Comparative Study of CNN, Vision Transformer, and Hybrid CNN–ViT Models for Indonesian Batik Pattern Classification

Batik is an Indonesian cultural heritage with unique visual characteristics and deep philosophical value. The complexity of motifs, color variations, and geometric details make batik classification an interesting challenge in the field of computer vision. This study conducted a comparative study between three deep learning approaches for classifying Indonesian batik motifs using Convolutional Neural Network (CNN), Vision Transformer (ViT), and a hybrid CNN–ViT model. The dataset used includes more than 3,000 batik images from various regions in Indonesia, with a variety of motifs such as Yogyakarta Kawung, Aceh, Ceplok, and Megamendung.Each model was trained with uniform parameters and augmentations to ensure fair evaluation, resulting in CNN accuracy of 94.43% F1-macro 93.45%, ViT accuracy of 91.55% F1-macro 89.78%, and Hybrid CNN-ViT accuracy of 94.04% F1-macro 92.91%. This is reinforced by the combination of modules (EfficientNet-B2 + CBAM + ArcFace) that can improve model performance furthermore, This study contributes to the development of an automated batik classification system and supports cultural preservation through artificial intelligence- based digitization.