CAJ | 학술논문

针对绿色农作物图像背景复杂且分割难的问题，提出一种基于改进均值漂移算法的分割方法。采用均值漂移算法对图像进行平滑和分割时，带宽的选择直接影响平滑和分割的结果。传统的均值漂移分割方法需要人为地设定空域带宽和值域带宽这2个参数。该文首先根据绿色作物图像的颜色特点，提取图像的颜色指数；然后采用均值漂移算法，将图像的颜色信息与空间信息结合起来，根据作物图像颜色分布的丰富程度定义自适应空域带宽，采用渐近积分均方差来获得自适应值域带宽，对图像进行平滑滤波；最后采用 Otsu 方法将平滑后的图像分成两部分：绿色部分和背景部分。试验结果表明，该方法能够有效地分割出绿色作物，并在分割性能上明显优于常规的颜色指数方法，作物图像的错分率均小于6.5%。
침대록색농작물도상배경복잡차분할난적문제，제출일충기우개진균치표이산법적분할방법。채용균치표이산법대도상진행평활화분할시，대관적선택직접영향평활화분할적결과。전통적균치표이분할방법수요인위지설정공역대관화치역대관저2개삼수。해문수선근거록색작물도상적안색특점，제취도상적안색지수；연후채용균치표이산법，장도상적안색신식여공간신식결합기래，근거작물도상안색분포적봉부정도정의자괄응공역대관，채용점근적분균방차래획득자괄응치역대관，대도상진행평활려파；최후채용 Otsu 방법장평활후적도상분성량부분：록색부분화배경부분。시험결과표명，해방법능구유효지분할출록색작물，병재분할성능상명현우우상규적안색지수방법，작물도상적착분솔균소우6.5%。
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.