MDGlass: Microscopic Defect Dataset for Glass with Polarized and Orthogonal Images

Abstract The glass defect dataset serves as the foundation for analyzing defects and applying artificial intelligence algorithms to classify them. The current glass defect datasets are very limited and cannot reflect the microscopic state of defects. To fill this gap, we built a microscopic glass defect dataset. This dataset comprises 4 types of defects, namely glassy state, surface spots, and so on, and can be categorized into a real subset and a generated subset. The real subset contains approximately 450 images collected directly from the production line. The defect sizes range from 0.1mm to 10mm, and the images are magnified 50-100 times using a Leica polarizing microscope under polarised and orthogonal illumination. As the real subset is small, we also built a generated subset with 1150 images using DcGAN, DDPM and Stable Diffusion algorithms. Furthermore, we evaluated the image quality of the dataset using 6 broad metrics, such as Contrast and NIQE, and established a classification baseline for the dataset by employing classical image classification algorithms.