A lightweight real-time method for detecting irregular potato surface defects for sorting lines

To address limited accuracy, excessive redundancy, and real-time deployment constraints in current potato surface defect inspection methods, this study developed a lightweight real-time detector based on YOLOv11n, called You Only Look Once for Potato Surface Defects (YOLO-PSD). First, the backbone and neck C3k2 blocks were replaced with C3k2-ConvFormer-CGLU (Convolutional Gated Linear Unit) modules, enhancing multi-scale texture and boundary representation, which yields higher detection accuracy at lower computation. Second, the baseline detection head was substituted with LSCD-LQE (Localization-Quality Estimation–Lightweight Shared Convolutional Detection Head), which employs a shared lightweight convolutional stem, decoupled regression and classification branches, and DGQP (Distribution-Guided Quality Predictor) for localization-quality estimation, suppressing high-score/low-quality boxes while improving separability under overlapping targets. Finally, Wise-IoU was adopted as the bounding-box regression loss; its adaptive focusing mechanism improves localization accuracy at high IoU thresholds and promotes stable late-stage convergence. On the potato-defect benchmark, YOLO-PSD attains 95.1% precision, 89.9% recall, and 95.6% mAP, surpassing the baseline by 4.4, 4.4, and 3.1 %age points, respectively. The computational cost and parameter count reach 5.4 GFLOPs and 2.16 M, making reductions of 15.6% and 16.3% compared with the baseline. In real industrial deployment on an intelligent sorting machine, YOLO-PSD achieves 12.6 ms latency, with online sorting accuracy of up to 94.7% under the tested conveyor-speed conditions, underscoring its practical robustness and readiness for engineering applications. Collectively, this work delivers a high-accuracy, lightweight, and low-latency solution for automated detection and sorting of potato surface defects, handling scalable adoption in advanced agricultural automation.