Research And Implementation Of Insect Target Detection System Based On Improved Faster R-CNN Model

At present,biological research,pest detection and other fields with the help of insect target detection technology,ushered in a new round of rapid development.Traditional insect target detection methods need to go through complex image preprocessing,and also need to design a specific feature extraction algorithm according to the specific task requirements.The model recognition accuracy trained by traditional methods is low and the generalization ability is weak.In this context,this paper uses the convolutional neural network in deep learning technology to extract image information,and conducts targeted optimization on it,so as to train a model with excellent detection and generalization performance.In this paper,an insect image data set was made according to the project requirements of the research group,and the performance of the optimized target detection model Faster R-CNN was verified through comparative experiments.Finally,an insect target detection system based on improved Faster R-CNN was designed on this basis.The main work of this paper is as follows:(1)Collect,expand,label and make insect image data sets.According to the requirements of the research group,this paper collected and labeled five types of images of common agricultural insects with different postures.Due to the insufficient number of samples,the original data set was expanded through image spatial transformation and grayscale operation.The expanded data set has a total number of 6670 samples,which meets the requirements of deep learning for the number of data sets.(2)The improvement and verification experiment of insect target detection model.In this paper,the target detection algorithms that are widely used at present are analyzed theoretically,and Faster R-CNN is selected as the target detection framework of the system.In addition,Fasters-RCNN was optimized in three aspects.First,after comparing and analyzing the advantages and disadvantages of common network structures,Res Net-50 was used to replace the original VGGNet,so as to enhance the expression ability of the model while ensuring the reduction of the number of model parameters.Secondly,In the process of feature extraction from shallow features to deep features,the possibility of image data loss will increase.In order to avoid the occurrence of similar situations,we introduced a multi-scale feature fusion method.Finally,the Anchor mechanism in target detection is optimized: on the one hand,the Online Hard Sample Mining strategy is added in batch processing to balance the positive and negative insect samples;on the other hand,Soft Non-maximum Suppression is used to replace Nonmaximum Suppression in filtering candidate target boxes to avoid missed detection of candidate target boxes of target insects.(3)The software and hardware design and implementation of insect target detection system.This system is mainly composed of embedded system,Web server and target detection server.In the aspect of embedded system,Raspberry Pi is used as the development platform.Its main work is to collect images regularly as the client,and complete the task of uploading images by establishing the socket connection between Raspberry Pi and the server.In terms of server,the server part of the system can be divided into Web server and target detection server.Firstly,by integrating Spring Boot,My Batis,Thymeleaf,Echarts and other frameworks,the Web server is designed to realize the basic functions of the Web client.Secondly,the target detection server is based on Django framework and equipped with the pre-trained Faster R-CNN target detection model to realize the insect target detection function.The improved insect target detection system based on improved Fasters-RCNN has reached the system design standard,and has a wide range of application value.