Sweet Pepper Picking Point Recognition Based on Improved YOLOv5s and Multi-Information Clustering

In this paper, we study a machine vision-based method for Sweet pepper picking point recognition which is an important part of the automatic picking process realized by the Sweet pepper picking robot. To obtain accurate picking point location coordinates, we propose an improved YOLOv5s algorithm and a multi-information K-means++ clustering algorithm to recognize pepper picking points. Firstly, the color image is subjected to Contrast Limited Adaptive Histogram Equalization to adjust the contrast of the image. the pepper fruits are identified by an improved YOLOv5s in which the loss function and NMS algorithm are improved. The improved YOLOv5s algorithm was compared with the original algorithm on the sweet pepper dataset, and the improved algorithm achieved the mAP of 89.9%, which was 0.7% higher than the mAP of the original algorithm. After that, the region of interest of the picking point is determined according to the location of the prediction box. The depth image is preprocessed and the depth information and color components are selected for K-means++ clustering to segment the target class and the background class. Finally, after morphological processing to filter out noise and image refinement, the accurate identification of pepper stem is realized, and the picking point is located to provide accurate picking point location information for the pepper picking robot and improve the success rate of pepper picking.