Rice canopy image segmentation based on statistical histogram k-means clustering

CHEN Ke-yin;WU Chong-you;GUAN Zhuo-huai;LI Hai-tong;WANG Gang;School of Physics and Electronic Engineering, Jiaying University;Nanjing Institute of Agricultural Mechanization,Ministry of Agriculture and Rural Affairs;

rice canopy image;;image segmentation;;statistical histogram;;k-means clustering

A rice canopy image segmentation method based on statistical histogram k-means clustering was proposed to solve the problems of existing k-means methods, such as sensitivity to the initial cluster center, error prone segmentation and low operation efficiency. Firstly, according to the image histogram, the pixel value with large difference in the number of pixels was selected as the initial cluster center of the rice canopy image. Secondly, the prior correspondence between the pixel value in the image histogram and the number of pixels in the image was used to weight the clustering objective function. Finally, the k-means clustering framework was used to cluster and segment the rice canopy image. In order to verify the effectiveness of the proposed method, comparative experiments were carried out with the feature pixel extraction methods of rice canopy images based on k-means, k-means++, k-mc2 and afk-mc2. The results showed that the canopy images of Lindao 20, Wuyunjing 32 and Zhongjing 798 were clustered and seymented at mid-maturing stage, and the average intra-class sequare errors were 36.6, 29.5 and 29.5 under normal light conditions, the average inter-class square errors were 67.5, 51.8 and 51.8, and the average running time was 0.2 s, 0.3 s and 0.3 s. Under strong light conditions, the average intra-class square errors were 15.2, 4.9 and 12.1, the average inter-class square errors were 30.9, 9.5 and 21.2, and the average running time was 0.3 s, 0.2 s and 0.3 s. The results of the method used in this study are better than the above four traditional clustering methods.