Identification Of Weed Community Characteristics In Field Images

In recent years,the use of machine vision to identify field weeds and variable spraying of herbicides has become the development direction of modern fine agriculture.However,traditional image processing techniques focus on the identification of weeds in single plants or small areas.For this problem,this paper proposes taking weed community as the main research object,three different methods for extracting weed community characteristics were proposed to solve the key problem of effectively extracting weed community characteristics.The main work includes the following parts:(1)Using the Non-Subsampled Shearlet Transform algorithm(NSST)to extract and identify the weed community.Firstly,the characteristics of the experimental sample images used are introduced.Then the principle of non-subsampled shearlet transform is briefly introduced.And the advantages of it in characterizing anisotropic information are analyzed.Meanwhile,the specific implementation steps of the non-subsampled shearlet transform algorithm in this chapter are introduced in detail,and the flow chart is given.At the same time for different grayscale images,different segmentation sizes,different scales and a detailed analysis of the effects of different recognition algorithms on the recognition rate is given.Finally,the conclusion that the HSV(H)gray image block size is 128×128 pixels is the best.The SVM classification recognition algorithm has the best recognition effect.The SVM algorithm is the most robust and its average recognition rate can be it reached 66.6% and the highest recognition rate can reach 71.7%.(2)The binary mask image in the ROI region is extracted based on the interaction method,and the weed community feature image is obtained by data aggregation and similarity matching.Different recognition algorithms are used to identify and classify the feature values in the feature image under different mask images.The average recognition rate of four algorithms,Decision Tree(Tree),Support Vector Machine(SVM),K Nearest Neighbor Algorithm(KNN),and Bagging algorithm,is between 87.0% and 89.4%.(3)Due to the poor real-time and portability of the manual extraction ROI region algorithm,it is necessary to manually extract some mask images in the ROI region as a template for data aggregation,which is subject to subjective factors.An algorithm for automatically acquiring ROI regions is proposed.This algorithm eliminates the manual extraction of mask images.It can extract weed community features faster,and the real-time performance of the algorithm is guaranteed.At the same time,the SVM classification and recognition algorithm is used to classify and identify the extracted feature values,and the recognition rate reaches 89.9%,which can meet the experimental expectation requirements of this topic.