Automated Textile Quality Control Using Deep Learning: Multi-Class Classification of Fabric Surface Defects

Abstract

Fabric defects directly influence the surface appearance and overall quality of textile products, which makes their reliable detection and classification an essential component of quality control in the textile industry. This study applies deep learning, specifically transfer learning with a pretrained convolutional neural network, to automatically recognize six fabric image categories: vertical defects, defect-free fabric, holes, horizontal defects, lines, and stains. A dataset of 2,739 fabric images was divided into 1,369 images for training and 1,370 images for validation. Prior to model input, all images were converted to three-channel colour format, resized, centre-cropped to 224 × 224 pixels, and normalized. An 18-layer residual neural network (ResNet18), pretrained on a large image corpus, was fine-tuned for five epochs using the Adam optimizer with a learning rate of 0.001 and a batch size of 32. Performance was evaluated on the validation partition, which was also reused for final testing, yielding an overall accuracy of 91.17%, a macro-averaged F1 score of 81.10%, and a support-weighted F1 score of 91.12%. Classification results differed noticeably across defect categories: defect-free images and stains achieved F1 scores of 98.12% and 92.27%, respectively, whereas holes proved more difficult to recognize, reaching an F1 score of 68.31% and a recall of 60.58%. These outcomes demonstrate that a fine-tuned deep learning model can effectively distinguish multiple categories of fabric surface condition within the evaluated dataset, while also exposing uneven recognition performance for certain defect types, particularly those with subtle or irregular visual patterns. By linking established textile defect categories with quantitative deep learning evaluation metrics, this work offers a preliminary baseline for automated, CNN-driven fabric quality assessment. The approach is relevant to textile manufacturing organizations, including those in Russia, seeking to integrate intelligent visual inspection systems into their production and quality assurance workflows.