Optimizing Convolutional Neural Networks for Real-Time Fabric Defect Detection

In pursuing flawless textile production, fabric defect detection has emerged as a critical challenge demanding sophisticated solutions. This study proposes a classification-based approach using convolutional neural networks (CNNs) to address the challenge of identifying fabric defects. We evaluated the performance of four CNN architectures - ResNet50, VGG16, MobileNetV2, and InceptionV3, on nine different categories of fabric defects. Our methodology involved the strategic removal of the top layers from each model, followed by their initialization with pre-trained weights and the incorporation of a custom output layer tailored for multi-class classification. The empirical results underscore the efficacy of our approach, with ResNet50 leading the pack by achieving an impressive accuracy of 99.39%. This performance notably exceeds VGG16 and MobileNetV2, which attained 98% accuracy, and InceptionV3, which recorded 93%. These findings demonstrate the superior capability of ResNet50 in fabric defect detection and highlight the potential of deep learning in revolutionizing quality control processes within the textile industry. The proposed method promises significant improvements in the early identification and rectification of fabric defects, thereby contributing to enhanced product quality, reduced waste, and increased operational efficiency in textile manufacturing. As the industry seeks higher standards, integrating such advanced CNN-based solutions could become a cornerstone of future quality assurance practices.