Research Of Weed Recognition Method Based On Support Vector Machine

Presently, extensive herbicide spraying pollutes the environment, destructs farmland and increases agricultural production costs. For solving this problem, domestic and foreign experts propose "precision agriculture".It uses machine vision technology to identify weed from agriculture crops and soil background, and then, spot-spray the herbicide quantitatively according to weed's distribution. Reviewing of relative research at home and abroad, for the study of seeding wheat and four weeds , this paper focuses on the weed detection according to color feature, combinative shape and texture features of weed. The image processing system of infield weed detection based on machine vision that was useful to detect weed and spot-spray herbicide in the field was developed.In line with the classification feature of support vector machine (SVM), weed classification in essence is a complex multi-class identification problem of small sample, high-dimensional and nonlinear. Thus, this paper presents a weed recognition method based on SVM. The main content and achievements are presented as follows:1. Background segmentation. Take L*a*b* as color-space and a* as characteristic variant according to the color difference between plant and soil. Compared with the effect of other two segmentation algorithms, improved thresh method of maximum classes square error was chosen to change gray image into binary image. Complete plant leaves were gained by dealing with the binary image through morphological and region labeling algorithms.2. Identification features extraction. Study the image features include geometric characteristics, nondimensional shape features and texture features. Through statistical analysis of feature data, three RST invariant (for image rotation, scale, shift changes are constant) shape parameters and three texture parameters were used as plant identifying features.3. Classifier design. Construct 4-fold cross-validation shape and texture classifiers with SVM .In the experiment, grid search method was used to optimize the kernel parameters and the punishment parameter C. Classification results of polynomial, RBF and Sigmoid kernel function were compared. The results showed RBF kernel function was most suitable for plant classification. The recognition rate of feature fusion SVM classifier is 95%(wheat), 87.5%(portulaca olercea), 92.5% (amaranthus retroflexus), 90% (cirsium segetum), 90%(digitaria ciliaris).This paper presents a weed classification method based on SVM with the combination of shape and texture features. This method distinguishes wheat from four types of weeds on the farmland with the accuracy of 91%. It will provide theoretical support and viable project for variable herbicide spraying. Compared with other classification methods, this SVM-based classification method has advantages in high accuracy, fast recognition speed, simple algorithm and strong generalization.