Algorithm for recognition and image segmentation of overlapping grape cluster in natural environment

Abstract：In order to solve the problem that it is difficult to identify and segment the overlapping grape clusters in the naturalconditions, this paper selected Xiahei grape as the experimental sample to carry out relevant research. Firstly, the RGB colorchannel was transformed into HSV color channel and YCbCr color channel. Through histogram analysis and comparison ofeach component of HSV color channel and YCbCr color channel, H component in HSV space was determined as theexperimental object of image preprocessing. Then, K-means clustering algorithm was used to extract the H component in HSVcolor space. To obtain the region of overlapping grape clusters, Sobel operator was used to extract the contour of theoverlapping grape clusters. Then the contour analysis method was used to analyze the contour of the overlapping grapeclusters, and the contour extreme points and the central point of the overlapping grape clusters were obtained. By analyzingthe aspect ratio of grape cluster region, we could judge the overlapping type of grape clusters. If it was a single grape cluster, itwould be identified directly. If it was a veritcally overlapping grape cluster, rotated the image 90° anticlockwise, and then usedthe same processing method as the left and right overlapping grape clusters proposed in this paper. The general target grapecluster area was obtained by contour analysis. After that, we could solve the quasi-central point by the central point betweenthe midpoint and the right extreme point, and the intersection point of the two lines between up and lower extreme points andleft and right extreme points in the target grape cluster region. The improved Chan-Vese model limited the iteration times ofthe Chan-Vese model by using the distance between the center point and the quasi-central point. After that, based on the ChanVese model and the quasi-central point as the origin, the iterative identification of target grape cluster was carried out. Thebinary image processing was carried out for the grape cluster area obtained by Chan-Vese, and the contour information of theoverlapping boundary was obtained. Finally, image fusion was used to fuse the grape cluster contour obtained by Chan-Vesemodel with the original contour. The test results show that the average recognition accuracy of the target grape cluster on theleft of the overlapping grape cluster was 92.19%, and the average false positive rate was 2.82%. The average recognitionaccuracy of the target grape cluster on the right of the overlapping grape cluster was 87.22%, and the average false positiverate was 6.03%. The average recognition accuracy of all the target grape cluster was 89.71%, and the average false positiverate was 4.24%. Compared with the existing segmentation method which only based on contour analysis, the method proposedin this paper could recognize and segment the whole target grape cluster, and greatly improve the accuracy of recognition andsegmentation. This method is more conducive to acquisition the information of the target grape cluster picking point，thestudy provide an efficient and stable recognition and segmentation algorithm for the grape picking robot to successfully pickoverlapping grape clusters.