Research Of Feature Extraction Oriented To Weed Identification

It is inevitable that intensifying efforts to promote precision agriculture progress, while weed is a main restriction factor for agriculture development. In order to realize variable volume spraying, technology of weed identification on the basis of image processing has become a focus in the study of weed identification. In weed identification based on machine vision, choice of feature extraction method affects the final recognition results, the accuracy of single feature is unable to accurately identify weeds in cotton fields. Therefore, the research on weed identification technology by feature fusion is of important theoretical and practical significance.For weed identification based on machine vision, the following work is done:Firstly, image segmentation. Image segmentation for weed image. Images(endives, eclipta prostrate, calystegia hederacea wall, Amaranthus retroflexus L and Amaranthus lividus L.) are grayed before segmentation of weed leaf image. By comparing results of image graying by average method, the weighted average method and maximum method, HSI paper chooses the weighted average method and uses median filtering method to filter gray images, Then, Images are segmented by gray histogram, the between-cluster variance method and maximum entropy threshold method, results show that the effect of maximum entropy threshold method is best. So HSI paper chooses maximum entropy threshold method. Finally, HSI paper separately using Sobel operator, Roberts operator, gaussian filtering(LOG) operator, Prewitt edge detection operator and Canny operator, and the experimental results show that the gaussian filter operator edge detection processing results is better, so HSI paper select Gauss-Laplacian operator when ? ?2.Secondly, multi-type features. are extracted and fused. In color feature extraction, by comparing, FMS, SMS and TMS in HSI are extracted by color moment and as color features. In shape feature extraction, by comparing, REC, RWL, CIR and SPH are extracted by geometric parameter method and are as shape features. In texture feature extraction, ASM, CON and COR are extracted by GLCM. Then this paper uses the relevant reserves combined with principal component analysis of dimension reduction techniques for feature space dimension reduction, according to the characteristic parameters of the correlation coefficient matrix retains the characteristic parameters of the correlation coefficient is less than 0.7 characteristic parameters for contrast, correlation, second moment, and then the principal component analysis is adopted to the rest of the fusion feature parameters for dimension reduction, access to two principal components. The feature parameters kept and principal components are as new features. COR, REC and ASM consist of feature space for weed identification.Lastly, For five types of weeds in 200 samples, 160 samples of the training set, 40 samples as test set, based on local weighted k- near the center of mass neighbor classification algorithm is used to identify the weeds and other traditional methods of comparative test. Three comparative experiments are carried out. Three experiments include identification of five kinds of weeds, three kinds of weeds and two kinds of weeds. The highest accuracy of identification by method proposed in HSI paper in three conditions are 88%,87.5% and 86.5% respectively, which is higher than state of the art. Accuracy of identification by HSI method is also higher than state of the art for every kind of weed.Weed identification based on multi-type features can solve the problem of weed identification in cotton fields. Experimental results show that method proposed in HSI paper is superior to state of the art and is suitable for identification of multi-class weeds. HSI method can also be applied in identifying weeds in other fields.