INTELLIGENT IMAGE PROCESSING AND MACHINE LEARNING FOR AUTOMATED FABRIC IDENTIFICATION AND QUALITY INSPECTION: A CRITICAL REVIEW

Abstract

Automated visual inspection has become a critical part of intelligent textile manufacturing, yet existing literature is quite scattered in terms of fabric material recognition, classification of weaves/texture, defect detection, localization and real-time implementation. The earliest methods used texture and color descriptors constructed by hand, while more recent research moved toward the use of convolutional neural networks, transfer learning, multiscale architecture, object detector networks, segmentation, attention-based networks and multimodal representations. The current review focuses on this evolution and not the sameness of all computer vision approaches. In particular, it addresses the problem of separating two related problems: recognition of what a piece of fabric is and recognition of if its surface is of acceptable quality. Woven fabrics' recognition research demonstrates the benefits of learned visual representations, while material identification literature shows the relevance of small-sample and lightweight models. Surface defect detection literature evolves from unsupervised image reconstruction and compact CNNs to increasingly advanced real-time detectors and segmentation algorithms. Multimodal literature suggests that texture, local microstructure and complementing surface descriptors may help in material identification. However, despite significant advances in the field, there are still several limitations: lack of datasets, unbalanced classes, poor control of fabric roll leakage, restricted laboratory conditions, poor consistency in performance measures, lack of cross-dataset evaluation and ignoring of uncertainty/explainability issues. An open direction for future research suggested by this review is a multimodal and robust framework of inspection of fabric identity, texture/weave, surface defects and quality.