VAD-MambaYOLO: Improvement and application of MambaYOLO for weld detection on metal roofing systems

Abstract The automated inspection of welds on metal roofing systems at large-scale infrastructures, such as airport terminals, is essential for long-term structural health monitoring. However, achieving a practical balance between high detection precision and computational efficiency remains challenging due to the complex visual characteristics of weld defects, including scale variation and low contrast under reflective metal surfaces. To address this problem, this study proposes VAD-MambaYOLO, a high-precision detection model that performs multi-dimensional optimizations on the MambaYOLO baseline architecture. It integrates an improved VisionClueMerge module, which combines multi-scale convolutions with attention mechanisms to strengthen feature extraction and contextual awareness. Additionally, a dynamic upsampling operator is introduced to adaptively refine feature reconstruction, simultaneously enhancing localization accuracy and processing efficiency. Evaluated on a real-world dataset of metal roof welds, VAD-MambaYOLO achieves a precision of 0.882, recall of 0.853, mAP50 of 0.892, and mAP50-95 of 0.432. Compared with various YOLO variants and other detectors, the proposed model achieves a significant improvement in detection accuracy while maintaining a favorable computational footprint, offering a robust and efficient solution for automated weld quality assurance in metal roofing construction.