Research On Identification Of Corn Seedlings And Weeds Based On Deep Learning

The treatment of weeds plays a vital role in the yield of corn,but the traditional spraying weeding is easy to cause environmental pollution,and precise spraying can solve the above problems.The accurate identification of corn and weeds is the basis of precise spraying.This paper studies the identification problems of corn seedlings and weeds,corn seedlings and weeds are taken as the research objects,The goal is to improve the recognition accuracy of corn and weeds,and reduce the size of model file,this study provides support for the precise spraying of weeds.The main research work of this paper is as follows:Firstly,aiming at the lack of database of corn seedlings and weeds at present.When corn is in three to five leaf stage,it is the best weeding period,and images are collected during this period,the images of corn and weeds at three to five leaf stage were collected by single shot,and more than 14000 images are obtained through preprocessing.The operation of data augmentation is used to increase the diversity of images.The processed images are marked in yolo format by labelimg software,the image data set of corn and weeds is obtained,which has more than17000 images.Secondly,the algorithm model of corn and weed based on yolov5 was studied in this chapter,and analyze the input terminal,backbone,neck and prediction of yolov5,and the feature map of the model is visually analyzed,it provides theoretical support for algorithm improvement.Through many fine-tuning experiments,the m AP@0.5value of yolov5 s model reaches 0.916,the size of the exported file of the model is13.7mb,and the AP values of corn and weed are 0.953 and 0.879 respectively.It shows that the modle can correctly identify corn and weeds,but the recognition performance of the model for weeds should be improved in the future,and the size of the exported model file should be reduced while ensuring the recognition performance.Thirdly,aiming at the problems that the yolov5 s model has a low recognition accuracy for weeds and a large output model,this chapter proposes an improved algorithm of yolov5s;The modle of SENet is embedded in yolov5 s,and the Focus structure of yolov5 s is replaced with Conv structure to facilitate the extraction of key features from the model;The C3 structure is fused with ghost module and the C3 module in yolov5 s is replaced with the fused structure,so that reduce the output file size of the model while ensuring the performance of the model;After many fine-tuning experiments,the value of m AP@0.5 for the improved model reaches0.930,the overall recognition performance of the modle is improved by 0.014 points,the AP value of weeds increased by 0.008 points,the AP value of corn increased by0.02 points,compared with the original model,the size of the output model is reduced by 2.9mb.The experiment shows that the overall recognition performance of the new model is improved and the file size of the exported model is reduced.Fourthly,this chapter explores the effect of the new model,and verify the robustness of the model by designing experiments of brightness,occlusion and noise;The new model is tested by using the test set,the value of the m AP@0.5 for the improved model reaches 0.934,the AP value of corn and weed is 0.992 and 0.875 respectively,the time of detecting each image reaches 3.1ms;the feasibility of precision spraying system was verified by building an indoor precision spraying platform;The precision spraying experiment of weeds in field was carried out,the identification accuracy for corn and weeds was 92% and 83%,respectively,and the probability of identifying weeds and spraying accurately reaches 81%,which verifies the recognition performance of the model for corn and weeds,and it shows that it is possible to carry out the operation of the precision spraying and weeding during the three to five leaf stage of corn.