Wheat leaf pest detection method based on YOLOv10

Abstract

According to the wheat leaf pest detection method based on YOLOv10, data processing and model architecture can be improved, the detection precision and speed can be improved, and the problems of data set limitation and insufficient model optimization in the prior art can be solved, so that the wheat pest recognition accuracy can be improved, and agricultural producers can be helped to monitor and prevent pests in real time. Compared with the prior art, the detection precision is high, the reasoning speed is high, the generalization ability is high, and computing resources are saved. According to the improved YOLOv10 model, a CoordAttention mechanism is introduced, so that the accuracy of disease and pest detection is remarkably improved. The model optimizes the reasoning process by removing the NMS, the reasoning speed of each image reaches 0.8 millisecond, and the real-time detection efficiency is greatly improved. Training is carried out in combination with multiple data sets, the model has good performance in different field environments, complex scenes of different illumination, angles, shielding and the like can be processed, and the applicability in actual production is improved. The optimization of the model enables the calculation resource consumption to be lower during reasoning.

Description

Technical Field

[0001] The present invention relates to the technical field of insect pest detection, and in particular to a wheat leaf insect pest detection method based on YOLOv10. Background Art

[0002] At present, the detection of wheat pests and diseases mostly relies on manual observation and manual statistics. This method is not only time-consuming and labor-intensive, but also often leads to misdiagnosis or missed diagnosis due to the subjectivity of human judgment, which affects the growth and yield of wheat. With the rapid development of deep learning technology, image-based automated detection methods have provided new solutions for agricultural production. Although existing technologies have been applied in the field of wheat pest and disease detection, they still have the following shortcomings: Dataset limitations: Most wheat disease detection methods use data sets from a single source, with a single image perspective, and most of the scenes are laboratory or fixed field environments. Due to the lack of diversity in the data set, the model shows poor generalization ability when dealing with complex field environments such as different lighting conditions and leaf occlusion, resulting in low detection accuracy.

[0003] Model complexity and computational overhead: Although existing models such as YOLOv3 and YOLOv4 have good detection accuracy, their complex model structure leads to excessive computational overhead, making it difficult to efficiently deploy them on edge devices such as drones and field monitoring equipment, limiting agricultural applications with high real-time requirements.

[0004] Post-processing dependency: The existing YOLO model uses non-maximum suppression (NMS) for target screening. Although this method can reduce detection redundancy, it will increase the delay of model reasoning and reduce detection efficiency in scenes with dense targets, such as complex environments with multiple leaves in the field. It is not suitable for real-time detection in large-scale fields. Summary of the invention

[0005] The purpose of the present invention is to provide a wheat leaf pest detection method based on YOLOv10, which can improve data processing and model architecture, enhance the accuracy and speed of detection, solve the problems of data set limitations and insufficient model optimization in the prior art, thereby improving the accuracy of wheat disease and pest identification, and helping agricultural producers to monitor and prevent diseases and pests in real time.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is: a wheat leaf pest detection method based on YOLOv10, comprising the following steps: Step 1, collecting pictures or video data of wheat leaves, and preprocessing the pictures or video data of wheat leaves; Step 2: Use the pre-processed wheat leaf images or video data obtained in step 1 as training images for model training. Select YOLOv10 as the model architecture, remove non-maximum suppression NMS, introduce dual allocation strategy, introduce CoordAttention mechanism, and introduce semi-supervised learning method. In the training process, add data enhancement strategy; use the multi-task loss function in YOLOv10 to optimize the model. Step 3: Export the YOLOv10 model trained in step 2 into a model format suitable for deployment and load it to the edge device to achieve real-time reasoning; receive the wheat image, preprocess it, and input it into the YOLOv10 model. The YOLOv10 model extracts the features of wheat leaves and ears, identifies and classifies the pest and disease areas, detects multiple pest and disease targets in one frame of image, and returns the location information and pest and disease category of each target; Step 4: The YOLOv10 model outputs the location information of each detected pest target and the predicted pest category. At the same time, it outputs a confidence score to indicate the YOLOv10 model's confidence in the pest detection result. In the detected image, the detected disease area is framed and the disease category and confidence are marked. Step 5: Summarize the test results, generate a heat map of disease distribution and a trend analysis report; recommend relevant prevention and control measures based on the test results.

[0007] Preferably, in step 1, a drone equipped with a high-resolution camera is used to fly over the farmland regularly to take pictures or videos of the wheat field; the collected data covers different angles, lighting and environmental conditions to ensure data diversity, and high-resolution equipment is used in conjunction with manual work to directly capture images of wheat plants on the ground, especially to supplement data in the early stages of the disease or in complex scenarios.

[0008] Preferably, the preprocessing of the picture or video data of wheat leaves in step 1 includes image cleaning and denoising, cropping and enhancement, and image annotation.

[0009] Preferably, in step 2, YOLOv10 adopts a single-stage detection method to simultaneously complete target detection and classification in one forward propagation.

[0010] Preferably, the data enhancement strategy in step 2 includes: randomly cropping different areas of the training image to simulate various scenes; enhancing the adaptability of the model to different lighting conditions by adjusting contrast, brightness and saturation; rotating and mirror-flipping the image to enhance the model's perception of different directional features of wheat in the image.

[0011] Preferably, in step 2, the multi-task loss function in YOLOv10 is used to optimize the model including positioning error, classification error and target confidence error, and the optimization process is performed by gradient update through Adam or SGD optimizer.

[0012] Preferably, the YOLOv10 model regularly acquires new data through edge computing, and updates the model in combination with the cloud to ensure that the model adapts to different farmland environments; the YOLOv10 model combines user feedback data to perform continuous semi-supervised learning and model fine-tuning to gradually optimize the detection accuracy of the model.

[0013] The beneficial effects of the present invention are: Compared with the existing technology, this solution has the following advantages: 1. High detection accuracy: The improved YOLOv10 model significantly improves the accuracy of pest and disease detection by introducing the CoordAttention mechanism, especially in the detection of wheat yellow rust, the model's accuracy exceeds 90%.

[0014] 2. Fast inference speed: The model optimizes the inference process by removing NMS. The inference speed of each image reaches 0.8 milliseconds, which greatly improves the efficiency of real-time detection and is suitable for deployment in edge devices.

[0015] 3. Strong generalization ability: The model is trained with a variety of data sets, and performs well in different field environments. It can handle complex scenes such as different lighting, angles, and occlusions, which improves its applicability in actual production.

[0016] 4. Save computing resources: The optimization of the model makes it consume less computing resources during inference, which is suitable for the deployment of low-power devices, especially for applications on drones and remote monitoring equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 Schematic diagram of the experimental image processing atlas of the present invention.

[0019] Figure 2 Schematic diagram of the steps of preparing and preprocessing the experimental data set of the present invention.

[0020] Figure 3 It is a YOLOv10 model structure diagram of the present invention.

[0021] Figure 4 Schematic diagram of the confusion matrix of the present invention.

[0022] Figure 5 Schematic diagram of the normalized confusion matrix of the present invention.

[0023] Figure 6 It is a schematic diagram of the F1-Confidence curve of the present invention.

[0024] Figure 7 It is a schematic diagram of the Precision-Confidence curve of the present invention.

[0025] Figure 8 Schematic diagram of the Precision-Recall curve of the present invention.

[0026] Fig. 9 Schematic diagram of the Recall-Confidence curve of the present invention.

[0027] Fig.10 Schematic diagram of the training and validation loss curves of the present invention.

[0028] Fig.11 It is a schematic diagram of the effect of the actual machine installation of the model of the present invention. DETAILED DESCRIPTION

[0029] The present invention is further described in detail below in conjunction with embodiments and with reference to the accompanying drawings.

[0030] The present invention discloses a wheat leaf pest detection method based on YOLOv10. The embodiment comprises the following steps: Step 1, collecting pictures or video data of wheat leaves, and preprocessing the pictures or video data of wheat leaves; Step 2: Use the pre-processed wheat leaf images or video data obtained in step 1 as training images for model training. Select YOLOv10 as the model architecture, remove non-maximum suppression NMS, introduce dual allocation strategy, introduce CoordAttention mechanism, and introduce semi-supervised learning method. In the training process, add data enhancement strategy; use the multi-task loss function in YOLOv10 to optimize the model. Step 3: Export the YOLOv10 model trained in step 2 into a model format suitable for deployment and load it to the edge device to achieve real-time reasoning; receive the wheat image, preprocess it, and input it into the YOLOv10 model. The YOLOv10 model extracts the features of wheat leaves and ears, identifies and classifies the pest and disease areas, detects multiple pest and disease targets in one frame of image, and returns the location information and pest and disease category of each target; Step 4: The YOLOv10 model outputs the location information of each detected pest target and the predicted pest category. At the same time, it outputs a confidence score to indicate the YOLOv10 model's confidence in the pest detection result. In the detected image, the detected disease area is framed and the disease category and confidence are marked. Step 5: Summarize the test results, generate a heat map of disease distribution and a trend analysis report; recommend relevant prevention and control measures based on the test results.

[0031] In the step 1, a drone equipped with a high-resolution camera is used to fly over the farmland regularly to take pictures or videos of the wheat field; the collected data covers different angles, lighting and environmental conditions to ensure data diversity, and the images of the ground wheat plants are directly taken with high-resolution equipment in cooperation with manual work, especially in the early stage of the disease or in complex scenes to supplement the data. The preprocessing of the picture or video data of the wheat leaves includes image cleaning and denoising, cropping and enhancement, and image annotation. The collected images may have noise or blur, so the images are first denoised to ensure the quality of the image. In order to improve the detection accuracy, the image will be cropped to a size of 640x640, focusing on the leaves and ears of wheat. In addition, the diversity of the data is increased by image enhancement techniques such as rotation, scaling, brightness adjustment, etc., and the images are blurred in some scenes to enhance the model's ability to cope with blurred scenes. Each image in the data set will be labeled with the disease type and disease location to ensure that the training can accurately guide the learning of the YOLOv10 model.

[0032] In step 2, YOLOv10 uses a single-stage detection method to complete target detection and classification in one forward propagation. The design goal is to achieve efficient and real-time pest and disease detection. Compared with the previous YOLO version, this project selected the L variant of YOLOv10, which combines the following improvements: 1. Continuous double allocation without NMS post-processing: The traditional YOLO model uses non-maximum suppression NMS to filter out redundant detection boxes. By removing NMS and introducing a double allocation strategy, the model is more efficient during inference and reduces inference latency.

[0033] 2. CoordAttention mechanism: The CoordAttention mechanism is introduced to combine position information and content information for attention calculation, which improves the model's ability to capture structural and sequential characteristics when processing images, especially the accurate identification of disease edges and spots.

[0034] 3. Semi-supervised learning: When some data are not labeled or insufficiently labeled, the semi-supervised learning method is used to improve the model's efficiency in utilizing unlabeled data, thereby enhancing its generalization ability for different scenarios.

[0035] During the training process, data enhancement strategies were added, including: 1. Random cropping: Randomly crop different areas of the training image to simulate various scenes.

[0036] 2. Color transformation: By adjusting contrast, brightness and saturation, the model's adaptability to different lighting conditions can be enhanced.

[0037] 3. Rotation and flipping: Rotate and mirror flip the image to improve the model's ability to perceive the different directional features of wheat in the image.

[0038] In step 2, the multi-task loss function in YOLOv10 is used to optimize the model, including the Bounding Box regression of the positioning error, the disease category predicted by the classification error, and whether the target confidence error has the target disease. The optimization process is updated by the Adam or SGD optimizer for gradient update.

[0039] The trained YOLOv10 model in step 3 is exported to a model format suitable for deployment, such as ONNX or TensorRT, and loaded into an edge device such as a drone equipped with a GPU to achieve real-time reasoning. After receiving the wheat image, it is first preprocessed and input into the YOLOv10 model. The YOLOv10 model extracts the features of wheat leaves and ears in the input image, identifies the pest and disease areas and classifies them. YOLOv10 can detect multiple pest and disease targets in one frame of image at the same time, and returns the location information bounding box and pest and disease category of each target. During the reasoning process, the CoordAttention mechanism introduces the combination of location information and content information, so that the model can better capture the structural relationship between wheat leaves and diseases in the image during reasoning. This is particularly important for wheat pest and disease detection, because pests and diseases are often reflected in specific spots, discolored areas and other tiny features of leaves and ears. Compared with the traditional YOLO model, this solution removes the post-processing step of NMS and directly outputs the detection box optimized by the dual allocation strategy. This greatly reduces the reasoning delay and improves the detection speed. In the pest and disease detection scenario, real-time performance is crucial, especially when the drone is flying quickly, the detection delay directly affects the detection effect and efficiency.

[0040] The model output in step 4 includes the location information bounding box coordinates of each detected pest target and the predicted pest category wheat yellow rust, septicemia, stem rust or powdery mildew. At the same time, the confidence score is output to indicate the degree of certainty of the model in the pest detection result. In the detected image, the detected disease area is framed and the disease category and confidence are marked. It can be displayed through edge devices or transmitted to the cloud for users to view and analyze.

[0041] The YOLOv10 model in step 5 regularly obtains new data through edge computing, and updates the model in combination with the cloud to ensure that the model adapts to different farmland environments; the YOLOv10 model combines user feedback data to perform continuous semi-supervised learning and model fine-tuning to gradually optimize the detection accuracy of the model.

[0042] In addition to drones, it can also be deployed in ground monitoring equipment or mobile devices to achieve multi-platform collaborative detection and improve coverage. It can be integrated with existing agricultural management platforms to provide comprehensive disease monitoring, prediction and decision support.

[0043] In the specific experiment, 2,647 pest and disease images were selected, including annotation information of four types of pests and diseases, and divided into training set, validation set and test set in a ratio of 6:3:1. The training environment used was a cloud server, configured with a 16-core CPU, 48GB memory and 2 GPUs with 24GB video memory. The model was set for 500 epochs during training, with each batch containing 50 images. Adaptive learning rate and early stopping strategy were used to improve training efficiency.

[0044] like Figure 1 Shown is a schematic diagram of the experimental image processing atlas: ①The picture in the upper left corner shows the number of instances of different disease categories. There may be multiple pest instances in one picture. The four diseases are wheat powdery mildew, wheat septoria, wheat stem rust and wheat yellow rust.

[0045] ②The upper right corner shows the distribution of all target bounding boxes in the image. Each blue box represents the bounding box of a disease target in the image.

[0046] ③The scatter plot (y vs x) in the lower left corner: This figure shows the position distribution of the center point of the disease target in the image, with the horizontal axis being the x coordinate and the vertical axis being the y coordinate.

[0047] ④ The scatter plot in the lower right corner (height vs width): This figure shows the distribution of the width and height of the target bounding box. The points in the figure represent the width and height of the marked disease target.

[0048] The last three figures show that the sampling perspective of the dataset is wide, which is beneficial to the training of the model.

[0049] like Figure 2 The following is a schematic diagram of the experimental dataset resume and preprocessing steps: ① The histogram on the diagonal shows the distribution of a single variable: x variable: The histogram shows that the values ​​of this variable are evenly distributed between 0 and 1, with the peak slightly concentrated in the middle, which means that the y values ​​of most data points are around 0.5. The histogram on the diagonal shows the distribution of a single variable: y variable: The histogram shows that the values ​​of this variable are evenly distributed between 0 and 1, with the peak slightly concentrated in the middle, which means that the y values ​​of most data points are around 0.5. Width variable: The histogram of width shows a typical triangular distribution, indicating that the width values ​​of most data are concentrated in the middle area, but the data are less at both ends. Height variable: The distribution of height shows a skewed pattern, with more concentrated data points on the right, indicating that most samples have relatively small height values, while larger height values ​​appear less frequently.

[0050] ② The scatter plot on the non-diagonal line shows the relationship between the two variables. The deeper the scatter plot, the greater the density of target points distributed at its location. The correlation graphs of x and width, y and height show that the closer the center point of the target box is to the center of the image, the larger the target box range is, which also meets the image sampling requirements. However, there is no strict correlation between width and height, y and width, x and height, and x and y.

[0051] like Figure 3 The following is the YOLOv10 model structure diagram: Figure 4 The figure shows a confusion matrix diagram, which shows the classification between different categories. The values ​​on the diagonal represent the number of correct predictions made by the model on each category, and the values ​​on the off-diagonal of a row represent the number of times the model misclassifies samples of that category as other categories. The off-diagonal values ​​of a column represent the number of times samples of other categories are misclassified as that category, and backgrad represents unrecognized samples. Figure 5 The figure shows a normalized confusion matrix diagram, which shows the correct classification rate of different categories. The values ​​on the diagonal represent the proportion of the model's correct predictions for each category, and the values ​​on the off-diagonal lines of a row represent the proportion of samples of that category that the model misclassifies as other categories. The off-diagonal values ​​of a column represent the proportion of samples of other categories that are misclassified as that category, and backgrad represents the proportion that is not recognized. The confusion matrix and the normalized confusion matrix correspond to the model accuracy analysis in Example 2.

[0052] like Figure 6The figure shows a schematic diagram of the F1-Confidence curve. The F1-confidence curve shows the F1 score of the model at different confidence thresholds. The curve shows that as the confidence increases, the F1 value first increases and then decreases. The curve shows the F1 scores of different categories of wheat diseases at different confidence levels. The F1 score of wheat-yellowrust is the highest among all categories, followed by wheat-septoria, wheat-stemrust, and wheat-powdery-mildew. The thick blue line represents the overall performance of all categories. The F1 score at a confidence level of 0.510 is 0.83, indicating that the overall performance of the model at this confidence level is good. The F1 scores of most categories peak at low confidence levels, and the F1 scores gradually decrease as the confidence level increases. This figure corresponds to the model effect analysis in Example 2.

[0053] like Figure 7 The figure shows a Precision-Confidence curve diagram: The precision-confidence curve shows the accuracy of the model at different confidence levels. Wheat-Fusarium head blight shows higher accuracy at higher confidence levels, while wheat-powdery mildew performs relatively poorly. The curves for all categories reach the highest accuracy (0.99) at a confidence level of 1.000, indicating that the model is very confident in some predictions.

[0054] like Figure 8 The figure shows a schematic diagram of the Precision-Recall curve. The precision-recall curve reveals the overall detection performance of various diseases. Wheat-yellow rust maintains a high recall rate at high confidence, while other categories such as wheat-powdery mildew have a faster drop in recall rate at high confidence. The curves of all categories show that the recall rate of the model is about 0.80 at low confidence (0.000), and the model is easier to correctly classify at low confidence.

[0055] like Fig. 9 The figure shows a Recall-Confidence curve diagram. The precision-recall curve reveals the overall detection performance of various diseases. Wheat-yellow rust maintains a high recall rate at high confidence, while other categories such as wheat-powdery mildew have a faster drop in recall rate at high confidence. The curves of all categories show that the recall rate of the model is about 0.80 at low confidence (0.000), and the model is easier to correctly classify at low confidence.

[0056] like Fig.10The figure shows the training and validation loss curves, which show the convergence of the model during the training process. As the number of training rounds increases, the classification loss and positioning loss gradually decrease, and the validation loss keeps pace with the training loss, indicating that the model is not overfitting.

[0057] Overall Performance: There are 794 validation images in total, with 3528 instances (the number of instances refers to the number of target objects annotated). Box P (Bounding Box Precision): The bounding box precision of the model is 0.931, indicating that the model has a high accuracy in locating the target object. R (Recall): The recall rate is 0.746, indicating that the model can detect most targets. mAP50: The average precision (with an IoU threshold of 0.5) is 0.782, indicating that the model performs well when the overlap rate between the predicted box and the true box reaches 50%. mAP50-95: Under a stricter IoU range (0.5 to 0.95), the average precision of the model is 0.723, indicating that the model still performs well in locating the target object with high accuracy.

[0058] Performance of each category: Wheat-powdery-mildew: Box P: 0.89, recall rate 0.677, mAP50 0.686, mAP50-95 0.63, indicating that the recall rate of this category is low and the precision is acceptable. Wheat-septoria: Box P: 0.959, recall rate 0.703, mAP50 0.757, mAP50-95 0.707, high precision but low recall rate. Wheat-stemrust: Box P: 0.934, recall rate 0.757, mAP50 0.804, mAP50-95 0.731, relatively balanced and good performance. Wheat-yellowrust: Box P: 0.941, recall rate 0.847, mAP50 is 0.88, mAP50-95 is 0.825, indicating that the precision and recall rate of this category are both high, and the model performs best in this category. Speed: The processing speed is 0.8 milliseconds for preprocessing, 0.8 milliseconds for inference, 0 milliseconds for loss calculation, and 0.2 milliseconds for post-processing per image. The overall speed is fast and the inference efficiency is high.

[0059] Judging from the verification results, the model performs well in the wheat disease classification task, especially in the wheat-yellowrust category, with high precision and recall. Some categories (such as powdery mildew and fusarium head blight) perform slightly worse and may need further optimization to improve the recall rate. In addition, the model still maintains good performance under the more stringent mAP50-95 evaluation index, indicating that it has good robustness in target positioning and bounding box prediction. Fig.11 The figure shows the effect of the model being installed on a real machine. It can be seen that the prediction effect of this model is very good.

[0060] After 3.75 hours of training, the model’s overall detection accuracy reached 80%.

[0061] For the detection of wheat yellow rust, a pre-annotated image set was used for verification. The average precision (mAP50) of the YOLOv10 model on this type of pests and diseases reached 0.88, and the recall rate was 0.847, indicating that the model can accurately detect wheat yellow rust in most cases. This detection result shows the advantages of the present invention in specific disease categories.

[0062] The trained YOLOv10 model was deployed on drone equipment, and the camera on the drone was used to monitor wheat in the field in real time. The model can quickly detect different types of pests and diseases under complex lighting conditions, with an inference time of less than 1 millisecond per image, meeting the needs of real-time field monitoring. In practical applications, the system helps agricultural producers detect pest and disease problems in advance, reduce the use of pesticides, and increase yields.

[0063] It should be noted that the parts not described in detail in the above embodiments are all prior art.

[0064] The above description is only a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention should be covered within the protection scope of the present invention.

Claims

1. A wheat leaf pest detection method based on YOLOv10, characterized in that: The following steps are involved: Step 1, collecting pictures or video data of wheat leaves, and preprocessing the pictures or video data of wheat leaves; Step 2: Use the pre-processed wheat leaf images or video data obtained in step 1 as training images for model training. Select YOLOv10 as the model architecture, remove non-maximum suppression NMS, introduce dual allocation strategy, introduce CoordAttention mechanism, and introduce semi-supervised learning method. In the training process, add data enhancement strategy; use the multi-task loss function in YOLOv10 to optimize the model. Step 3: Export the YOLOv10 model trained in step 2 into a model format suitable for deployment and load it to the edge device to achieve real-time reasoning; receive the wheat image, preprocess it, and input it into the YOLOv10 model. The YOLOv10 model extracts the features of wheat leaves and ears, identifies and classifies the pest and disease areas, detects multiple pest and disease targets in one frame of image, and returns the location information and pest and disease category of each target; Step 4: The YOLOv10 model outputs the location information of each detected pest target and the predicted pest category. At the same time, it outputs a confidence score to indicate the YOLOv10 model's confidence in the pest detection result. In the detected image, the detected disease area is framed and the disease category and confidence are marked. Step 5: Summarize the test results and generate a heat map and trend analysis report of disease distribution; Recommend relevant prevention and control measures based on the test results.

2. The wheat leaf insect pest detection method based on YOLOv10 according to claim 1, characterized in that: In step 1, a drone equipped with a high-resolution camera is used to fly over the farmland regularly to take pictures or videos of the wheat field; the collected data covers different angles, lighting and environmental conditions to ensure data diversity, and high-resolution equipment is used in conjunction with manual work to directly capture images of wheat plants on the ground, especially to supplement data in the early stages of the disease or in complex scenarios.

3. The wheat leaf insect pest detection method based on YOLOv10 according to claim 1, characterized in that: The preprocessing of the wheat leaf image or video data in step 1 includes image cleaning and denoising, cropping and enhancement, and image annotation.

4. The wheat leaf insect pest detection method based on YOLOv10 according to claim 1, characterized in that: In step 2, YOLOv10 adopts a single-stage detection method to simultaneously complete target detection and classification in one forward propagation.

5. The wheat leaf insect pest detection method based on YOLOv10 according to claim 1, characterized in that: The data enhancement strategy in step 2 includes: randomly cropping different areas of the training image to simulate various scenes; enhancing the adaptability of the model to different lighting conditions by adjusting the contrast, brightness and saturation; rotating and mirror-flipping the image to enhance the model's perception of different directional features of wheat in the image.

6. The wheat leaf insect pest detection method based on YOLOv10 according to claim 1, characterized in that: In step 2, the multi-task loss function in YOLOv10 is used to optimize the model including positioning error, classification error and target confidence error. The optimization process is performed by gradient update through Adam or SGD optimizer.

7. The wheat leaf insect pest detection method based on YOLOv10 according to claim 1, characterized in that: The YOLOv10 model regularly acquires new data through edge computing and updates the model in combination with the cloud to ensure that the model adapts to different farmland environments; the YOLOv10 model combines user feedback data to perform continuous semi-supervised learning and model fine-tuning to gradually optimize the detection accuracy of the model.

Images