A deep learning-based method and device for detecting citrus psylla

Abstract

The application discloses a kind of citrus psylla intelligent detection method and device based on deep learning, method includes the following steps: S1, acquisition citrus psylla color image under natural orchard scene, pre-treatment is carried out to acquisition image, and citrus psylla dataset is made;S2, improvement YOLOv5-s network model, using citrus psylla dataset is iteratively trained to improved YOLOv5-s network model, and optimal model is obtained and used as psyllid detection model;S3, using 2 pairs of servo steering engine is formed rotatable holder and carries variable zoom camera and forms image collector;Psyllid detection model is deployed to embedded device, and using the equipment control image collector;S4, the image obtained by image collector is input into psyllid detection model and is identified, determines whether there is citrus psylla in the image photographed and its position.The method and device of the application can be deployed in natural orchard, and the citrus plants in a certain distance range are efficiently and intelligently monitored to determine whether there is citrus psylla.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, specifically relating to a method and device for intelligent detection of citrus psyllids based on deep learning. Background Technology

[0002] Citrus Huanglongbing (HLB) is a devastating disease affecting citrus, severely impacting yield and quality, and even causing tree death. Detection and control of citrus psyllids are crucial for controlling the spread of HLB. Manual field inspection is currently the most common method for detecting citrus pests, such as tapping branches with plastic sticks for sampling and using yellow sticky traps or nets for counting. However, these methods rely on visual inspection, are inefficient, and cannot provide timely assessment of citrus pest conditions in orchards, hindering accurate and effective management of large citrus-growing areas.

[0003] In recent years, with the rapid development of machine vision technology, convolutional neural networks have been widely used in the detection of various crop diseases and pests, becoming a method to replace manual identification and traditional machine vision recognition. Korean scholars have proposed a citrus disease and pest classification method that can be installed on mobile devices for real-time detection using compressed convolutional neural networks, training and validating it on a dataset of 16,528 images containing 14 types of citrus diseases and pests. Chinese scholars have constructed a dataset of 2,141 images of citrus psyllids, improving the neck network of the YOLOv4-Tiny model, utilizing the detail information of shallow networks, and employing cross-mini-batch normalization to build a citrus psyllid recognition model. However, due to the small size of citrus psyllids, the above studies are mainly applied to the classification or detection of citrus diseases and pests at close range. Similar to traditional detection methods, it still requires manual handling of mobile devices close to the fruit trees to collect images for classification or detection, and it still cannot significantly improve the detection efficiency of citrus psyllids. Therefore, how to use intelligent detection methods to completely replace manual identification and location of citrus psyllids in the complex environment of orchards is a problem that professionals urgently need to solve. Summary of the Invention

[0004] The main objective of this invention is to overcome the shortcomings and deficiencies of the existing technology and propose a deep learning-based intelligent detection method and device for citrus psyllids. Based on the improved YOLOv5-s target detection network model, it can efficiently and intelligently detect all citrus plant images collected within a certain range of the image collector, thus solving the problem of mid-to-long-distance detection of citrus psyllids in natural orchard scenarios.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A deep learning-based intelligent detection method for citrus psyllids includes the following steps:

[0007] S1. Collect color images of citrus psyllids in natural orchard scenes, preprocess the collected images, and create a citrus psyllid dataset;

[0008] S2. Improve the YOLOv5-s network model. Use the citrus psyllid dataset to iteratively train the improved YOLOv5-s network model to obtain the optimal improved YOLOv5-s network model and use it as the psyllid detection model.

[0009] S3. Use two servo motors to form a rotatable gimbal and mount a variable-focus camera to form an image acquisition device; deploy the psyllid detection model onto an embedded device and use the device to control the image acquisition device;

[0010] S4. Input the image acquired by the image acquisition device into the psyllid detection model for identification to determine whether the captured image contains citrus psyllids and their location.

[0011] This invention also includes a deep learning-based intelligent detection device for citrus psyllids. The device employs the detection method provided by this invention and includes:

[0012] The image acquisition module is used to acquire images of citrus fruits in a natural orchard setting where citrus psyllids are to be detected. It includes an image acquisition device and a rotatable gimbal.

[0013] Image acquisition equipment, specifically a zoom camera or a DSLR camera;

[0014] The rotatable gimbal consists of two servo motors and is equipped with image acquisition equipment.

[0015] An embedded edge computing device is equipped with a trained psyllid detection model, which is used to control the image acquisition module to acquire citrus images in a natural orchard setting.

[0016] The psyllid localization module is used to determine whether citrus psyllids exist in the captured image and output their location based on the detection results of the psyllid detection model.

[0017] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0018] 1. This invention improves the network structure of the YOLOv5-s model by using dynamic convolution instead of traditional convolution, adding a second-order channel attention mechanism for small target detection, and using Wise-IoU instead of C-IoU loss function to improve the model's detection accuracy for small targets. During training, mixup data augmentation is used to reduce the model's dependence on all target features. Ultimately, while maintaining the advantages of the original model's simplified structure and fast detection speed, the model's target detection accuracy is improved, enabling the identification of citrus psyllids on trees at a distance in complex field environments.

[0019] 2. In view of the fact that the detection coverage of ordinary cameras is limited due to the large area of ​​the orchard, this invention designs a variable-focus image acquisition device with 360-degree rotation function, which consists of a variable-focus camera, multiple rotating servo motors and a controller, providing a hardware foundation for collecting images of multiple citrus trees in natural orchards. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention;

[0021] Figure 2 This is a diagram showing the overall structure of the improved YOLOv5-s network model in the embodiment.

[0022] Figure 3 This is a graph showing the mAP values ​​of the model tested using the validation set in the embodiment.

[0023] Figure 4 This is a schematic diagram of the intelligent detection device for citrus psyllids in the embodiment;

[0024] Figure 5 This is a schematic diagram of the input and output images of the present invention. Detailed Implementation

[0025] like Figure 1 As shown, this invention provides an intelligent detection method for citrus psyllids based on deep learning, comprising the following steps:

[0026] S1. Collect color images of citrus psyllids in natural orchard scenes, preprocess the collected images, and create a citrus psyllid dataset; specifically:

[0027] S11. Use a tripod to fix the shooting equipment to take pictures of citrus plants, simulating the image acquisition of color images of citrus plants by an image acquisition device in a real application scenario; the color images include citrus psyllid images with different pixel sizes taken at different focal lengths; the shooting equipment can be a zoom camera, SLR camera, etc.

[0028] S12. Preprocess the collected color images and divide the psyllid dataset into training, validation, and test sets according to a preset ratio; specifically:

[0029] S121. Use annotation software to annotate the citrus psyllids in the collected color images, including the psyllids' positions and border sizes. Specifically, use annotation software such as labelImg to annotate the citrus psyllids in the images. The label information, such as the location of the citrus psyllids and the smallest rectangle containing the psyllids, is stored in an XML file in VOC format to create a dataset. The dataset includes color image files and corresponding annotation information files.

[0030] S122. Divide the labeled color image set into a training set, a validation set, and a test set in a ratio of 7:1:2.

[0031] S123. Perform four operations on the color images in the training set: brightness adjustment, contrast adjustment, noise addition, and rotation to amplify the data.

[0032] S2. Improve the YOLOv5-s network model. Iteratively train the improved YOLOv5-s network model using the citrus psyllid dataset to obtain the optimal improved YOLOv5-s network model and use it as the psyllid detection model; specifically including:

[0033] S21. Improve the YOLOv5-s network model by replacing traditional convolution with dynamic convolution ODConv and adding a second-order channel attention mechanism for small object detection; replace the C-IoU loss function with Wise-IoU to obtain the improved YOLOv5-s network model; specifically:

[0034] The original YOLOv5-s network model uses traditional convolution to perform convolution operations on input image features. The weights of the convolution kernels are predefined. During the convolution process, the convolution operation between the kernels and the input data is fixed and independent of the input data. This convolution operation smooths the features, and small targets can easily be submerged in background noise, affecting the model's performance. To improve this, channel-space full-dimensional dynamic convolution is used to replace the traditional convolution method of the original YOLOv5-s network model. A multi-dimensional attention mechanism is introduced through a parallel strategy. When performing convolution operations, the four dimensions in the kernel space composed of n convolution kernels participating in the operation are fused, and the n convolution kernels are linearly weighted in parallel to avoid over-smoothing the features and improve the ability to express local details of the features.

[0035] S212. The original YOLOV5-s network model is prone to false detections and false negatives when dealing with small targets. In order to make the model learn the features of small targets such as citrus psyllids more effectively, a second-order channel attention mechanism is introduced into the original YOLOV5-s network model. Based on the correlation and covariance between channels, each channel in the feature map is weighted to better extract the interaction information between different channels, strengthen the interaction ability between different channels in the convolutional neural network, and thus improve the detection performance of the network model for small targets such as citrus psyllids.

[0036] S213. Replace the C-IoU of the original YOLOV5-s network model with Wise-IoU, and introduce a dynamic focusing mechanism. A dynamic focusing coefficient is calculated based on the difference between the distance and aspect ratio between two bounding boxes. The dynamic focusing coefficient is applied to the regression loss of the bounding box coordinates and aspect ratio information, so as to pay more attention to the regression of different parts of the bounding box.

[0037] The improved YOLOv5-s network model consists of a feature extraction module, a neck module, and a detection head module, totaling 25 layers. Layers 1 to 9 primarily utilize the CSPDarknet53 network for feature extraction, extracting useful features from the input image layer by layer while compressing the dimensionality of the feature vectors. Layers 10 to 23 form the neck module, employing dynamic convolution and adding a second-order channel attention mechanism for small object detection in layer 24. The neck module also includes a FastSpatial Pyramid Pooling (SPPF) module, which can handle different proportions and sizes of objects in the image. Layer 25 is the detection head module, comprising a regression head and a classification head, using Wise-IoU instead of the traditional C-IoU loss function. The classification head uses Focal Loss to focus on hard samples, improving detection performance; the regression head is responsible for regressing the bounding box coordinates.

[0038] By introducing a second-order channel attention mechanism, after the last layer (layer 23) of the original neck network structure completes the feature extraction, the second-order channel attention mechanism will perform weighted calculations on each channel in the obtained feature map in layer 24 in order to better mine the interaction information between different channels, and then input it into the detection head module in layer 25 for detection.

[0039] like Figure 2 The diagram shown is a schematic of the improved YOLOv5-s network model.

[0040] S22. Set the network model parameters for the improved YOLOv5-s network model; specifically:

[0041] S221. Use the k-means++ algorithm to cluster the target boxes in the training set; set the k value of the clustering algorithm according to the number of target boxes output by the model, and use the clustering results as the target box parameter values ​​of the improved YOLOv5-s network model; including:

[0042] S2211. Randomly select one sample from the training set as the first initial cluster center;

[0043] S2212. Calculate the shortest distance between each sample and the existing initial cluster centers;

[0044] S2213. Calculate the probability that each sample is selected as the next initial cluster center, and select the next initial cluster center according to the roulette wheel method;

[0045] S2214. Repeat steps S2212 and S2213 until k initial cluster centers are selected.

[0046] S2215. Calculate the distance between each sample and the k cluster centers, and assign it to the cluster center with the smallest distance to form k clusters;

[0047] S2216. Update the cluster center to the mean of all samples belonging to the cluster;

[0048] S2217. Repeat steps S2215 and S2216 until the cluster centers no longer change, and obtain the clustering results as the target box parameter values ​​of the improved YOLOv5-s network model.

[0049] Specifically, step S221 is as follows:

[0050] (1) Collect the training dataset, including images and corresponding bounding box information.

[0051] (2) Normalize the width and height of all annotation boxes to the range of 0-1. Calculate the width and height of each annotation box: width = xmax - xmin, height = ymax - ymin, where (xmin, ymin) represents the coordinates of the upper left corner of the annotation box, and (xmax, ymax) represents the coordinates of the lower right corner of the annotation box;

[0052] (3) Select the number of anchor frames to use (generally, select 9 anchor frames).

[0053] (4) Randomly select 9 anchor boxes as the initial cluster centers; 9 annotation boxes can be randomly selected from the dataset as the width and height of the initial cluster centers.

[0054] (5) Use the k-means++ algorithm to cluster the width and height of the normalized bounding boxes, assigning each bounding box to the nearest cluster center (i.e., anchor box). Calculate the distance between each bounding box and each cluster center, using Euclidean distance or other distance metrics; assign each bounding box to the nearest cluster center to form a cluster.

[0055] (6) Calculate the new center point (average of width and height) for each cluster center (anchor box). For each cluster, calculate the average width and height of all labeled boxes as the new cluster center.

[0056] (7) Repeat steps (5) and (6) until the cluster centers no longer change or the maximum number of iterations is reached. In each iteration, recalculate the distance between each bounding box and the new cluster center, and update the bounding box assignments.

[0057] (8) By denormalizing the width and height of the cluster centers to the original image size, the anchor frame parameters can be determined. Multiply the normalized cluster centers by the width and height of the original image to obtain the actual values ​​of the anchor frame parameters.

[0058] S222, Set the pixels, batch size, weight decay rate, initial learning rate, and number of model iterations of the input image for the convolutional neural network.

[0059] S23. Iteratively train the improved YOLOv5-s network model using preprocessed partial color images; specifically, place the improved YOLOv5-s network model with the parameters set into a computer with the configured environment, train the model using the training set, test the model's mAP value using the validation set for a preset number of iterations, and save one model. For example... Figure 3 As shown, for example, each iteration uses a validation set to test the model's mAP value and saves the model's weights to a specified path; the mAP value is displayed as... Figure 3 The broken line at the top.

[0060] During iterative training, before each iteration begins, the deep learning framework not only reads images from the training set but also generates new images through mixup data augmentation. The newly generated images and the read images are combined into training samples and input into the improved YOLOv5-s network model for training. After each round of training, the recognition accuracy of the model is tested using a validation set based on the model weights obtained from that round of training.

[0061] Mixup data enhancement specifically includes:

[0062] Two images (image A and image B) and their corresponding bounding boxes are randomly selected from the training dataset. A mixing parameter λ is randomly selected, which follows a Beta distribution. Each pixel in image A and image B, along with the bounding box information, is mixed according to formulas (1) and (2). The mixed image is used as the training sample, and the mixed bounding box information is used for object detection training.

[0063] Repeat the above steps until the required number of training samples is reached or one round of training is completed. Through mixup data augmentation, YOLOv5 can increase the diversity of the dataset, making the model more robust and capable of generalizing to different inputs.

[0064] New_image=λ*Image_A+(1-)*Image_B(1)

[0065] New_bounding_box =

[0066] *_box_A+(1-)*Bounding_box_B(2)

[0067] S24. The improved YOLOv5-s network model with the highest recognition accuracy during iterative training is used as the final woodlice detection model.

[0068] S3. A rotating gimbal is constructed using two MG995 servo motors, and an image acquisition unit is built with a variable-focus camera. The psyllid detection model is deployed on an embedded Jetson nano device, which is then used to control the image acquisition unit. Specifically, such as... Figure 4 As shown, two MG995 servo motors and a suitable bracket structure are used. One MG995 servo motor is mounted on the bracket structure as a horizontal rotation motor. The output shaft of this servo motor is parallel to the horizontal rotation direction, enabling it to control horizontal rotation. A second MG995 servo motor is mounted on the output shaft of the horizontal rotation motor as a vertical rotation motor. The output shaft of this servo motor is perpendicular to the vertical rotation direction, enabling it to control vertical rotation. A zoomable USB camera (such as a GUCEE G05 18X camera) is mounted on the vertically rotating servo motor and connected to the embedded device Jetson nano.

[0069] To drive an MG995 servo motor using a PCA9685 servo controller, connect the red wire of the MG995 to the V+ pin of the PCA9685 servo controller, connect the brown wire to the GND pin of the PCA9685 servo controller, and connect the yellow wire (signal wire) to the PWM output pin of the PCA9685 servo controller.

[0070] Connect the PCA9685 servo controller to the Jeston nano device pins. According to the PCA9685 servo controller pinout, connect VCC and GND to the Jeston nano's 3.3V and GND pins respectively, and connect SDA and SCL to the corresponding I2C interfaces on the Jeston nano.

[0071] Install and initialize the PCA9685 and OpenCV libraries on the Jeston Nano. Use the PCA9685 library in Python to control the PCA9685 servo controller to control the servo motors. In the program, first initialize the PCA9685 servo controller, then set the frequency and duty cycle of each PWM output port to control the servo motor movement. Use the OpenCV library in Python to control the camera's image capture and focal length changes.

[0072] S4. Input the images acquired by the image acquisition device into the psyllid detection model for identification to determine whether citrus psyllids exist in the captured images and their location. Specifically, use a Jetson nano to control the camera and servo motors to acquire images of a designated citrus plant. Send the acquired images into the psyllid detection model for target detection. Based on the detection results, determine whether psyllids exist in the acquired images, and determine the specific location of the citrus psyllids based on the rotation angles of the two servo motors and the current focal length of the camera. For example... Figure 5 The image shown is a schematic diagram of the input and output images.

[0073] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0074] Example

[0075] All data in this embodiment were collected in natural citrus orchards, originating from Guangzhou City, Guangdong Province. A zoomable camera and a DSLR camera were primarily used to capture RGB color images at a focal length of 50mm-200mm from a distance of 300cm from citrus psyllids on several branches and leaves. Blurry images were removed, and 4824 images were ultimately retained.

[0076] The image annotation tool LabelImg was used to label the citrus psyllids in the images. The label information, such as the location and border size of the citrus psyllids, was stored in an XML file in VOC format. 3377 images were randomly selected from the dataset as the training set, 482 images were randomly selected from the remaining images as the validation set, and the remaining 965 images were used as the test set.

[0077] The images in the training set were augmented using four operations: brightness adjustment, contrast adjustment, noise addition, and rotation. The final training set contained 13,508 images. Furthermore, the size of citrus psyllid targets in the images was estimated, defining individuals of citrus psyllids occupying less than 32×32 pixel areas as small targets.

[0078] The k-means++ algorithm was used to cluster the target boxes in the training set. The k value of the clustering algorithm was set to 9 based on the number of anchor boxes output by the model. The final clustering results are: (5, 8), (13, 17), (15, 20), (25, 26), (35, 49), (40, 56).

[0079] The hardware platform used is 16GB of RAM, an Nvidia GeForce RTX 2080Ti GPU, an Intel Core i7-10700@2.90GHz processor, and an Ubuntu 18.04 operating system. The method of this embodiment is implemented using Python as the programming language on the Pytroch deep learning framework.

[0080] The initial learning rate was set to 0.00341, the learning rate decay factor was 0.1, the model was saved once every 10 iterations, and the maximum number of iterations was 400. The anchor box parameters of the improved YOLOv5-s network were: (5, 8), (13, 17), (15, 20), (25, 26), (35, 49), (40, 56).

[0081] During model training, mixup data augmentation is used to enhance the dataset and generate new training images. In the YOLOv5 training model, the dynamic convolution method ODConv is used instead of the traditional convolution method, a SOCA attention mechanism for small object detection is added, and Wise-IoU is used instead of C-IoU loss function. The total loss value of the model is observed, and training is stopped when the total loss converges or the maximum number of iterations is reached, resulting in the improved YOLOv5-s network model.

[0082] The improved model achieved an average accuracy of 86.16% in detecting citrus psyllids, a 2.14% improvement compared to the basic training model, and a 4.29% improvement in the average accuracy for detecting small targets. The improved model enhanced detection accuracy while maintaining roughly the same detection speed, and also largely overcame the difficulty of detecting small targets.

[0083] The method provided in this invention improves the YOLOv5-s model by modifying the network structure and methods that were originally unfavorable for small object detection. For example, it uses the dynamic convolution method ODConv instead of the traditional convolution method, adds the SOCA attention mechanism for small object detection, and uses Wise-IoU instead of the C-IoU loss function, all of which effectively improve the model's detection accuracy for small objects. Furthermore, during training, mixup data augmentation is used to reduce the model's dependence on all target features. Ultimately, while maintaining the advantages of the original model's streamlined structure and fast detection speed, the method improves the model's target detection accuracy, enabling the identification of citrus psyllids on trees at considerable distances in complex field environments.

[0084] In another embodiment, a deep learning-based intelligent detection device for citrus psyllids is also provided. The device employs the intelligent detection method for citrus psyllids described in the above embodiments, and includes:

[0085] The image acquisition module is used to acquire images of citrus fruits in a natural orchard setting where citrus psyllids are to be detected. It includes an image acquisition device and a rotatable gimbal.

[0086] Image acquisition equipment, specifically a zoom camera or a DSLR camera;

[0087] The rotatable gimbal consists of two MG995 servo motors and is equipped with image acquisition equipment;

[0088] The embedded edge computing device, specifically the Jetson nano, is equipped with a trained psyllid detection model, which controls the image acquisition module to capture images of citrus fruits in a natural orchard setting.

[0089] The psyllid localization module is used to determine whether citrus psyllids exist in the captured image and output their location based on the detection results of the psyllid detection model.

[0090] It should also be noted that, in this specification, terms such as "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0091] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A deep learning-based method for intelligent detection of citrus psyllids, characterized in that, The method comprises the following steps: S1, collecting color images of citrus psylla in a natural orchard scene, preprocessing the collected images, and making a citrus psylla dataset; S2, improving the YOLOv5-s network model, using the citrus psylla dataset to iteratively train the improved YOLOv5-s network model, obtaining the optimal improved YOLOv5-s network model and using it as a psyllid detection model, specifically including: S21, improving the YOLOv5-s network model, using dynamic convolution instead of traditional convolution, and adding a second-order channel attention mechanism for small target detection; using Wise-IoU instead of C-IoU loss function to obtain the improved YOLOv5-s network model; specifically: S211, using channel-space full-dimensional dynamic convolution to replace the traditional convolution method of the original YOLOV5-s network model, and when performing convolution operation, fusing four dimensions in the kernel space composed of n convolution kernels participating in operation, performing parallel linear weighting on n convolution kernels, avoiding excessive smoothing of features, and improving the local detail expression ability of features; S212, introducing a second-order channel attention mechanism into the original YOLOV5-s network model to enhance the interaction ability of information between different channels in the convolutional neural network, thereby improving the detection performance of the network model on small targets; S213, using Wise-IoU to replace C-IoU of the original YOLOV5-s network model, introducing a dynamic focusing mechanism, calculating a dynamic focusing coefficient according to the distance between the two boundary boxes and the difference between the height-width ratio, and applying the dynamic focusing coefficient to the regression loss of the coordinates and height-width information of the boundary box, thereby paying more attention to the regression of different parts of the boundary box; S22, setting network model parameters for the improved YOLOv5-s network model; S23, iteratively training the improved YOLOv5-s network model using part of the preprocessed color images, and after each round of iterative training, using the validation set to verify the recognition accuracy; S24, using the improved YOLOv5-s network model with the highest recognition accuracy during iterative training as the final psyllid detection model; S3, using two servo steering engines to form a rotatable gimbal and a variable focal length camera to form an image collector; deploying the psyllid detection model to an embedded device and using the device to control the image collector; S4, inputting the images obtained by the image collector into the psyllid detection model for identification to determine whether citrus psylla exists in the captured image and its position.

2. The deep learning-based method for intelligent detection of citrus psyllids according to claim 1, characterized in that, Step S1 specifically includes: S11, fixing the shooting device by a tripod to shoot citrus plant images, simulating the image collection of the image collector in the actual application scene; the color images include citrus psylla images with different pixel sizes shot at different focal lengths; the shooting device uses a variable focal length camera or a single-lens reflex camera; S12, preprocessing the collected color images, and dividing the psylla dataset into a training set, a validation set, and a test set according to a preset ratio. 3.The deep learning-based method for intelligent detection of citrus psyllids according to claim 2, characterized in that, Step S12 specifically includes: S121, the citrus psylla in the collected color image is labeled by using a labeling software, including the position of the psylla and a minimum rectangular frame containing the psylla; S122, the color image after labeling is divided into a training set, a verification set and a test set according to a ratio of 7:1:2; S123, the color image in the training set is subjected to four kinds of operations of brightness adjustment, contrast adjustment, noise addition and rotation to expand the data.

4. The deep learning-based method for intelligent detection of citrus psyllids according to claim 1, characterized in that, Step S22 is specifically: S221, the size of the target frame in the training set is clustered by using a k-means++ algorithm; the k value of the clustering algorithm is set according to the number of target frames output by the model, and the result of clustering is taken as the target frame parameter value of the improved YOLOv5-s network model; S222, the pixel, batch size, weight decay rate, initial learning rate and model iteration number of the input image of the convolutional neural network are set.

5. The deep learning-based method for intelligent detection of citrus psyllids according to claim 4, characterized in that, Step S221 is specifically: S2211, one sample is randomly selected from the training set as the first initial clustering center; S2212, the shortest distance between each sample and the existing initial clustering center is calculated; S2213, the probability of each sample being selected as the next initial clustering center is calculated, and the next initial clustering center is selected according to the roulette method; S2214, steps S2212 and S2213 are repeated until k initial clustering centers are selected; S2215, the distance between each sample and the k clustering centers is calculated, and each sample is divided into the clustering center with the smallest distance to form k clusters; S2216, the clustering center is updated to the mean value of all samples belonging to the cluster; S2217, steps S2215 and S2216 are repeated until the clustering center no longer changes, and the result of clustering is obtained as the target frame parameter value of the improved YOLOv5-s network model.

6. The deep learning-based method for intelligent detection of citrus psyllids according to claim 1, characterized in that, Step S23 further includes: In the iterative training process, before each iteration starts, the deep learning framework reads images from the training set, generates new images through mixup data enhancement; The newly generated images and the read images are combined into training samples and input to the improved YOLOv5-s network model for training; After each round of training is completed, the model weight obtained based on the round of training is used to verify the recognition accuracy of the model based on the verification set.

7. The deep learning-based method for intelligent detection of citrus psyllids according to claim 1, characterized in that, In step S3, two MG995 servo steering engines are used to form a rotatable gimbal, and the embedded device is specifically a Jetson nano.

8. A deep learning-based intelligent detection device for citrus psylla, characterized in that, The detection method and device of any one of claims 1-7 comprise: An image acquisition module for acquiring citrus images in a natural orchard scene to be detected for the presence or absence of citrus psylla, including an image acquisition device and a rotatable gimbal; The image acquisition device is specifically a zoom camera or a single-lens reflex camera; The rotatable gimbal is composed of two servo steering engines and carries the image acquisition device; An embedded edge computing device that has a trained psylla detection model deployed therein, for controlling the image acquisition module to acquire citrus images in a natural orchard scene; A psylla positioning module for determining whether the captured image contains citrus psylla based on the detection result of the psylla detection model and outputting the position of the psylla.

Images