An Improved YOLOv8-Based Method for Detecting Appearance Defects in Hull Closure Welds

This paper addresses the challenges of low efficiency and poor reliability in manual inspection of weld defects during ship manufacturing, proposing an improved YOLOv8-based detection method for hull closure weld appearance defects. By constructing a weld dataset containing five typical defects including cracks and pores, we innovatively integrate a Spatial and Channel Reconstruction Convolution (ScConv) module into the YOLOv8 backbone network. The spatial reconstruction unit and channel reconstruction unit effectively reduce feature redundancy while enhancing feature representation capability for micro-defects. Experimental results demonstrate that the improved model achieves 87% mAP50 on our proprietary dataset, with accuracy and recall rates reaching 87.8% and 86.4% respectively. Field deployment tests confirm a detection speed of 10.4s/m, showing 98.6% improvement in efficiency compared to traditional manual inspection (20.7s/m), thereby meeting real-time requirements. This method provides an automated solution for ship weld quality inspection with significant engineering application value.