An Automated Framework for Fabric Defect Detection in Textile Inspection

Abstract

Fabric quality is of paramount importance in the textile industry. As a result, defect detection in fabric is critical for quality assurance to rigorous standards. Traditional methods of fabric inspection can suffer from human error and be quite labor-intensive. Therefore, further automated systems are needed to support this process. This paper describes a novel deep learning, and machine vision framework using Detectron2, state-of-the-art computer vision package. The framework is based on the Mask R-CNN model for automatic fabric defect detection. The architecture uses a ResNeXt-101 backbone, Feature Pyramid Networks (FPN), and Region of Interest (RoI) Align to improve defect detection for multiple textures, and sizes of fabric. The model achieves excellent performance, with a mean average precision (mAP) between 93% and 97%, near 97% recall rate, and F1-scores of between 95%- 96%. This automated framework not only improves the fabric inspection process but also reduces human error. Therefore, reducing costs and providing product quality assurance. By incorporating deep learning into fabric defect detection, this research represents a significant advancement towards improving efficiencies, and quality control, across the textile industry.