Abstract: Aiming at problems of low efficiency and strong subjectivity in manual potato sorting, insufficient defect recognition accuracy under complex illumination, weak generalization ability of small-sample detection, and difficult deployment of edge devices, a potato surface defect detection method integrating multi-layer image preprocessing and lightweight YOLO26s model was proposed．This model applied image preprocessing to extract and eliminate background interference, effectively enhancing defect features in complex environments．At the model level, lightweight YOLO26s network architecture was adopted, and transfer learning was implemented with pre-trained weights loaded, which alleviated overfitting risk in small-sample training and balanced detection accuracy and edge-side inference efficiency．Experimental results showed that proposed method achieved a mean average precision at mAP50 of 0.881．Inference time for a single image was 8.5 ms on training platform and only 186 ms on the Raspberry Pi 5 edge device．Compared to baseline models YOLO11n and YOLOv8n, comprehensive detection performance has been improved by 10% and 9% respectively．This method boasted excellent detection stability and high accuracy in small-sample scenarios, with sound adaptability to edge devices．It could provide technical reference and solution support for industrial application of intelligent potato sorting equipment．