A Fully Convolutional Neural Network for Fast Detection, Classification, and Segmentation of Fabric Defects

Abstract

The integration of artificial intelligence and image processing for fabric defect detection is gaining prominence due to its practical significance in enhancing production quality. This study proposes a fast and accurate convolutional neural network (CNN) designed to detect defects in fabric with minimal computational complexity. The model processes input images of size 256 × 256 and generates defect masks of size 64 × 64. To improve detection accuracy, the model incorporates techniques such as ResNet, scheduled learning rate policies, data augmentation, and a weighted cross-entropy loss function. Trained on a diverse dataset of 2,681 defect samples from four fabric types and defect classes (holes, oil stains, color stains, and roller marks), the model achieved an accuracy of over 96%, a loss value below 0.1, and high recall, precision, and F1-Score. Compared to other state-of-the-art models, the proposed model delivers competitive performance with significantly faster prediction times, making it suitable for real-world fabric inspection applications. © The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2025.