基于全卷积网络与凹点搜索的重叠葡萄分割算法  

Overlapping grape segmentation algorithm based on full convolutional network and concave point search

作　　者：雷旺雄 卢军[1] LEI Wang-xiong;LU Jun(College of Mechanical and Electrical Engineering,Shaanxi University of Science&Technology,Xi'an 710021,China)

机构地区：[1]陕西科技大学机电工程学院,西安710021

出　　处：《光电子．激光》2021年第3期231-240,共10页Journal of Optoelectronics·Laser

基　　金：陕西省科技厅工业攻关基金项目(2016GY-049)资助项目。

摘　　要：为解决自然环境下重叠葡萄难以分离的问题,提出了一种基于改进的FCN全卷积网络和凹点搜索的重叠葡萄分割方法。主要步骤如下:第一,对重叠葡萄图像进行H分量提取,然后通过改进的FCN全卷积网络得到二值图,对二值图进行质心位置求解和轮廓凸包检测;第二,根据矢量夹角最大法进行单串葡萄轴线检测,分离前的轮廓凸包与二值图像相减得到凹区域,然后根据轴线做对称凸包筛选出分割凹点所在的凹区域;第三,建立重叠葡萄分界线几何计算模型,再根据最短距离进行分割凹点搜索。此外,对采集的400幅重叠葡萄图像进行验证,平均分割标准误差为4.03°,检测成功率为92.5%,算法消耗时间为0.40~0.59 s,为采摘机器人提供了一种新的重叠葡萄分割方法。To solve the problem that overlapping grape is difficult to separate in natural environment,a method of overlapping grape segmentation based on improved FCN and concave point search was proposed.The main steps are as follows:First,a comparative analysis of different color spaces and different color channels was conducted for the collected summer black overlapping grape images.It was found the contrast between the target and the background was large in the H component under the HSV color space,so the H component was extracted from the overlapping grape images.The H component of the overlapping grape was input into the convolutional multiscale fusion FCN to obtain the binary graph,and then the centroid of the overlapping grape was solved and the contour convex hull was detected.Second,axis of a single bunch of grapes was detected according to the maximum angle of the vector.The concave region was obtained by subtracting the binary image from the convex hull of the contour before separation.Symmetrical convex hull was made according to the axis,and then the concave region containing segmented concave points was obtained.Third,the geometric calculation model of overlapping grape boundary was established,the contour of the selected concave region was traversed,and then the segmented concave point was searched according to the shortest distance.In addition,the 400 overlapping grape images were verified,the average standard segmentation error was 4.03°,the detection success rate was 92.5%,and the algorithm consumed time of 0.40~0.59 s.Compared with the existing methods,the accuracy of overlapping grape segmentation in this paper was not easily affected by the details of contour inflection points,and the accuracy of recognition was improved.It provided a new overlapping grape segmentation method for picking robot.

关 键 词：重叠葡萄分割 全卷积网络 凹点搜索 凸包检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]