Comparative Analysis of Model Architectures Using Transfer Learning Approach in Convolutional Neural Networks for Traditional Ulos Fabric Classification

Ulos cloth is a traditional woven fabric of the Batak tribe in North Sumatra, valued for its aesthetic and symbolic significance in various ceremonies. The diversity of ulos motifs presents challenges in preservation due to their unique patterns and functions. This study aims to develop an accurate method for classifying ulos motifs using Transfer Learning on Convolutional Neural Network (CNN) architectures. Five popular models—VGG16, VGG19, MobileNetV3, Inception-V3, and EfficientNetV2—were evaluated on a dataset of 962 ulos images across six motif categories.The results show that Inception-V3 outperformed other models with an average validation accuracy of 98.13% and the lowest loss of 5.67%. Inception-V3 also demonstrated superior generalization, achieving the highest K-fold validation accuracy, while VGG16 and VGG19 exhibited overfitting at higher learning rates. Two-way ANOVA analysis confirmed significant performance differences among the models and highlighted the interaction between model type and training methods. This research recommends Inception-V3 as the optimal model for ulos motif classification, offering an efficient and reliable tool to support cultural preservation through advanced image recognition technology.