Image segmentation method for green crops using improved mean shift

Wu Yanlian,Zhao Li,Jiang Haiyan,Guo Xiaoqing,Huang Fen

DOI: https://doi.org/10.3969/j.issn.1002-6819.2014.24.019

2014-01-01

Abstract:Digital image processing technology has received considerable attention in many aspects of agriculture, some typical examples including estimation physiological status of crops, disease and insect pest identification, vegetation-cover estimation, and quality detection for agricultural products. One of the most important and essential tasks is the crop image segmentation which separates the green crop material or region of interest from the background. In recent years, green crop image segmentation has been an important research topic and several methods have been proposed. However, green crop image segmentation is still a difficult problem since the green crop images usually involve complicated backgrounds. To deal with this problem, we propose in this paper a novel segmentation method based on Mean shift and color index. Mean shift is an iterative procedure that shifts each data point to the average of data points in its neighborhood. The performance of Mean shift depends heavily on the size of bandwidth which means that bandwidth selection is a key issue in mean shift-based image smoothing and segmentation. Classical Mean shift segmentation method needs spatial bandwidth and range bandwidth to be initialized, which usually leads to lower segmentation precision. We present an improved Mean shift algorithm by using adaptive spatial and range bandwidth. Firstly, color index was extracted according to the color feature of the green crop image in the RGB color space. Secondly, with the extracted color index, images were smoothed and segmented by Mean shift algorithm. The proposed improved Mean shift algorithm employs an adaptive bandwidth strategy where the adaptive spatial bandwidth is determined according to the color distribution of the images by combining color information and spatial information. It means that a small spatial bandwidth is suitable for images containing much more detailed information, while images containing large flat areas require a larger bandwidth. This approach smoothed the images without the loss of detailed information. In addition, adaptive range bandwidth can be obtained by Asymptotic Mean Integrated Square Error (AMISE). Finally, with Otsu method, the images were classified into two parts:green and non-green. In order to verify the performance of the proposed method, the comparison experiments have been carried out. Different test images containing green crops were utilized to compare the proposed method with the color index-based segmentation methods such as ExG and CIVE methods, which have been widely used recently. These test images were acquired under field conditions and natural light conditions, covering different crop and soil types. Experiments showed that the results of the proposed method were superior to that of ExG and CIVE. Compared to the ExG and CIVE methods, there are less small black and white regions and segmentation errors in the segmentation results of the proposed method, particularly for the images that included strongly shadowed parts and some crop straws. Experiment results also demonstrated that our method was more insensible to soil types and illuminant variations compared with the ExG and CIVE methods and that the average segmentation errors of green crop images were less than 6.5%. In summary, the proposed segmentation method in this paper can segment the green crop effectively and obtain better performance than the traditional color index methods.