and block each other, when annotating wheat ears with scab, the entire ear is included in the frame as much as possible, and only the ears with more than one-third exposed are labeled.

[0061] S3. Build a lightweight network model that improves YOLOv8.

[0062] The improved YOLOv8 network model is set to five parts, namely the input layer, the backbone network layer, the neck connection layer, the head detection head and the output layer. The constructed YOLOv8 network model is as follows Figure 4 shown.

[0063] The input layer is responsible for receiving image data and preprocessing the image, including image resizing, image normalization, and data enhancement.

[0064] Based on the YOLOv8 model, using the StarNet network (such as Figure 5 StarNet completely replaces its backbone network (Backbone) with a 4-level hierarchical architecture, using convolutional layers for downsampling and improved modules for feature extraction. In order to meet efficiency requirements, layer normalization is replaced by batch normalization and placed after the deep convolution; and a deep convolution (DW-Conv) is added at the end of each block. The channel expansion factor is always set to 4, and the network width is doubled at each stage.

[0065] In the neck network (Neck) part, a new lightweight convolution GSCOnv module (such as Figure 6 ) instead of the normal convolution operation, and the VoV-GSCSP module (as shown Figure 7The VoV-GSCSP module replaces the original C2f module with a sparse convolution (SC), a depthwise separable convolution (DSC), and a shuffle. The VoV-GSCSP module combines the theoretical concepts of DensNet, VoVNet, CSPNet, etc., uses a lightweight convolution method GSConv to replace the standard convolution, and further introduces GSBottleneck on top of GSConv to enhance the learning ability of the model, reduce computational complexity, simplify the network structure, and maintain sufficient accuracy.

[0066] The Head part is responsible for the final target detection and classification tasks, including a detection head and a classification head:

[0067] Detection head: Contains a series of convolutional layers and deconvolutional layers to generate detection results.

[0068] Classification head: Global Average Pooling is used to classify each feature map, and the probability distribution of each category is output by reducing the dimension of the feature map.

[0069] S4. Input the original image dataset into the improved YOLOv8 network model for training.

[0070] The dataset is divided into training set, validation set and test set according to the ratio of 8:1:1. The pre-trained network weights are used as the initial weights, and the training set is used to train the improved YOLOv8 model using the transfer learning method.

[0071] The pre-trained weights were iterated for 100 epochs. The backbone network was frozen in the first 50 epochs of training, with the learning rate set to 1e-4 and the decay rate set to 0.96. After 50 epochs, the backbone network was unfrozen, with the learning rate set to 1e-5 and the decay rate set to 0.96.

[0072] The validation set is used to verify the trained improved YOLOv8 model. In order to avoid overfitting and evaluate the generalization ability of the model, the trained model is applied to the validation set, and the classification performance of the model is evaluated using indicators such as mAP and mIoU, and the model that performs well in the validation set is selected.

[0073] S5. Test the improved YOLOv8 network model and update the learning parameters of the improved YOLOv8 network model;

[0074] The improved YOLOv8 network model is tested and optimized using the divided test set to determine the learning parameters of the improved YOLOv8 network model, and various parameters of the improved model, the original model, and the comparison model (YOLOv9, YOLOv10), such as the amount of calculation, the amount of parameters, and the accuracy, are recorded. The following are the test results:

[0075] Table 1

[0076] Para / MB GFLOPS / GB P / ％ R / ％ MAP50 / % F1 YOLOv8 11.12 28.4 85.4 86.9 91.4 0.86 YOLOv9 15.57 67.4 87.3 86.6 93.6 0.87 YOLOv10 8.03 24.4 88.4 86.3 93.5 0.87 Improving the model 5.68 14.0 86.8 84.8 92.1 0.86

[0077] The parameter amount (Para) and computational effort (GFLOPS) of the improved YOLOv8 model are reduced by 48.9% and 50.7% respectively compared with the original YOLOv8 model. At the same time, the precision (precision) and mean average precision (mAP) are improved by 1.4 and 0.7 percentage points respectively compared with the original model, which is better than other comparative target detection algorithms and can provide a reference for the deployment and application of mobile detection equipment such as wheat disease detection drones.

[0078] S6. Use the updated improved YOLOv8 network model to detect the wheat fusarium wilt data to be detected.

[0079] Input a single wheat fusarium image to be tested; click the test button to get the image test result, see Figure 8 .

[0080] Embodiment 2:

[0081] Based on Example 1, but different in that, a lightweight system for wheat fusarium head blight detection based on drone remote sensing and improved YOLOv8 for the method in Example 1 is further proposed, comprising:

[0082] An image input unit, used for inputting and detecting an image to be detected;

[0083] Detection module, improves the YOLOv8 model, including input layer, Backbone network layer, Neck connection layer, Head detection head and output layer. The StarNet model is used to replace the backbone network of the YOLOv8 network structure, GSCOnv is used to replace the ordinary convolution operation in the neck network part, and the VoV-GSCSP module is used to replace the original C2f module; it is used to detect images;

[0084] The output unit is used to output the wheat fusarium wilt detection result.

[0085] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A wheat scab detection method based on UAV remote sensing and improved YOLOv8, characterized in that: The following steps are involved: S1. Use drones to collect original images of wheat during the grain filling period in natural field environments; S2, screening, data enhancement and annotation of the original image dataset to establish a wheat fusarium rust image dataset; S3. Build a lightweight network model that improves YOLOv8; S4, input the original image data set into the improved YOLOv8 network model for training; S5. Test the improved YOLOv8 network model and update the learning parameters of the improved YOLOv8 network model; S6. Use the updated improved YOLOv8 network model to test the wheat fusarium data to be tested, and evaluate the complexity and accuracy of the model.

2. The wheat scab detection method based on unmanned aerial vehicle remote sensing and improved YOLOv8 according to claim 1, characterized in that: The S1 specifically includes the following contents: S101, selecting a wheat field in the grain-filling period in a natural field environment that is infected with fusarium smut; S102. Use a drone to plan an "S"-shaped flight route, take directional photos at the same altitude and intervals, and complete data collection.

3. The wheat scab detection method based on unmanned aerial vehicle remote sensing and improved YOLOv8 according to claim 1, characterized in that: The S2 specifically includes the following contents: S201, filter out three types of image data of wheat scab, namely, small, medium and dense data, to ensure that various samples exist in the data set to meet the diversity of data; S202, performing data enhancement by rotating images and adjusting image parameters to enrich the data set; S203. Labelimg is used to label each wheat ear with scab in the established image dataset according to the disease condition of wheat in each image, and the labels are labeled as one category, namely Scab. When labeling the wheat ear with scab, the entire wheat ear is included in the frame, and the wheat ears with more than one-third exposed are labeled.

4. The wheat scab detection method based on unmanned aerial vehicle remote sensing and improved YOLOv8 according to claim 1, characterized in that: The S3 specifically includes the following contents: Based on the YOLOv8 model, the backbone network is completely replaced by the StarNet network; the StarNet network adopts a 4-level hierarchical architecture, uses convolutional layers for downsampling, and uses improved modules for feature extraction; replaces layer normalization with batch normalization and places it after the depthwise convolution, and adds a depthwise convolution at the end of each block; sets the channel expansion factor to 4, and doubles the network width at each stage; In the neck network part, a new lightweight convolution GSCOnv module is used to replace the ordinary convolution operation, and the VoV-GSCSP module is used to replace the original C2f module; the GSConv module is a hybrid convolution composed of sparse convolution, depthwise separable convolution and shuffle; the VoV-GSCSP module combines DensNet, VoVNet, and CSPNet, uses the lightweight convolution method GSConv to replace the standard convolution, and further introduces GS bottleneck on top of GSConv to enhance the learning ability of the model, reduce the computational complexity, and simplify the network structure.

5. The wheat scab detection method based on unmanned aerial vehicle remote sensing and improved YOLOv8 according to claim 4, characterized in that: The S4 specifically includes the following contents: S401, dividing the data set into a training set, a validation set, and a test set according to an 8:1:1 ratio; S402, using the pre-trained network weights as initial weights, and using the transfer learning method to train the improved YOLOv8 model using the training set; The pre-trained weights were iterated for 100 epochs. The backbone network was frozen in the first 50 epochs of training, with the learning rate set to 1e-4 and the decay rate set to 0.

96. After 50 epochs, the backbone network was unfrozen, with the learning rate set to 1e-5 and the decay rate set to 0.

96. S403. Use the validation set to validate the trained improved YOLOv8 model.

6. The wheat scab detection method based on unmanned aerial vehicle remote sensing and improved YOLOv8 according to claim 5, characterized in that: The S5 specifically includes the following contents: Use the divided test set to test and continuously optimize the improved YOLOv8 network model, determine the learning parameters of the improved YOLOv8 network model, and record various parameters of the improved YOLOv8 model, the original model, and the comparison model, such as the amount of calculation, the number of parameters, and the accuracy.

7. The wheat scab detection method based on unmanned aerial vehicle remote sensing and improved YOLOv8 according to claim 6, characterized in that: The updated improved YOLOv8 network model described in S6 is used to detect the wheat scab data to be detected, which specifically includes the following contents: S601, click input image to select a single wheat scab image to be detected; S602, obtaining image detection results; S603, return to the main interface.

8. A wheat scab detection system based on UAV remote sensing and improved YOLOv8 applied in any of the methods described in claims 1-7, characterized in that: include: An image input unit, used for inputting and detecting an image to be detected; The detection module improves the YOLOv8 model, including the input layer, the backbone network layer, the neck connection layer, the head detection head and the output layer. The StarNet model is used to replace the backbone network of the YOLOv8 network structure, the GSCOnv is used in the neck network part to replace the ordinary convolution operation, and the VoV-GSCSP module is used to replace the original C2f module for image detection; The output unit is used to output the wheat fusarium rust detection results and display the detection results to the user using a visual interface.

9. A computer device, characterized in that: The computer device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the instruction, program, code set or instruction set is loaded and executed by the processor to implement the wheat fusarium detection method based on drone remote sensing and improved YOLOv8 as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one instruction, at least one program, code set or instruction set, and the instruction, program, code set or instruction set is loaded and executed by the processor to implement the wheat fusarium detection method based on drone remote sensing and improved YOLOv8 as described in any one of claims 1-7.

Images