Edge-Enhanced and Content-Aware Feature Reconstruction for Road Object Detection in Adverse Weather
TL;DR
WRA-YOLO, an improved YOLO11-based detector integrating edge enhancement and content-aware feature reconstruction, and MPDIoU loss is employed to improve bounding box regression by constraining corner-point distances.
Abstract
Road images captured under adverse weather often suffer from low contrast, blurred boundaries and rain/snow noise, which degrades the robustness of object detectors. To address this issue, this paper proposes WRA-YOLO, an improved YOLO11-based detector integrating edge enhancement and content-aware feature reconstruction. First, a depthwise wavelet transform convolution module (DW-WTConv) is introduced into the early backbone to explicitly decompose low-frequency structures and high-frequency edge details, thereby reducing detail loss caused by downsampling. Second, a content-adaptive feature recalibration module (BlockAttnRes) is designed to dynamically fuse multi-level features using learnable pseudo-query vectors and RMSNorm, suppressing noisy semantic responses in adverse weather. Finally, MPDIoU loss is employed to improve bounding box regression by constraining corner-point distances. Experiments show that WRA-YOLO improves mAP@0.5 from 72.28% to 75.42% on DAWN and from 63.40% to 65.56% on RTTS, demonstrating better adaptability to road scenes under fog, rain, snow and sandstorm conditions.