Research On Recognition Of Agricultural Insect Images Based On Deep Learning

With the development of society and science,image recognition,as an important part of computer vision,has made great progress.Traditional target recognition task requires complicated image-preprocessing,so special methods have to be designed to extract target features.However,the accuracy and adaptability of these traditional methods are not ideal.In order to tackle the shortcomings of traditional methods in target recognition,this thesis proposed an improved convolutional neural network with powerful feature learning ability to implement insect image recognition.This thesis aimed to common agricultural insects,built an insect image dataset,designed an deep learning method to figure out insect recognition task.The improved target detection algorithms,Faster R-CNN and RefineDet,were proposed which performance were verified in agricultural insect identification by experiments,based on deep learning.The main work is described as below:(1)Since there is no public insect image dataset that is suitable for our research,common agricultural dataset with 24 species insect images was manually collected and labeled in real natural environment.In addition,the dataset was augmented by rotating images and adding image noise.As a result,this dataset contains a total of 10,560 images,which is appropriate for deep learning.(2)According to the research of insect recognition,target recognition algorithm and the training way of algorithm were analyzed and chose in accordance with current research topic in theoretical perspective.Faster R-CNN was selected as the main research framework.By studying specific object of insects,an improved network Faster R-CNN(VGG19)was proposed,which was further verified on insect image dataset.(3)An improved algorithm based on RefineDet network architecture was proposed.The flaws of Faster R-CNN(VGG19)were summarized and analyzed and two improvements were proposed on Faster R-CNN.First,the original dataset was further augmented into larger-scale one.Second.RefineDet algorithm was adopted as main framework for insect image recognition.This thesis also proposed new ARM network architecture which combined with newly constructing anchor frame.Experimental results showed that improved anchor frame scheme and new ARM network architecture can improve the performance of insect recognition.The final RefineDet(VGG19)model achieved a recognition rate of 92.8%.