Research On Crop Pest Detection Method Based On Improved Convolutional Neural Network

In recent years,my country's economic investment in agricultural production has been increasing,of which grain production accounts for the largest proportion of the total input,but the output of grain has not been significantly improved,and one of the important reasons affecting grain production is crop pests.With the rapid development of agricultural information technology,more intelligent treatment is needed in the management of crop pests,instead of large-scale spraying of pesticides.The pesticide dosage should be strictly controlled to achieve precise spraying to control crop pests and ensure the efficiency of pesticide spraying.And comprehensively realize precision and intelligent agriculture.This article takes 6 major crop pests photographed against a natural background as the research objects.The main contents are as follows:Firstly,it analyzes the characteristics of the major methods in the target detection field before 2012,from the manual feature period to the artificial neural network period,and then from the artificial neural network to the convolutional neural network.It mainly analyzes the three most popular convolutional neural network models in the current target detection field.One is the RCNN series algorithms based on candidate regions and deep learning classification,and the other is the SSD and YOLO series algorithms based on deep learning regression.After analyzing multiple algorithm models,this article will use YOLOv3 as the basic model.Secondly,based on the image data of 6 kinds of crop pests,and then grab some related pest image data through web crawler technology,use the two parts of crop pest image data as basic data,and construct the crop pest data set needed in this research CPXJ-Datasets,and preprocess it.In the preprocessing process,a non-overall labeling method is proposed,which is compared with the traditional overall marking method,and the effective pixel proportion(EPP)is proposed to measure the number of non-integral marking methods.This situation.After preprocessing,multiple training data sets required by convolutional neural networks are obtained,which prepares data for subsequent experiments.Finally,it is improved on the basis of YOLO-v3,and uses a more advanced clustering algorithm ISODATA to obtain anchor boxes,which provides better initial values ??for the prediction of bounding boxes.Analyzed the feature extraction network Darknet53,and added two Dense Block structures at the end of its structure,increasing the information from the original 3 scales to 5 scales,increasing the amount of information of the features,and also improving the characteristics of the target detection The representational power.In order to adapt to the image input of different scales,the SPP network structure is added before the first fully connected input in the YOLO-v3 network structure to increase the multi-scale adaptability of the model to the input image.In order to make the model more robust,this study uses the ELU activation function to replace the Leaky Re LU activationfunction in the original model.By conducting convolutional neural network experiments on six crop pests,the improved convolutional neural network model of this research and the current mainstream three target detection algorithm models are compared and analyzed for crop pest image detection effects,and then based on the data set constructed in this paper CPXJ-Datasets compares and analyzes the experimental results of the non-holistic marking method and the traditional marking method,and proves the reliability and validity of the non-holistic marking method and the improved convolutional neural network model proposed in this paper.