A denim fabric defect dataset for closed-loop YOLOv8 detection (holes, missing yarn, oil stains, abrasion marks).

A purpose-built dataset of 1,506 annotated denim fabric images (1920×1080, colour) for real-time defect detection, accompanying the paper "Detect–Decide–Act: Coupling Real-Time Defect Detection to Physical Actuation in Denim Inspection" (Md. Nazimus Sakib, Chaon Ronjon Kormokar, Nafis Ahmad; submitted to the Journal of The Textile Institute, 2026). The dataset covers four industrially relevant defect classes — Holes, Abrasion_Mark, Oil_Stain, and Missing_Yarn — annotated in YOLO format (one .txt per image: class_id, normalised x_center, y_center, width, height). It contains approximately 1,988 annotated instances, split into 1,206 training and 300 validation images (≈80:20) at the source-image level so that augmented variants of a given source remain in the same subset. Denim samples were manually modified to reproduce the morphology of the four classes, then imaged on a laboratory roller rig under constant LED illumination. An initial 307 source captures were expanded by offline augmentation (180° rotation, horizontal and vertical flips, brightness increase/decrease); blurred frames and duplicates were removed to yield the final 1,506 images. Annotation was performed in CVAT with tight bounding boxes. Note on scope: all defects were induced under controlled laboratory conditions on a single denim substrate. Performance on genuine, morphologically diverse factory defects and other fabric types remains to be validated.