Implementasi Algoritma Hybrid CNN-ViT (LeViT-128) dalam Klasifikasi Motif Batik Indonesia

Abstract

Batik is an Indonesian cultural heritage featuring a wide variety of motifs with high inter-class visual similarity. The visual and manual identification process is highly subjective and inefficient. To address this classification problem, this study implements a deep learning approach using a hybrid Convolutional Neural Network (CNN) and Vision Transformer (ViT) architecture, specifically the LeViT model. This integration aims to synergize CNN's ability to extract local features—such as lines and dots—with ViT's self-attention mechanism to capture global spatial context efficiently. The research methodology follows the Cross-Industry Standard Process for Data Mining (CRISP-DM) framework. The model was trained on a dataset of 5,292 images covering 39 batik motif classes. To address data imbalance, class weight adjustments and image augmentation techniques were applied. Based on the final evaluation using the One-vs-Rest approach on the test set, the model achieved an overall accuracy of 90.05% and a Weighted F1-Score of 89.84%, demonstrating the effectiveness and reliability of the hybrid CNN-ViT algorithm in classifying batik motifs with high visual complexity.