Research And Realization Of Winged Insects' Statistics And Recognition System Based On Machine Vision And Yellow Board Induction

Rapid and accurate statistics and identification of pests is of great importance for crop pest control.Traditionally,pests' detection mainly relies on plant protection experts through naked eye for manual statistics and identification,which is subjective,ambiguous,time-consuming and laborious.With the development of computer technology,people gradually applied image processing technology and pattern recognition technology to the research and identification of pests,and established the identification system to improve the recognition accuracy and efficiency.In this paper,typical winged insects collected from field and greenhouses were used as objects to study.We constructed the image acquisition system with yellow board and raspberry camera,and used raspberry as the development platform to complete the construction of statistics and identification system.Firstly,in order to distinguish targets between impurity such as excrement,water droplets and mud points on the collected images,method of detecting and counting of winged insects based on connected region markers and YOLO depth learning network were studied.Then,traditional pattern recognition technology and depth learning recognition technology were compared.In the study of traditional pattern recognition methods,global features such as color features,geometrical features and texture features and local HOG features were extracted.According to the extracted features,SVM and BP neural network were used to build classification model and classify the specified species.Meanwhile,classification and recognition rate of the classifier based on multivariate feature combination was discussed.In the study of depth learning classification method,the recognition technology based on YOLO depth learning network was analyzed.Finally,this paper completed the build of raspberry development environment based on Python,OpenCV and Darknet to complete the proposed system.The performance of the system was tested with the pictures collected at crop planting base and strawberry shed,and the artificially identified species and quantity were used as the standard.The experimental results show that,compared with the connected area mark method,YOLO depth learning network owns higher count accuracy and stronger anti-interference ability.For classification results,due to the number of samples used for training is not enough,YOLO's classification accuracy is not high.Classification accuracy of SVM classifier is better than BP neural network.Singular features such as color,texture,shape and local features cannot distinguish the pests very well.The combination of global feature with local feature has a certain improvement in the classification accuracy,but the calculation speed is obviously slow.Therefore,counting method based on YOLO depth learning network and SVM classifier based on global feature is chosen for statistics and recognition system in this paper.The counting average correct rate of test picture is 93.45%,the average accuracy of classification is 92.24%,which basically meets the design requirements of this paper.The results of this paper provide a theoretical and practical basis for pest statistics identification based on machine vision.