Steel Surface Defect Detection Based on Denoising Diffusion Implicit Models with Data Augmentation

Due to the intricacy of steel production, surface defects unavoidably arise during the process. Detecting these defects is crucial for ensuring the production of high-quality steel. While some defect detection methods have been developed based on target detection models, simply applying object detection models without considering their efficiency often leads to subpar performance. This study addresses two key challenges in steel surface defect detection: insufficient data and detection efficiency. To tackle these challenges, we employed denoising diffusion implicit models (DDIMs) to augment the original dataset. Additionally, we introduced a novel method for defect detection in steel, termed YOLO-DSC, built upon YOLOv8 architecture. DDIMs were specifically leveraged to learn and train the publicly available NEU-DET dataset, generating reliable sample data. To enhance feature fusion, we integrated distribution shifting convolution into the bottleneck, effectively reducing the parameter count and computational load of YOLO-DSC. Experimental results conducted on NEU-DET demonstrate that YOLO-DSC exhibits robust performance and is well-suited for deployment on edge devices.