Automatic road damage recognition based on improved YOLOv11 with multi-scale feature extraction and fusion attention mechanism

Abstract

Rapid urbanization and growing traffic volumes have increased the demand for efficient and accurate road damage detection to ensure traffic safety and optimize maintenance. Traditional manual and vehicle-mounted inspection methods are often inefficient, costly, and prone to error. Deep learning-based approaches have made progress but still face challenges in detecting small objects, handling complex backgrounds, and meeting real-time requirements due to high computational costs and limited generalization. This study proposes an improved road damage detection method based on YOLOv11, incorporating a Tiny Object Detection Layer for enhanced small object recognition through high-resolution and multi-scale feature fusion. A Global Attention Mechanism is integrated to emphasize critical regions and suppress background noise. Additionally, lightweight convolution modules (C3k2CrossConv and C3k2Ghost) optimize the network to reduce computational complexity and improve inference speed. Experimental results on the RDD2022 dataset show that the YOLOv11-ATL model achieves 3.2% and 3.1% gains in mAP@0.50 and mAP@0.50:0.95, respectively, demonstrating robust performance in complex environments while maintaining a favorable balance between accuracy and efficiency. Overall, the proposed approach offers a practical and effective solution for intelligent road damage detection, supporting urban infrastructure management and intelligent transportation systems.