Fabric Anomaly Recognition Dataset

This dataset has been specifically developed to address critical challenges in fabric defect detection within the textile industry of Bangladesh. Recognizing the importance of maintaining high-quality standards in textile exports, minimizing material waste, and ensuring customer satisfaction, this dataset offers a comprehensive, real-world collection of fabric defects captured directly on production floors in Bangladeshi textile manufacturing units.
The dataset focuses on three principal defect types that are most prevalent and industrially significant in textile quality control:
• Hole (Class ID-0): Physical punctures or missing material resulting from mechanical damage during weaving, finishing, or handling stages. These defects undermine both the fabric’s structural integrity and its visual appeal, often leading to rejection of the product.
• Broken End (Class ID-1): Breakage of warp yarns during weaving or finishing, causing visible gaps or thin lines on the fabric surface. These occur due to mechanical failures, excessive yarn tension, or handling errors, affecting fabric durability.
• Oil Stains (Class ID-2): Surface discolorations and blemishes caused by oil or lubricant leaks from machinery during production. Such stains can alter fabric appearance and interfere with subsequent dyeing or finishing operations.
The dataset consists of a total of 2,157 images, carefully curated to maintain near-class balance and reduce bias during machine learning model development. The class distribution is:
• Hole: 649 images
• Broken End: 666 images
• Oil Stains: 842 images
This balanced representation ensures that models trained on this dataset can detect defects equitably across all classes, enhancing overall performance.
All images were collected in situ on factory floors across various textile units in Bangladesh, capturing defects in authentic production environments. Images exhibit a wide range of illumination conditions and background variations, reflecting real operational challenges encountered in industrial settings.
Annotation was performed using LabelImg, a widely adopted open-source graphical tool for object detection labeling. Expert annotators manually drew bounding boxes tightly enclosing each defect instance, providing precise spatial localization critical for training and validation of deep learning models. Each annotated object is labeled with a numerical class ID consistent with widely used detection model conventions, such as YOLO:
• Hole → 0
• Broken End → 1
• Oil Stains → 2
The labeling process involved rigorous quality control, including multiple review cycles and cross-validation by different annotators, to minimize subjective errors and ensure annotation accuracy. This meticulous annotation effort is essential to prevent error propagation during model training, which could otherwise degrade detection performance.