Abstract

The accurate detection of potato orientation and bud eye positions is critical for guiding the end-effector of intelligent cutting robots. This study introduced Potato Orientation Detection You Only Look Once (POD-YOLO), a novel lightweight model based on YOLOv8n, designed for fast and precise detection of potato orientation and bud eye locations. Key innovations include replacing the Cross Stage Partial Dark Network (CSPDarkNet) with the Cross Stage Partial and Dual Partial Network (CSPDPNet) to reduce parameter count and improve detection accuracy. Additionally, the "no more strided convolutions or pooling" approach replaced downsampling modules in the backbone and neck, enhancing detection of small targets and low-resolution images. The regression loss function was further optimized by substituting Kalman Filtering Intersection over Union (KFIoU) for improved rotated bounding box performance. Experimental results showed that POD-YOLO achieved a mean Average Precision (mAP) of 97.2%, with a precision of 95.2%, recall of 94.0%, and detection time of 9.01 ms. With only 1.75 million parameters, POD-YOLO was lightweight and efficient, meeting real-time requirements. This research offers a robust and effective solution for automated potato orientation and bud eye detection, laying the groundwork for advanced agricultural automation.