Defect Detection Performance Improvement Method Using Conditional Diffusion Model and Pseudo-label based Semi-supervised Learning

Defects in steel products pose significant risks to safety and reliability in the steel industry. Although supervised deep learning has been implemented for defect inspection, its reliance on large annotated defect samples for training remains time-consuming and costly, limiting its practicality and generalization to diverse real-world scenarios. To address this challenge, we propose a semi-supervised framework that augments steel surface defect images using a conditional latent diffusion model and leverages pseudo-labels generated by a pre-trained segmenter to train a defect segmentation model. Using the public dataset NEU-seg, we augment data and evaluate the segmentation performance of our method for defects and defect regions against conventional geometric image augmentation techniques. Experimental results demonstrate the proposed method’s superior segmentation accuracy for both defect types and regions compared to comparative approaches, as measured by mean intersection over union (mIoU) and mean Dice coefficient (mDice).