Research Of Feature Extraction And Recognition Method Of Weed Image Based On Machine Vision

In traditional agriculture, under the condition that distributing of weed and crop doesn't be uncertain of, the overlapping spraying operation will waste too much herbicide, and result in the environment pollution. For solving this problem, it is one of the key technologies for research of precision farming that identify the weeds from agriculture crops and soil background by means of machine vision technique, spot -spray the herbicide quantitatively. Presently the theory, feature parameters, research advances and problems using position, the shape, texture, color and multi-spectral features to identify weed were discussed, respectively. Reviewing of relative research at home and abroad, for the study of weeds in wheat and cotton fields, a few new methods were studied to detect weed according to color feature, combinative shape and color features of crop and weed. The image processing system of infield weed detection based on machine vision that was able to detect weed and spot-spray herbicide in the field was developed.Aim at the severe occluding of leaves of weed and wheat, this paper presents a weed identification method that color feature is utilized to distinguish plants and background: using a method that takes YIQ as color-space and I as characteristic variant and improved method of maximum classes square error as criterion; color feature is utilized to distinguish wheat and weed: using a method that takes HSI as color-space and homogeneity of I and s as characteristic variant separately and hierarchical homogeneity segmentation as criterion; ultimately morphological feature is utilized to obtain weed: using a method that combines morphological opening and closing filter and AND operation algorithm. This paper proposes the use of hole filling algorithm and the small area of elimination algorithm on regional markers so as to slove the problem that some binary images have a little small white area in black background or have a little small black holl in plant leaves. The proposed methods together with a chemical weeding system are simulated and the efficiency of the overall systems are evaluated theoretically. Experimental results show that these methods are effective. Various weeds of the particular cotton fields were analyzed, calculate four RST invariance shape parameters of weed leaf: Roundness, Elongation, Dispersal and Compactness, and ther shape parameters of weed leaf: Area, Perimeter, etc. The shape feature parameters database of weeds is established primarily. Aim at the severe occluding of leaves, the algorithms of morphological operators and label watershed segementation based on the mathematical morphology were put forward to resolve the problem that parameters extraction of shape features is more difficult.This paper proposes some methods of feature extraction and recognition for infield weeds, which will provide the theoretical support and viable project for variable herbicide spraying. The results in this research are of definite instruction to development of the automatic killed-weed system based machine vision.