RAW-FABRID: RAW FABRic Image Dataset for defect detection
Description
RAW-FABRID (RAW FABRic Image Dataset) is a comprehensive collection of high-resolution industrial textile images designed to develop, test, and benchmark anomaly detection and semantic segmentation algorithms. Sourced from diverse manufacturers to ensure realistic variability in texture and weaving, this dataset supports robust research in automated visual inspection. DATA ACQUISITION & FORMAT Images were acquired using a custom-built machine equipped with a Basler line-scan camera (4 px/mm) and dual LED illumination. All images are provided as 8-bit grayscale PNGs. DATASET ORGANIZATION The data is structured to support both full-resolution analysis and patch-based benchmarking: 1. images directory (709 images) Contains the high-resolution original images. To ensure data quality, the raw sensor captures were cropped (starting at coordinate X=120, with a width of 1792 pixels) to remove non-fabric dark borders and uneven peripheral illumination. All images have a resolution of 1792×1024 pixels with a single grayscale channel. This folder includes both defect-free and defective samples. 2. masks directory (204 masks) Contains the pixel-level ground truth binary masks corresponding to the defective images. The mask files share the exact same names as their corresponding defective images in the images folder. Masks use two intensity levels: 0 for non-defective background areas and 255 for defective regions. 3. MVTec directory To facilitate seamless integration into existing Artificial Intelligence (AI) anomaly detection frameworks, the high-resolution images were cropped into 256×256-pixel patches. These patches are organized strictly following the standard MVTec Anomaly Detection benchmark directory structure: - train/good (14,196): Contains only non-defective patches intended for model training. - test/good (4,969): Contains non-defective patches for evaluation. - test/defect (687): Contains defective patches for evaluation. - ground_truth/defect (687): Stores the corresponding 256×256 binary masks for the test/defect patches, maintaining identical filenames. METADATA & ANNOTATIONS (COCO & CSV) To ensure full data traceability and reproducibility, four auxiliary files are included in the root directory: Metadata (CSV): RAW_FABRID_HighRes_Metadata.csv and RAW_FABRID_Patches_Metadata.csv provide structured information (filename, fabric origin, defect area). The patches metadata explicitly maps each 256x256 patch back to its high-resolution source image. Annotations (JSON): RAW_FABRID_HighRes_COCO.json and RAW_FABRID_Patches_COCO.json provide precise bounding box and polygon annotations for all defects, strictly formatted as COCO-compatible JSON. USE CASES Ideal for training, evaluating, and comparing traditional computer vision and Artificial Intelligence (Deep Learning) algorithms for the textile industry.
Files
Institutions
- Universitat Politecnica de ValenciaValenciana, Valencia