Research And Application Of Rice Disease Identification Based On Improved YOLOV5

During the growth and development of rice,there are often diseases,which will cause rice death and affect rice yield if not controlled in time.Identification of rice leaves can reduce the impact of disease on rice yield.However,the traditional artificial control method is not only heavy workload,low efficiency,and accidental,can not achieve real-time monitoring,seriously limited the development of modern agriculture.Therefore,using target detection algorithm to identify rice diseases is an important research direction in the agricultural field.Traditional machine learning algorithms based on artificial feature extraction,such as decision tree,random forest and support vector machine,need a large number of original images for training,and the recognition rate of disease images is not high,so it is necessary to study an effective method to improve the recognition effect of rice leaf disease.In this paper,rice disease images obtained from the experimental rice field in Jinzhai County,Lu ’an City,Anhui Province were used as the experimental data set,and a rice disease recognition method based on improved YOLOV5 was proposed to realize the prediction and application of rice disease.The main research work in this paper is summarized as follows:(1)Acquisition and pretreatment of rice disease image data set.Rice disease images were taken in the rice experimental base of Jinzhai County,Lu ’an City,Anhui Province,including rice leaf blast,rice bacterial blight and rice hemp leaf spot,with a total of 3234 images,each with 2400*1080 pixels.Images were preprocessed to delete those out of focus and lacking of disease features,and the images were normalized.The normalized images were input into the YOLOV5 model to enhance the rice disease data set based on the method of hyperparameter evolution.(2)Construction of rice disease recognition model based on YOLOV5.First,the YOLOV5 model was compared with Faster RCNN,YOLOv3,YOLOv4 and YOLOR models,and the experimental results showed that YOLOV5 model had the best comprehensive performance.Then,Shuffle Net V2 lightweight network YOLOV5-V2 for short was added on the basis of YOLOV5 model.Delete the Focus layer in YOLOV5 and perform channel clipping on the YOLOV5 header.Compared with YOLOV5-V2 model,the image prediction speed of YOLOV5-V2 model is increased by 3 times,and the weight file size is reduced to 3.2MB,which hardly reduces the accuracy of rice and reduces the calculation amount of backbone network.Finally,PP Pico Det network is transplanted based on YOLOV5,yolo V5-P for short.The experimental results show that the calculation amount and weight file size of Yolo V5-P are reduced,and the weight file is reduced to only2.2MB.(3)Design and implementation of rice disease target recognition system.Based on the real-time recognition of the camera on the mobile terminal,the real-time detection on the mobile terminal is realized under the same server,and the recognition results are saved on the host with a delay of about 5 seconds.Based on computer end rice disease recognition system using Py Qt5 design page,add camera recognition,image recognition,video recognition three modules.Traditional network models pursue higher model recognition effect,which often takes up a large amount of calculation,leading to better model evaluation indicators but not suitable for industrial application.This paper solves the problems of slow recognition speed,frame drop or high delay caused by the large weight files and computation of the model in mobile deployment.