Deep-Learning-Based Probabilistic Graphical Models for Automated Defect Detection in Smart Textile Manufacturing

Guang Gao, Chuangchuang Chen

Article
2025 / Volume 8 / Pages 925-939
Received 8 August 2025; Accepted 28 August 2025; Published 10 December 2025
https://doi.org/10.31881/TLR.2025.925

Abstract
Smart textile manufacturing is challenged by unique and subtle defects that are difficult to detect using traditional inspection methods. To address these challenges, we propose a novel hybrid framework, Deep-CRF, that integrates the feature-extraction capabilities of deep learning (DL) with the contextual reasoning of probabilistic graphical models (PGMs) for high-accuracy automated defect detection. Our approach uses a ResNet-50 model, fine-tuned on a custom Smart Textile Defect Dataset (STDD), to produce an initial defect probability map. This map is then processed by a fully-connected conditional random field (CRF), which refines the segmentation mask by modeling spatial dependencies — thereby improving boundary accuracy and reducing prediction noise. Experimental results demonstrate the superiority of the Deep-CRF framework over traditional methods and standard DL models such as U-Net. Our model achieved a state-of-the-art mean intersection over union (mIoU) of 93.7% and an F1 score of 94.1%. The CRF refinement stage proved crucial, improving the mIoU by 3.5 percentage points over the ResNet-50 baseline. This work presents a robust and accurate solution that can significantly enhance quality control and pave the way for reliable, large-scale production of smart textiles.

Keywords
defect detection, deep learning (DL), convolutional neural networks (CNNs), probabilistic graphical models (PGMs), conditional random field (CRF)