APM-DETR: adaptive perceptual multi-scale DETR for road defect detection in complex scenes
TL;DR
This study proposes an adaptive perceptual multi-scale detection transformer (APM-DETR), a detection network customized for real-world complex road environments that achieves highly competitive detection accuracy, outperforming existing models on multiple defect types.
Abstract
Road defect detection is crucial for automated infrastructure inspection and traffic safety, yet it remains challenging due to extreme scale variations in defects, irregular morphologies, and complex background interference. To tackle these issues, this study proposes an adaptive perceptual multi-scale detection transformer (APM-DETR), a detection network customized for real-world complex road environments. Taking the real-time DETR as the baseline model, the proposed framework integrates two targeted enhanced modules for road defect detection tasks: a deformable frequency-spatial perception backbone (DFSA-Net), which is designed to enhance the discriminative feature representation of irregular road defects and effectively suppress redundant road background noise; a multi-scale adaptive enhancement feature fusion pyramid network for enriching multi-scale feature expression of road defects with varying scales and morphologies. Extensive experiments and visualization analyses on the UAV-PDD2023 and China-D datasets validate the effectiveness of the proposed method, and confirm that APM-DETR achieves highly competitive detection accuracy, outperforming existing models on multiple defect types. The relevant code is available at: https://github.com/qmlb908/APM-DETR.