Classification of Indonesian Batik Motifs using CNN VGG16 with Transfer Learning and Fine-Tuning

Abstract

Indonesian batik is a cultural heritage passed down through generations and officially recognized by UNESCO. However, the identification of batik motifs still relies heavily on manual assessment by experts. This study aims to develop an automated batik motif classification system using a Convolutional Neural Network (CNN) based on the VGG16 architecture with a transfer learning approach. The dataset was obtained from the Indonesia Batik Motifs repository on Kaggle and consists of three batik motif classes: Batik Bali, Batik Kawung, and Batik Megamendung. To address class imbalance, data augmentation was applied to produce a balanced dataset of 1,500 images, with 500 images per class. The dataset was then divided into training, validation, and testing sets using an 80:10:10 ratio. The preprocessing stage included grayscale image conversion to reduce computational complexity. The proposed model, fine-tuned on the last eight layers of VGG16, achieved a test accuracy of 98.00% with an F1-score of 0.98. Among the three classes, Batik Megamendung achieved the highest F1-score (0.99), followed by Batik Kawung (0.98) and Batik Bali (0.97). Comparative experiments showed that the proposed VGG16 transfer learning model outperformed both MobileNetV2 with transfer learning (97.33% accuracy) and VGG16 trained from scratch (33.33% accuracy). This study contributes to the development of an accurate batik motif classification system that can be deployed in real time through a Streamlit-based web application.