Research On Detection And Recognition Method Of Agricultural Light-trap Similar Pests Based On Deep Learning

Intelligent detection and accurate identification of agricultural pests is an important prerequisite for rapid and accurate forecasting of agricultural pests.It is of great significance for reducing crop losses and improving the green and sustainable development of agricultural economy.The agricultural lamp trapped by the intelligent insect forecasting lamp has a variety of species,diverse body postures,and pest scales falling off.The images of the lamp trapped pests have similarities and differences within species,which increase the difficulty of identifying the pests trapped by the agricultural lamp.This paper uses the deep learning target detection algorithm to realize the detection of agricultural lamp pests,and proposes a fine-grained image recognition method of agricultural lamp pests based on bilinear attention network to improve the recognition rate of agricultural light-trap similar pests.The research content and results are as follows:(1)Research on the detection algorithm of agricultural lamp trap pests based on deep learning.The two-stage Cascade R-CNN and the one-stage YOLO algorithm are used to detect the pest images of agricultural lights.Among them,Cascade R-CNN uses Res Net-50 and Res Net-101 as the feature extraction backbone network,and YOLO-V3 and YOLO-V4 use darknet53 and CSP-darknet53 as feature extraction networks,respectively.Taking 6 categories and 19 kinds of agricultural lights to attract similar pests as the research objects,comparative experiments were carried out under the same training set and test set.The test results show that the agricultural lamp trap detection model based on YOLO-V4 is more effective,with a recall rate of93.7% and a detection rate of 79.8%.(2)Research on the fine-grained image recognition model of similar pests induced by agricultural lights based on bilinear attention network.Aiming at the problem that YOLO-V4 has low accuracy in detecting agricultural lamp-trapped pests,this paper adopts a one-layer detection and two-layer classification cascade model to improve the recognition rate of agricultural lamp-trapped pests,and proposes a bilinear attention network based on The agricultural lamp attracts similar pest identification model(Bilinear-Attention Pest net,BAPest-net).The BAPest-net model adopts a bilinear network structure,modifies the down-sampling position of the feature extraction network,and adds an attention mechanism module.(3)Analysis and evaluation of detection and identification results of similar pests induced by agricultural lights.Comparing VGG19,Densenet,Resnet50,BCNN and BAPest-net’s light-trapped pest recognition models,the results show that BAPest-net has the best recognition effect,with an average recognition rate of 94.9%.The accuracy rate of the entire cascade system is 87.6%,which is 7.8% higher than the detection and recognition accuracy of the first layer.This paper proposes a detection and recognition model for similar pests induced by agricultural lights,which combines one-level detection and two-level classification.The bilinear attention network is used to reduce misjudgments due to similar posture,shape,and scale features.This method provides a theoretical basis for the further development of intelligent forecasting technology of insect situation in my country.