基于区域生长均值漂移聚类的苗期作物行提取方法

Seedling crop row extraction method based on regional growth and mean shift clustering

为解决传统机器视觉方法检测苗期作物行时不同作物种类、不同生长背景和不同作物行数导致的作物行提取精度低的问题,该研究提出一种基于区域生长和均值漂移聚类的苗期作物行提取方法。首先,通过Lab颜色空间中a、b双颜色分量最大熵法选取最优阈值进行图像分割;其次,通过垂直投影获取均值漂移的聚类窗口带宽,均值漂移时以聚类窗口边缘为种子点进行区域生长来归类和标记每一行作物,之后遍历所有作物行获取聚类中心点;最后,通过最小二乘法拟合聚类中心点得到作物行直线。试验结果表明,该方法对大蒜、玉米、油菜、水稻和小麦5种作物的苗期作物行提取精度较高,5种作物的平均行识别率为98.18%,平均误差角度为1.21°,每张图片的平均处理时间为0.48 s。该方法的作物行提取性能明显优于Hough变换方法,为田间环境多因素影响下的苗期作物行提取提供了一种更具鲁棒性的方法。

Automatic navigation can be used to significantly improve the operation accuracy and efficiency of agricultural machinery. Particularly, machine vision-based automatic navigation can greatly contribute to crop row detection. In this study,a novel crop row extraction was proposed using regional growth and mean-shift clustering, especially for higher accuracy of crop row extraction under different crop types, the number of crop rows, and growing backgrounds. Firstly, the a and b components of an image were obtained in the Lab color space, and then the maximum entropy values of a and b components were calculated for the optimal segmentation threshold, after which the image was segmented by the threshold for the binarization image. Secondly, the vertical projection operation was performed on the top strip of the binary image, where the mean value of the vertical projection curve was calculated to distinguish crop and non-crop areas. The minimum distance between crop areas was selected as the bandwidth of the crop clustering window. The top center pixel of the whole image was selected as the initial center point of the clustering window. The clustering center point moved from the center to both sides of the top of the image with the iteration of crop row clustering, where the shift vector was calculated in the clustering window.The clustering center point moved along the shift vector in single row clustering, where the edge of the clustering window was used as the seed point for regional growth. As such, all crop rows were obtained by the movement of clustering window and regional growth, while, the clustering center points of each crop row were grouped into a cluster. Lastly, least-squares fitting was performed on these clustering center points to obtain crop row lines. A total of 170 seedling images of five crop varieties were obtained to verify the feasibility of the method, including garlic, corn, oilseed rape, rice, and wheat. Hough transform and projection-proximity classification were also used to extract crop rows for comparison. Experimental results showed that more satisfactory performance of segmentation was achieved for the images with less significant color difference between crops and growing background using the maximum entropy of a and b components in the Lab color space, compared with the conventional segmentation using an excess green index. Furthermore, the crop row extraction for tested five crops performed better than that of Hough transform and projection-proximity classification fitting, in terms of row recognition rate, mean error angle, and mean processing time. The mean row recognition rate for the 170 tested images was 98.18%, the mean error angle of extracted straight lines of all crop rows was 1.21°, and the mean processing time for each image was 0.48 s. This finding can provide a more robust for crop row extraction under the influence of multi factors in the field using machine vision,particularly on real-time embedded platforms in practical applications.