YOLO-Defect: A New Multi-Scale YOLO-Based Deep Neural Network with Feature Enhancement for Multiclass Bridge Surface Defect Detection

Bridges continuously face traffic pressure and environmental influences, leading to surface defects that need accurate detection for optimized maintenance and sustainability. The appearance and shape of bridge surface defects vary greatly, and multiple overlapping defects can occur in small areas. Current automatic detection methods also struggle with detection accuracy due to these complexities. Most studies focus on specific defect types and cannot fully assess a bridge's condition. We present YOLO-Defect, a YOLOv5-based multi-scale deep neural network designed to detect various complex bridge surface defects. It features two innovative modules: the global depthwise spatial pyramid pooling fusion module, which integrates multi-scale defect information, and the feature enhancement module, which enhances the differentiation between defects and background elements. YOLO-Defect is evaluated on the ZJU SYG crack and CODEBRIM datasets, outperforming benchmarks like Faster R-CNN, SSD, RetinaNet, and YOLOv5. On the ZJU SYG crack dataset, it achieves 78.2% mAP@0.5 and 56.9% mAP@0.5:0.95, while on the CODEBRIM dataset, it achieves 39.5% mAP@0.5 and 17.8% mAP@0.5:0.95.