FIVORA:Dataset for Fabric Classification and Defect Detection

FIVORA: A Dual-Phase Dataset for Automated Textile Inspection The FIVORA dataset is formulated for automated textile inspection research and is divided into two primary subsets: Fabric Classification and Fabric Defect Detection. 1. Fabric Types Classification Dataset (Fabric_Types.zip) Total Data (Images): 3,601 images in total. Number of Classes: 9 distinct fabric classes (Chiffon, Cotton, Denim, Linen, Polyester, Silk, Velvet, Viscose, Wool). Device Collected: Captured using a Samsung Galaxy M21 smartphone (4GB RAM, 64GB internal storage). File Type: Primarily.jpg format, with some images in .jpeg format specifically within the Denim and Silk classes Folder Structure: Organized into 9 class-named root folders (Chiffon, Cotton, Denim, Linen, Polyester, Silk, Velvet, Viscose, Wool), each directly containing the respective image files. Data Distribution per Class: Chiffon: 400 Cotton: 400 Denim: 401 Linen: 400 Polyester: 400 Silk: 400 Velvet: 400 Viscose: 400 Wool: 400 Intended Use: Designed for training and evaluating image classification deep learning models (such as EfficientNet, MobileNetV2, and ResNet50) to accurately identify and categorize different textile materials based on their visual texture and weaving patterns. 2. Fabric Defect Detection Dataset (Fabric_Defects.zip) Total Data (Images): 2,938 images in total (2,058 Training, 587 Validation, and 293 Testing). Number of Classes: 7 distinct defect categories (Holes, Lines, Objects, Oil Spot, Pinched Fabric, Stain, Thread Error). Device Collected: Captured using a Xiaomi Redmi 13 smartphone running Android 16, equipped with a 108 MP main camera and 256 GB internal storage. Annotation Tool: Roboflow (Bounding box annotations manually assigned to localize defects). File Types: Images: .jpg format Annotations: .txt format (YOLO bounding box information) Configuration: data.yaml Folder Structure: The root directory contains 7 class-named folders. Each class folder is divided into train, valid, and test subdirectories. Each subset is further organized into images/ (containing .jpg files) and labels/ (containing .txt files). A data.yaml configuration file is included to define paths and classes. Data Distribution per Class: Holes: 349 Train | 99 Valid | 50 Test Lines: 302 Train | 87 Valid | 43 Test Objects: 298 Train | 85 Valid | 42 Test Oil Spot: 270 Train | 77 Valid | 38 Test Pinched Fabric: 280 Train | 80 Valid | 40 Test Stain: 281 Train | 80 Valid | 40 Test Thread Error: 278 Train | 79 Valid | 40 Test Intended Use: Specifically structured with precise bounding-box annotations for training, validating, and evaluating object detection models (including YOLO architectures and Real-Time Detection Transformer Large - RT-DETR-L) for automated fabric defect localization. The fabric samples used to create the dataset were collected from four locations in the Central and Western Provinces of Sri Lanka. Three different fabric vendors in Kandy, in the Central Province, and one garment factory in Veyangoda, in the Western Province, were the sources of the fabric samples.