A Comparative Study of Supervised and Unsupervised Deep Learning Models for Fabric Defect Localization in Smart Textile Manufacturing

Abstract

The increasing demand for automated visual inspection in textile manufacturing has intensified the demand for strong and efficient defect detection systems. Traditional manual inspection methods are often inconsistent and resource-intensive, while the inherent variability of fabric textures creates obstacles for conventional machine vision techniques. To overcome these constraints, this study presents a comparative analysis of supervised semantic segmentation and unsupervised anomaly detection models for fabric defect localization using the ZJU-Leaper benchmark dataset. A total of nine deep learning architectures were evaluated across three methodological categories: (i) CNN-based supervised models (U-Net, PSPNet, FPN, DeepLabV3, U-Net++, and DeepLabV3+), (ii) Transformer-based supervised models (UPerNet and SegFormer), and (iii) an unsupervised anomaly detection model (EfficientAD). All models were trained on standardized preprocessed images and optimized under identical hyperparameter settings to ensure fair benchmarking. The evaluation followed the multi-level protocol defined in the ZJU-Leaper paper, incorporating sample-, pixel-, and region-level metrics, alongside efficiency indicators such as FLOPs, parameter count, and FPS to assess deployment feasibility. The results show distinct trade-offs between accuracy and efficiency across the examined architectures. Transformer-based models demonstrated superior generalization on complex textures, achieving balanced performance between global context capture and computational cost. CNN-based models excelled in fine-grained localization, offering high Dice and IoU scores at moderate inference speeds. Conversely, the unsupervised EfficientAD model achieved competitive detection accuracy without relying on labeled data, underscoring its potential for defect-scarce industrial scenarios. Overall, this research establishes a unified evaluation framework for benchmarking supervised and unsupervised deep learning models in fabric inspection. The results show into the architectural and operational considerations necessary for developing scalable, real-time, and label-efficient quality control systems in smart textile manufacturing environments.