Image Recognition Of Rape Pests Based On Deep Learning

Rapeseed has a large area of planting in China and plays an indispensable role in the development of national economy.However,frequent insect threats affect the yield of rapeseed.In order to improve the yield of rapeseed and ensure its excellent quality,and promote the development of national economy,it is of great significance to carry out prevention and control measures for rapeseed-related pests.With the rapid development of deep learning technology,image-based identification has become a hot topic in the field of pest identification.In order to meet the actual needs of rape pest detection and identification,this paper selects seven common rape pests such as Cabbage butterflies and Aphids as the research objects,and conducts an improved comparative experiment on the Faster-RCNN and YOLOv5 target detection framework,so as to achieve the purpose of identifying rape pest images,and constructs a Web-based rape pest identification system.The research contents of this paper are as follows:(1)The laboratory specimen data,the public Image Net data set and the Kaggle reptile data were collected,cleaned and expanded.The images of training,verification and testing were stored in the folder according to the ratio of 8:1:1 and marked.The rape pest data set was constructed and completed,including seven types of pests,with a total of 5217 images.(2)The algorithm was improved based on the two-stage model Faster-RCNN,and the Mpest-RCNN rape pest target detection model was constructed.The Res Net101 feature extractor was used to replace VGG16,and a new convolution network structure with small anchor points was proposed to extract features.The m AP value was 92.67 %,which was5.8 % higher than that of Faster-RCNN,and the detection speed was 67.8 ms higher.Therefore,the recognition accuracy of rape pests was improved,and the recognition of small pests was better adapted.(3)The algorithm was improved on the basis of single-stage YOLOv5,and the SEYOLOv5 rape pest target detection model was constructed.The SE module was used to optimize the YOLOv5 m network model.The m AP value was 96.5 %,which was 9.63 %,3.86 % and 5.55 % higher than that of Faster R-CNN model,Mpest-RCNN model and YOLOv5 model,respectively,and the recognition accuracy of the target detection model was improved.In addition,the prediction time of SE-YOLOv5 model is 69.52 milliseconds,263.18 ms faster than Mpest-RCNN,thus achieving the goal of efficient and accurate identification of rape pests.(4)The experimental comparison of Mpest-RCNN and SE-YOLOv5 algorithm verifies that SE-YOLOv5 algorithm has faster detection speed and higher recognition accuracy in the recognition process of rape pest image.Based on this,this paper designs and builds a rape pest recognition system,which realizes the functions of pest recognition,pest query and pest display,and promotes the prevention and control of rape-related pests.