RSF-DETR：空频增强与上下文重构的路面损伤检测

打开文本图片集

中图分类号：TP391.41 文献标识码：Adoi：10.37188/OPE.20253322.3549 CSTR：32169.14.OPE.20253322.3549

Abstract： Aiming at the problems of various pavement-damage forms，low detection accuracy and high miss-detection rate，this paper proposed an improved method based on the RT-DETR model. First，inspired by the joint idea of high-frequency edge enhancement in the spatial domain and global feature extraction in the frequency domain，the spatial- frequency dual-domain feature-enhancement module FreSCal was designed to strengthen the model's ability to extract target and edge information and to improve its capacity to distinguish target regions from background. Secondly，drawing on the context-guided feature-reconstruction concept of the CGRSeg network，the context-guided spatial feature-reconstruction pyramid network RSDFPN was proposed. By building a scale-aware semantic pyramid and a dynamic feature-fusion mechanism，the model's capability to fuse features for multi-scale targets was significantly enhanced. Finally，through dynamic group convolution shufling and the global modeling capacity of Transformer，efficient spatial-domain feature enhancement and frequency-domain context fusion were achieved，raising the model's detection accuracy for target recognition. The experimental results show that the improved method in this paper has achieved significant improvement on both RDD2022 and UAV-PDD2023 mainstream datasets，mAP @0.5 Compared with the baseline method，the indicators are increased by 1.9% and （204号 3.7% respectively，which can provide an effective technical support for pavement damage detection.

Key Words： pavement damage detection；Real-Time Detection Transformer（RT-DETR）； space fre quency dual domain；context guided reconstruction；Dynamic Group Convolution Shuffle Transformer（DGCST）

1引言

在全球城市化加速进程中，道路设施面临高频荷载与环境侵蚀的多重威胁，传统人工检测因效率瓶颈难以满足需求，而基于深度学习的检测算法通过精准识别路面裂缝及坑洼损伤，为构建道路健康智能监测系统开辟了新路径。（剩余20769字）