Study On Identification For Field Weed In Rape Seedling Stage By UAV Vision

Associated weeds in farmland will affect the growth and quality of crops.At present,weed control mainly relies on large-area spraying of chemical herbicides in China,which is simple and efficient but causes environmental pollution and does not conform to the development concept of green environmental protection.Therefore,designing a crop and weed identification system to guide variable spraying will help reduce the economic cost of weeding and protect the ecological environment.This topic explores the segmentation of plant images and the classification of rapeseed and weeds in the actual field environment.The feasibility study of field weed density detection and variable spraying is carried out.The main content on this dissertation are as follows:(1)The crop emergence period,the degree of influence of weeds on the crops and the quality of the subsequent processing on the image quality were investigated to determine the time of image acquisition in the field and the angle and height during the image acquisition by the UAV.After comparison,the ExG method is used for grayscale processing;the bilateral filtering not only effectively removes the noise,but also preserves the edge details well,which is the most suitable image filtering method for this experiment.(2)The RGB color space is briefly introduced,and according to the actual field environment of this subject,the image automatic segmentation method of G-0.8R+0 threshold color index is proposed.By comparing the actual effects of the three methods of NDI+Otsu,ExG+Otsu and G-0.8R+0 thresholds,it is concluded that the G-0.8R+0 threshold method has the best effect on segmenting green plants,and the method can adapt to the actual situation in the field.,to achieve automatic segmentation of the image,and without using the Otsu method to obtain the threshold.(3)In the morphological processing,the closed operation has the best effect on the filling of the holes in the rape leaves,which can keep the area of the region of interest basically unchanged,and the image of the edge and joint of the rape leaves is clear.When the canny operator is used to extract the edge of plant leaves,it has the advantages of edge coherence and leaf information hiding.(4)The shape and texture data of the plants were obtained,standardized and input into the SVM and BP neural network classification models to identify and classify rapeseed and weeds.The recognition accuracy of the three shape features input into the SVM and BP neural network classification models was 94.44% and 90.38%,respectively.The recognition accuracy of the six texture features input into the SVM and BP neural network classification models was 83.33% and 89.25%,respectively.(5)In order to improve the accuracy of the recognition model and shorten the running time of the program,the shape feature parameters and the texture feature parameters are merged and then reduced by the PCA.The six new variables were input into BP neural network and SVM for classification.The recognition accuracy in the SVM classification model was 94.44%,and the program running time was 0.0178 s.The recognition accuracy in the BP neural network classification model is 96.28%,and the program running time is 3.61 s.