LIDet-Net: a framework for weld defect detection in low-light industrial environments

This study presents a novel LIDet-Net (low-light illumination defect detection network) framework for weld defect detection in low-light industrial environments. To address challenges of insufficient illumination, blurred defect boundaries, and small-target detection, the framework integrates a dynamic illumination enhancer Retinex-WeldNet with a multi-scale detector RE-YOLO. Retinex-WeldNet enhances brightness and suppresses noise based on the Retinex theory, while RE-YOLO employs a re-calibrated feature pyramid network (RC-FPN) to sharpen the boundary representation and improve small-target detection. Experimental results demonstrate that LIDet-Net achieves a mean average precision (mAP) of 91.1% on the low-light weld (LL-WELD) dataset, outperforming the baseline YOLOv11 model by 8.3%. Moreover, the framework demonstrates strong robustness and generalisation across steel and PCBs defect datasets, proving to be an efficient and accurate solution for industrial inspection under variable lighting.