RADNet: Adaptive Spatial-Dilation Learning for Efficient Road Crack Detection

Road crack detection is crucial for infrastructure maintenance and traffic safety, yet existing methods struggle to balance detection accuracy and computational efficiency due to complex texture similarities between cracks and road surfaces. In this paper, we propose RADNet, a lightweight framework for efficient road crack detection through adaptive spatial-dilation learning. Our approach introduces three key innovations: an adaptive spatial learning module (ASDown) for efficient feature extraction, a multi-scale dilation strategy (C2f-MSD) for comprehensive crack pattern capture, and a lightweight detection enhancement mechanism. Experiments on the RDD 2022 dataset demonstrate that RADNet achieves state-of-the-art performance with <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$86.6\%$ </tex-math></inline-formula> precision, 78.7% recall, and 83.8% mAP50, while maintaining a lightweight architecture of 2.47M parameters and 9.4 GFLOPs at 370 FPS. Ablation studies show that the integration of ASDown and C2f-MSD significantly enhances both detection accuracy and computational efficiency, improving mAP50 by 5.6% over the YOLOv8n baseline while reducing computational complexity.