ELW–YOLO: A next–generation lightweight model for pavement defects detection

Abstract

: Reliable pavement–distress detection remains difficult because defects vary greatly in size, geometry, and visual appearance. We present ELW–YOLO, a lightweight detection framework built on YOLOv2026n with three targeted additions: (1) an Enhanced Multi-Scale Attention (EMA) block at P3/P4 that captures spatial–channel cues at several scales and suppresses background interference; (2) a Localization Quality Estimation (LQE) head that predicts localization quality alongside class scores to improve NMS ranking; and (3) a Wise–IoU loss with attention-based outlier suppression for high aspect–ratio variance. On RDD2022 China (26,620 images, four defect classes), ELW–YOLO reaches 86.0% mAP@0.5 and 55.9% mAP@0.5:0.95, beating the baseline by 3.6 and 4.4 points while using only 2.7M parameters, 6.5 GFLOPs, and 74 FPS.