Research On Lawn Weed Recognition Method Based On Machine Vision

As an essential element of green space,lawn has a major impact on urban landscapes and ecological civilisation.Weeds compete with lawn for growing space and breed pests and diseases,which is one of the most important factors causing lawn degradation.Herbicides have become the main way of lawn weed management with the advantages of economy,efficiency and labor saving.However,the widespread use of herbicides not only exacerbates the problem of weed resistance,but also poses a threat to the health of people when they are in public areas,such as parks and sports grounds.The lawn weed identification system developed in this paper based on machine vision can help plant protection robots achieve precise application of herbicides,so as to greatly reduce the environmental damage of herbicides and prolong the formation cycle of the weeds’ resistance to herbicide.The main research contents are as follows.(1)Establishment of lawn weed identification system.Through the modular analysis of the functions of the lawn weed recognition system,the mechanical structure design of the weed recognition system and the hardware selection of the image acquisition system were completed.Combined with the design of lawn weed control car and target recognition method,the hardware module workflow and recognition algorithm flow design of lawn weed recognition system were completed.(2)Research on image preprocessing algorithm of lawn weeds.Firstly,the vegetation area in the image was extracted by using the extra-green segmentation operator,and the presence of weeds in the image was judged by the number of great peaks in the histogram,so as to improve the overall recognition efficiency of the recognition system.Two pre-processing methods,a pre-processing algorithm based on local variance and a multi-channel color fusion pre-processing algorithm using a non-linear fusion strategy,were proposed to expand the feature differences between weeds and lawns.(3)Research on lawn weed image segmentation algorithm.An improved Retinex enhancement algorithm incorporating the location relationship of the image target pixels was proposed.To obtain the spatial information of the required part of the pixels more accurately,the pixels were divided into three categories: foreground,background and pixels to be subdivided using multi-threshold segmentation and open-operator differencing.The spatial information of the pixels to be subdivided is extracted using local density method.The sigmoid function was used to incorporate the local density information and optimize the gray transformation coefficient of the reflection component to obtain the enhanced image,to realize the purpose of directly segmenting weeds by OTSU algorithm.A fuzzy C-mean segmentation algorithm that considers the spatial information of the gray distribution of the image was proposed.The region area is used to constrain the filtering range,smooth the gray maximum in the preprocessed image,and reduce the gray level loss of the target region caused by median filtering.The gray distribution difference characteristics in different directions around pixels were introduced to realize lawn weed segmentation in the clustering process.Both improved algorithms can effectively segment the lawn weeds as the detection target in this paper.Based on the results of weed segmentation and camera calibration,the precise application area of lawn weeds was determined.(4)Research on lawn weed classification algorithm.To realize the species recognition of lawn weeds,the problem of weed species identification was transformed into the problem of weed sample classification.For the case of a small number of weed samples,the extracted high-dimensional features of lawn weeds were reduced by principal component analysis,and a one versus rest support vector machine was constructed to complete the model training of the dimensionnally reduced features.For the case of high number of weed samples,a PPLCNet lightweight neural network was built to classify the weed samples,which improved the accuracy of weed recognition while taking into account the real-time performance of the classification algorithm.