A novel road damage detection framework with multi-scale feature enhancement

Abstract

The health condition of road transportation infrastructure is directly related to driving safety and transportation efficiency. However, existing automated detection technologies still struggle to meet the requirements of large-scale, high-precision, and real-time inspection when confronted with challenges such as significant scale variation of damage targets, strong background texture interference, and fine cracks. This paper proposes a road defect detection model comprising C2f_PBA, CARAFE, and WPIoU, based on the you only look once version 8 (PCW-YOLO). The C2f_PBA module was designed and applied to the backbone network, employing a multi-branch parallel attention mechanism to extract features at different scales. This effectively enhances the model’s ability to capture heterogeneous damage. In the neck network, a CARAFE content-aware upsampling operator is introduced to replace conventional nearest-neighbor interpolation, enabling precise recovery of semantic details and edge information in feature maps via an adaptive kernel prediction mechanism, thereby addressing feature distortion during upsampling. In addition, a variant of wise intersection over union based on minimum point distance intersection over union is proposed WPIoU. The WPIoU loss function combines geometric constraints with a dynamic focusing mechanism to optimize the bounding box regression process, significantly improving the localization accuracy for targets with extreme aspect ratios and low-quality samples. Evaluations using the RDD2022 dataset demonstrate that the PCW-YOLO framework outperforms the traditional YOLOv8 architecture, achieving improvements of 2.1%, 1.1%, and 2.2% in precision, recall, and mAP, respectively. To verify the model’s generalization capabilities, subsequent evaluations were conducted using the SVRDD dataset. The results showed that the optimized network achieved a 5.2% increase in precision, a 1.4% increase in recall, and a 3.8% increase in mAP, thereby providing a more precise and efficient solution for road defect detection.