Steel Surface Detection Based on Conditional Diffusion Model and Dual-Decoder Architecture

In manufacturing industries, the accurate detection of steel surface defects is critical for ensuring product quality and safety. However, supervised learning-based defect detection systems face significant challenges due to data scarcity and class imbalance in industrial environments. In this paper, we propose a pipeline that combines diffusion model-based data augmentation with pseudo-labeling and a dual-decoder architecture. Our approach first utilizes a finetuned conditional diffusion model to generate synthetic defect images with diverse characteristics and backgrounds. To obtain supervision for these unlabeled synthetic images, they are processed through a pretrained segmentation model to create pseudo-labels. These pseudo-labels are then combined with original labeled data to form an augmented training dataset. A dual-decoder segmentation network is trained on this augmented dataset, performing both multi-class and binary segmentation tasks simultaneously. Experimental results on the Magnetic Tile dataset demonstrate the effectiveness of our method, achieving significant improvements of 2.46 % in mIoU compared to baseline approaches, validating the practical applicability in real industrial environments.