Batik is a significant Indonesian cultural heritage with a vast diversity of motifs, making manual classification a challenging task. This research provides a comparative analysis of two prominent deep learning architectures, the Convolutional Neural Network (CNN), represented by VGG16, and the Vision Transformer (ViT), represented by DeiT, for the classification of Tulungagung batik images. A balanced dataset of 2,400 images, comprising two classes (Bangoan and Majanan), was utilized. The experiment was conducted using three distinct training-to-testing split ratios (80:20, 70:30, and 60:40) to evaluate model robustness. Performance was assessed using accuracy, precision, recall, F1-score, and the confusion matrix. The results indicate that the CNN (VGG16) model consistently outperformed the ViT (DeiT), achieving its peak accuracy of 96% on both the 80:20 and 60:40 split ratios, showcasing high stability. The ViT (DeiT) model was more sensitive to the data split, reaching a peak accuracy of 94% with less consistent performance. We conclude that for this specific classification task, the VGG16 architecture is more robust, stable, and effective than the DeiT architecture.
Mochammad et al. (Thu,) studied this question.