Performance Comparison Between YOLO v4 and Mask R-CNN in Detecting Potato Tubers Running on Post-Harvest Conveyors

This research paper investigates the performance of YOLO v4 and Mask R-CNN models for potato tuber detection and segmentation. The study analyzes the outcomes obtained using the trained YOLO v4 model, revealing limitations in densely clustered scenarios due to difficulties distinguishing individual tubers from the background. Grayscale image processing did not significantly improve detection accuracy, underscoring the need for more sophisticated approaches. On the other hand, the Mask R-CNN model showcased superior segmentation quality with consistently high average mean intersection of union, which ranged between 73% and 80% depending on the density of the pile of potatoes.