YOLO-3DCA: Lightweight Road Defect Detection with Multi-Directional Stereo Convolution

Road defect detection is crucial for ensuring traffic safety and reducing maintenance costs. However, existing deep learning models struggle with multi-scale target recognition and background interference in complex road scenarios. To address these challenges, this paper proposes YOLO-3DCA, a lightweight framework for road defect detection. First, a Multi-Directional Stereo Convolution (MDSC) module is designed by integrating dilated, diagonal, and cross convolutions to expand the receptive field, enhancing feature extraction for irregular cracks and multi-scale potholes. Second, the Convolutional Block Attention Module (CBAM) is introduced to dynamically focus on defect regions through dual channel-spatial attention, suppressing background interference. The YOLOv5 architecture is optimized with Multi-Directional Stereo Convolution and Convolutional Block Attention Module, achieving efficient collaboration between shallow details and deep semantics. Experiments on the GRDDC2020 dataset demonstrate that the proposed method outperforms mainstream models, providing a high-precision, cost-effective solution for road maintenance.