A semantic‐enhanced transformer with adaptive fusion for road damage detection
Yuan Dai, Tingwei Zhang, Wei Zhou, Kaixiang Kuang, Kejun Long, Xi Lu, Shaofei Wang
Abstract
Road damage detection faces significant challenges including extreme scale variations, complex visual interference from road textures, diverse orientational patterns, and irregular boundaries. This paper proposes a semantic-enhanced and adaptive fusion detection transformer to address these domain-specific challenges through two synergistic innovations. The semantic enhancement attention module exploits distinctive frequency-domain characteristics of road damages through learnable spectral processing, where damaged regions exhibit 50.5% higher high-frequency energy, compared to intact surfaces, enabling effective discrimination between structural defects and background interference. The adaptive information fusion module implements a three-stage progressive architecture: loss-less transmission establishes information integrity across extreme scales through amplitude-aware upsampling and attention-driven fusion; omnidirectional pattern capture via multi-directional convolutions addresses diverse damage orientations; dual-path processing optimizes computational efficiency. Comprehensive evaluation across four datasets demonstrates state-of-the-art performance with significant improvements: 83.4% mean average precision at intersection over union threshold 0.5 on UAV-PDD2023 (+3.4% over previous best), 31.2% on CNRDD (+1.3%), 61.9% on RDD2020 (+3.0%), and 90.2% on nighttime NPD (+0.6%), while achieving superior efficiency with 62 giga floating-point operations, 20 million parameters, and 51 frames per second inference speed for real-time processing.