LE-YOLO: A Lightweight and Enhanced Algorithm for Detecting Surface Defects on Particleboard

Abstract

Current algorithms for surface defect detection in particleboard often encounter limitations such as high computational complexity and excessive parameter scale. To address these challenges, this study proposes the LE-YOLO model, which incorporates a normalized Wasserstein distance into the loss function to enhance the detection capability for minute surface defects. A dynamic mixed convolutional network module is introduced to construct a lightweight backbone architecture. Moreover, the Shared Dilated Feature Pyramid (SDFP) module is employed in the neck network, effectively reducing computational overhead while preserving detection accuracy. A lightweight detection head was further designed, integrating shared convolutional operations with a distribution-aware loss function, thereby substantially improving detection performance in complex textured environments. Experimental evaluations conducted on the Chipboardv1.0 particleboard surface defect dataset demonstrated that compared to the baseline YOLOv11n model, LE-YOLO achieved a 5% improvement in recall, a 1% increase in F1 score, a 4% enhancement in mAP@50, a 6% gain in mAP@50–95, a 12.69% acceleration in inference speed, and an 18.6% reduction in parameter count. Compared with other models, the proposed approach not only improved detection precision but also effectively reduced model complexity, achieving a lightweight and efficient detection framework.