Research And Implementation Of Sex Pheromone-baited Intelligent Forecasting System For Cnaphalocrocis Medinalis Based On Images

Crops in our country are deeply harmed by agricultural pests,causing immeasurable economic losses to the country.Therefore,real-time monitoring of pest population dynamics as well as effective prevention and timely control are the prerequisite to ensure food security.Sex pheromone-baited forecast is an important method widely used in agricultural pest monitoring,with strong sensitivity and specificity.At present,sex pheromone-baited traps based on machine vision can free forecasters from the task of identifying and counting pests and have become a hot spot in the research on intelligent pest monitoring.However,there are some problems in the image of the sticky pest board photographed by the sex pheromone-baited trap,such as complex image background,pest adhesion,and interference of similar non-target pests.As a result,this method cannot obtain satisfactory detection results.It is urgent to further improve the accuracy of pest detection and counting.In this paper,taking the rice pest Cnaphalocrocis medinalis(Guenee)as the research object,the detection model of Cnaphalocrocis medinalis based on deep learning is studied,and a sex pheromone-baited intelligent forecasting system for Cnaphalocrocis medinalis based on images is developed.The real-time,accurate and intelligent forecast of the Cnaphalocrocis medinalis is realized,and the historical data can be traced back.The main research contents and results are as follows:(1)An intelligent sex pheromone-baited trap based on machine vision is designed and implemented.To solve the problems in traditional sex pheromone-baited traps such as non-traceability of data,an intelligent sex pheromone-baited trap based on machine vision is designed and built,including a machine vision module and an image transmission module.Among them,the machine vision module is responsible for pest trapping and regular image shooting;the image transmission module is responsible for camera SDK control and remote image uploading.In this paper,a total of 712 images of sex pheromone-baited sticky pest board containing Cnaphalocrocis medinalis are collected by intelligent sex pheromone-baited trap to produce the image data set of Cnaphalocrocis medinalis.(2)Established and compared two deep learning-based detection models of the Cnaphalocrocis medinalis.To improve the accuracy of the detection and counting of the Cnaphalocrocis medinalis in the sticky insect board image,two deep learning models,Faster R-CNN and YOLOv3,were trained and tested based on the Cnaphalocrocis medinalis image collected by the intelligent sex pheromone-baited trap.The test results show that on the same test set,the Faster R-CNN and YOLOv3 detection models achieve91.9% and 91.1% precision,94.2% and 92.4% recall,respectively.Although the two detection effects are similar,YOLOv3 has faster target detection speed,stronger portability,and is more suitable for practical applications and embedded system transplantation.Therefore,the YOLOv3 model is selected as the sex pheromone-baited trapped Cnaphalocrocis medinalis detection model.(3)A detection model of Cnaphalocrocis medinalis based on YOLO-DB was established.Since there are some false detection and missed detections in the detection results of YOLOv3,this paper improves on the basis of YOLOv3.By adding the Drop Block regularization method,the false detection caused by the over-fitting of the model is reduced,and the DIo U-NMS non-maximum suppression is used to reduce missed detection caused by target adhesion.The improved YOLOv3 model achieved 94.0%precision and 94.2% recall for Cnaphalocrocis medinalis,but the artificial synthetic sex pheromone baited of Cnaphalocrocis medinalis has trapped non-target similar pests,which lead to interspecies misdetection of the improved YOLOv3.Aiming at this situation,a double-layer detection model YOLO-DB of Cnaphalocrocis medinalis is proposed.The first layer is the improved YOLOv3 model,and the second layer is the deep bilinear transformation classification network DBTNet-101.The output terminal of the first layer is cascaded with the input terminal of the second layer,and the false detection results of similar pests in the first layer are corrected through the second layer network to improve the accuracy of the model.The results showed that YOLO-DB’s detection precision and recall of the Cnaphalocrocis medinalis in the image reached 97.6%and 98.6%,respectively,detection FPS reached 2.0 frames per second.(4)Realize and test a sex pheromone-baited intelligent forecasting system for Cnaphalocrocis medinalis based on images.The system includes an intelligent sex pheromone-baited trap,a server and a Web client.Use Django framework to develop server program;use My SQL database to access information;use Vue framework to develop Web client;use HTTP request to connect intelligent sex pheromone-baited trap,server and Web client,when intelligent sex pheromone-baited trap and Web client send an image upload request,the server receives the request and calls the detection model,saves the detection and counting results to the database and feeds back,realizing the connection between the intelligent sex pheromone-baited trap,the Web client and the server.In order to verify the reliability of the system,the system is deployed and tested on the cloud server.The results showed that the functions,browser compatibility,system performance and other indicators of the sex pheromone-baited intelligent forecasting system for Cnaphalocrocis medinalis can all meet the expectations.The average FPS of the online detection of a single image is 0.205 frames per second,and the recognition speed and accuracy can meet the needs of intelligent forecasting of agricultural sex pheromone-baited trapped insects,providing ideas and reference for timely and accurate forecasting of sex pheromone-baited trap.