Automated recognition of Thai fabric patterns using transfer learning with ResNet-50

Abstract

Traditional Thai fabric patterns are valuable cultural heritage, but expert knowledge of these patterns is declining. This study uses a convolutional neural network (CNN) with transfer learning (ResNet-50) to classify traditional Thai fabric patterns automatically. We collected 961 images across 19 pattern categories, split into training (57%), validation (11%), and test (32%) sets. Using ImageNet pre-trained weights and progressive fine-tuning, the model achieved 99.68% test accuracy with macro-averaged F1-score of 99.35%. Ablation studies validated our approach: augmentation improved accuracy by 2.26%, fine-tuning outperformed frozen backbone by 4.84%, and backbone comparison showed ResNet-50 achieves higher F1-score than MobileNetV2 while MobileNetV2 offers 90% parameter reduction for mobile deployment. Controlled stress tests demonstrated robustness under image degradations typical of smartphone photography. Unlike previous studies using controlled conditions, our dataset includes real-world variations. These results show that transfer learning with small datasets can match expert-level pattern recognition for cultural heritage preservation.