Research On Weed Identification Based On Support Vector Machine In Corn Field

Agricultural robot is one of the important equipments for precision agriculture in the21st century, and smart spraying of herbicide achieved by agricultural robot becomes a direction for the development of intelligent agricultural machinery in the future. Weed identification is the first and important step for smart spraying. After studying the weed identification methods in the field at home and abroad, this paper creatively proposed background segmentation and weed identification for weed identification in complex environment of corn field. Then neighborhood histogram, multi-level shape analysis and support vector machine were deeply studied. The main research conclusions are as follows:1. The extra-green(2G-R-B) images were got from the same environment, and the neighborhood histogram was took as the feature vectors. Then the binary images achieved by SVM were compared with images labeled by hands to pick out ideal kernel function and neighborhood window models.2. The difference of images was caused by different environment, training samples without labels were extracted automatically from the images to be segmented, then semi-supervised support vector machine and k-means clustering algorithm were applied to train classifiers using the existing training samples with labels and training samples without labels. Then the most appropriate model for the environment was found out to improve the quality of image segmentation with transductive support vector machine (TSVM).3. This paper proposed a method for corn-weed recognition based on the combination technique of multi-level shape analysis and SVM. The difference in the shape of whole plants and structure distribution of leaves between com and weed was considered, then multi-scale ring shape features and multi-angle shape features were proposed. At last, the SVM classification method was applied to classify weed and corn. The experimental results showed the feasibility of presented method.The proposed algorithm could improve the accuracy and reliability for weed identification, and provided a technical basis for accurate weed identification and herbicide spraying.