Surface defect dataset for sanitary ceramics
Description
This dataset consists of high-resolution RGB images of ceramic sanitary ware products, including both defective and defect-free samples. It was designed for training and evaluating surface defect detection and segmentation models.
Files
Steps to reproduce
The dataset was constructed using 106 high-resolution images (67 defective, 39 non-defective) of ceramic sanitary ware products collected from the production line of a leading manufacturer. Instead of using time-consuming polygon-based manual labeling, a weak annotation strategy was employed. Defect regions were painted manually using solid green (RGB: 0,255,0) over the original images. These painted images were then processed to generate binary masks by thresholding the green pixels. Both the painted and original images underwent identical augmentation via 90°, 180°, and 270° rotations and were resized to 512×512 pixels. This process yielded 18,560 image patches in total, which were split into training, validation, and test sets.
Institutions
- Eskisehir Teknik Universitesi