A Real-Time Road Damage Detection Method Based on Detection Transformer

2025-10-23 Hits:

Journal:PRCV2025

Abstract:Road surfaces are prone to cracks and potholes due to environmental variations and material aging. Existing object detection algorithms exhibit inadequate adaptability to morphological diversity, feature ambiguity, and environmental interference in road damage, coupled with a persistent reliance on post-processing steps. To address these challenges, we propose an end-to-end real-time road damage detection
algorithm named RDD-DETR. To address feature ambiguity and background interference, an image enhancement method named M-USM and a residual context enhancement module (RCE Block) are designed to strengthen feature extraction capabilities in complex scenarios. A deformable attention-based feature interaction (DAIFI) module is constructed to optimize high-level feature interaction, while a cross-guided attention fusion (CGAF) module is established to achieve adaptive multiscale fusion through bidirectional feature complementarity. The end-toend
architecture is implemented to eliminate the dependency on NMS
post-processing inherent in existing YOLO-series algorithms. Experiments demonstrate that RDD-DETR achieves a 5.2% improvement in accuracy over baseline while maintaining speed comparable to YOLOseries algorithms.

Note:the 8th Chinese Conference on Pattern Recognition and Computer Vision (PRCV 2025), Shanghai, October 15-18, 2025

Co-author:Zhao Zhuoxuan,Fan Shenghua

Indexed by:Other

Correspondence Author:Chen Xi,Liu Haowen

Discipline:Engineering

Document Type:C

Translation or Not:no

Date of Publication:2025-10-15

Included Journals:SCI

Date of Publication:2025-10-15