Identifying Wood Infected By Pine Wilt Disease Based On Deep Learning-Method

Pine wilt disease(PWD)is highly harmful and spreads rapidly.Timely detection of pine wilt disease victims can lay the foundation for taking corresponding treatment measures.The continuous development of deep learning technology and computer vision technology provides a more effective solution for the identification and monitoring of pine wilt disease.How to identify the infected wood quickly and efficiently in large-scale pine forests is of great significance for the prevention and control of pine wilt disease and the protection of forest resources.In this thesis,the RGB image data of pine wilt disease infected wood was collected by UAV equipped with consumer-grade digital camera.After image preprocessing,classification,labeling,data enhancement and other operations,the sample data set of pine wilt disease infected wood was constructed.The four models of YOLOv5 s improved by the attention mechanism and the original YOLOv5 s model are trained,verified,and tested respectively.The results show that the accuracy and speed of the improved model recognition are improved,which can meet the accuracy requirements of the identification of the infected wood in the actual epidemic prevention,and the identification and detection system of the infected wood of pine wilt disease is constructed according to the model.The main work and results of this thesis include:(1)Image acquisition and dataset construction of wood infected by pine wilt disease.Due to the long period of satellite remote sensing and the fast propagation speed of pine wilt disease,the timeliness of remote sensing inversion of pine wilt disease is poor.The UAV RGB image acquisition has the characteristics of short acquisition cycle and easy access to the target area image,which can obtain a large area of pine forest image in a short time.In this thesis,the RGB image data of the pine forest area in the study area are collected by the UAV equipped with lens.After image preprocessing operations such as distortion correction,image enhancement and image matching,the collected images are enhanced by brightness,contrast adjustment and rotation cutting.In the data set labeling stage,Labeling is used to label the data set.It lays a foundation for improving the performance of the infected wood identification model.(2)Comparison and selection of target detection algorithms.This thesis introduces and analyzes the structural framework,advantages,and disadvantages of the Faster R-CNN network model in the two-stage object detection algorithm and the SSD and YOLOv5 s network models in the single-stage object detection algorithm.VGG16 and Res Net50 are selected as the feature extraction network for Faster R-CNN,and VGG16 and Mobile Net V2 are selected as the feature extraction network for SSD to construct the network model and uses these three object detection algorithms for comparative experiments.The training and testing of the model are carried out through the pine wood dataset.The experimental results show that the YOLOv5 s network model has the best effect on the detection and recognition of pine wilt disease in these three different target detection algorithms.(3)Identification method based on improved YOLOv5 s.In this thesis,the YOLOv5 s model is improved and optimized by introducing an attention mechanism(SE,ECA,CBAM and CA)that can improve the feature extraction and selection of infected wood.Through training tests,the evaluation indexes such as recognition accuracy and recognition speed of different models are compared.The results show that the introduction of four different attention mechanism models has a certain degree of improvement in the identification of pine wilt disease.The YOLOv5s-CA model with direction-aware and position-sensitive CA attention mechanism module has the best comprehensive performance.Compared with the original YOLOv5 s model,the accuracy of the YOLOv5s-CA model is improved by 3.30 %.The detection speed FPS is increased by 20 frames per second,and the model weight is 14.4 MB.(4)Design and implementation of a system for identifying and detecting pine wilt disease.In this thesis,we have completed the design and development of a pine wilt disease infected wood recognition and detection system using the improved YOLOv5 s recognition model proposed in the previous thesis,including the design of the GUI human-computer interaction interface and the functional implementation.The system can select the input file type and adjust the Io U parameters,confidence parameters and latency parameters,and can display both the original input image(video)and the image(video)after recognition detection and save the detection results in a folder.The system can complete accurate identification detection of images and videos of pine wilt disease infected wood.