YOLOv11-ODConv for Cross-Domain Micro-Scratch Detection in Photovoltaic Glass Manufacturing

Abstract

Cross-domain micro-scratch inspection is difficult in photovoltaic-glass manufacturing because the observable defect signal changes with coating type, furnace batch, camera response, illumination angle, and line contamination. This work develops a YOLOv11-ODConv detector that couples omni-dimensional dynamic convolution with domain-balanced feature learning and morphology-aware fusion. The model adapts convolutional kernels along spatial, input-channel, output-channel, and kernel-index dimensions, while a domain consistency objective constrains scratch representations without suppressing line-specific texture. A boundary-weighted localization term and hard-negative curriculum further separate faint elongated scratches from roller marks, dust trails, and specular streaks. Under leave-one-domain-out evaluation, the proposed model reaches 91.8% mAP50, 67.4% mAP50:95, 89.6% recall, and 12.7 false alarms per 1,000 images, improving the YOLOv11 baseline by 4.9, 5.7, and 4.1 percentage points in the first three metrics while reducing false alarms by 31.0%. It processes 1024 x 1024 images at 46.3 frames per second on an RTX 4060 and 21.8 frames per second on an industrial edge GPU. The results indicate that dynamic kernels are most useful when combined with domain-balanced supervision and morphology-aware localization, providing a practical route to robust surface inspection across manufacturing lines.