Research And Implementation Of Intelligent Identification System Of Pests Trapped By Searchlight Based On Machine Vision

Rice migratory pests cause severe yield and economic losses to rice food every year.It is important that these pests are timely and accurately monitored for controlling them.Searchlight trap is specially developed to use light to trap migratory pests at high altitudes.The searchlight trap can help us to know when pests move in,how much they move in and other relevant information.Compared with insect radar,the searchlight trap has the advantages of large insect quantity,obvious pest fluctuation curve and low application cost.But the trapped pests need to be manually identified from a large number of non-target insects,which is inefficient and labor-intensive.At the same time,the data can not be traced and other shortcomings also lead to the difficulty of the searchlight trap to meet the needs of large-scale migratory pest monitoring.In order to replace manual identification of migratory pests,this paper design and add a machine vision module to the traditional searchlight trap.An automatic identification model of migratory pests based on machine vision is established,using three migratory pests:Cnaphalocrocis medinalis,Nilaparvata lugens,Sogatella furcifera,and non-migratory Chilo suppressalis.An automatic identification system of migratory pests is established and realizes automatic identification and real-time monitoring of searchlight trap trapping pests.The main research contents and results are as follows:(1)Design and establish the machine vision module of the searchlight trap.In order to solve the problem of low automation of searchlights trap,this paper designs a machine vision module with a multi-layer distributed pest structure on the basis of traditional lights.This module uses an Android industrial tablet as the control center,uses a multi-layer insect structure to disperse insects,and calls an industrial camera to take photos of the dispersed insects.At the same time,the searchlight is used to collect image data,and the Label Img tool is used to annotate and form the insect image dataset.(2)Four identification models based on deep learning are established and compared.In order to select an identification model suitable for searchlight trap trapping pests,four target identification model(Faster R-CNN,YOLOv3,YOLOv4 and YOLOv5)are trained and tested.The test results show that YOLOv4 has better comprehensive performance than the other three models.The YOLOv4 model ran at 0.95 frames per second,and the accuracy of the model for Chilo suppressalis,Cnaphalocrocis medinalis,Sogatella furcifera,and Nilaparvata lugens reached 77.73%,73.66%,66.72% and71.27%,respectively.YOLOv4 model shows the characteristics of fast identification speed,high identification accuracy and strong portability.Yolov4 is more suitable for application needs,and hence is selected as the basic identification model.(3)The identification model of searchlight trap trapping pests based on YOLO-LPNet is established.To improve the identification performance of model,this paper proposes a new model called YOLO-LPNet by improving the YOLOv4 model.Aiming at the problem that small target pests occupy a small proportion of area in an image and the feature extraction network can extract effective features poorly,the interleaved cross sampling method is introduced into the model to improve the missing identification problem.At the same time,the SENet is introduced to improve the mutual false detection between target pests and disturbing pest.The identification accuracy of YOLO-LPNet for Chilo suppressalis,Cnaphalocrocis medinalis,Sogatella furcifera,and Nilaparvata lugens is 85.76%,90.94%,74.80% and 71.67%,respectively,representing improvements of 8.03,17.28,8.08,and 0.40 percentage points.The corresponding recall rates were 90.8%,86.01%,73.28%,and 81.39%,respectively,representing improvements of 19.85,25.88,18.04,and 22.15 percentage points.(4)An intelligent pest identification system based on machine vision is implemented and tested.The system includes a machine vision based searchlight trap,a server and an Android client.Using Django framework to develop server-side programs.Access information using My SQL database.Develop an Android client using Android Studio.After the system development was completed,the system was deployed on a cloud server and tested to verify its reliability.The results show that the system meets the expectations with respect to the stability of the Android terminal,system performance,and other indicators.The system model for identifying an image has an FPS of 0.65 frames per second,which means the identification speed and accuracy are sufficient to meet the needs of intelligent identification and reporting of agricultural pests.