MOD-YOLO: Improved YOLOv5 Based on Multi-softmax and Omni-Dimensional Dynamic Convolution for Multi-label Bridge Defect Detection

Xinyi He, Ping Ma, Yiyang Chen, Yuan Liu

Abstract: Object detection methods provide efficient and accurate solutions for industrial production and quality control in defect detection. Since the overlap of bridge defect categories often occurs simultaneously, it is difficult for target detection methods targeting a single label to achieve accurate bridge defect detection. This paper proposes a bridge defect detection scheme YOLOv5 based on multi-softmax and omni-dimensional dynamic convolution (MOD-YOLO), which combines the proposed multi-softmax classification loss function with omni-dimensional dynamic convolution (ODConv). MOD-YOLO is evaluated on codebrim dataset and achieves the highest performance compared to existing SOTA models such as YOLO series and transformer-based series.