Study On Identification Of The Field Weed Based On Computer Vision

The breeding of farmland associated weeds would decline the crop yields,using traditional chemical herbicides could control weeds effectively,but there were drawbacks such as water and soil pollution,pesticide residues,and personnel poisoning.In this paper,combined with optical image processing technology,in-depth analysis of crop information monitoring,researched a field weed identification method which based on computer vision technology,identified the weed and crop identification through BP neural network model.The main work of this paper included the following aspects:(1)In this experiment,three kinds of weeds were selected as the object of image collection and processing: digitaria sanguinalis,dogtooth and amaranth.The corn seedling and weeds were taken for 50 images and stored in groups.The open source computer vision library(OpenCV)was used to preprocess the collected images.The preprocessing algorithm included color space conversion,image filtering,threshold segmentation and morphological processing.(2)After pretreatment experiments,the contour information of weeds was obtained,and extracted the contour features of weeds.The extracted features included dimensionless parameter aspect ratio r,roundness R,and the first invariant moment feature S.(3)This paper used BP neural network as the weed recognition model.Through orthogonal experiment on BP neural network,the optimal network structure was determined: the learning rate was 0.5,the hidden layer node was 6,the input node was 3,and the output node was 1.The weed contour characteristic data was used as the input vector of BP neural network.The BP neural network was trained with 160 groups of sample data,and the network was simulated by Matlab.Updating the weight and bigotry,the correct rate of weed recognition was 95.42%,using the remaining 40 groups of sample data as test set to verify the BP neural network,the correct rate of weed identification was 93.33%.