Klasifikasi Motif Batik Nusantara Menggunakan Transfer Learning EfficientNetB0

Abstract

This study develops an image classification system to recognize 20 Nusantara batik motifs using EfficientNetB0 transfer learning. The research problem arises from the difficulty of distinguishing visually similar batik motifs under variations in color, lighting, camera quality, and shooting angle. The local dataset was derived from the Kaggle Batik Nusantara dataset and contains 800 images: 560 training images, 80 validation images, and 160 independent test images. Validation was used for model selection, while final generalization was assessed on the independent test set. The method covers dataset preparation, 224 x 224 image preprocessing, numerical augmentation, two-stage training, quantitative evaluation, error analysis, and Flask web integration. The model was trained through feature extraction and fine-tuning using categorical crossentropy with 0.1 label smoothing, Adam optimizer, ModelCheckpoint, EarlyStopping, and ReduceLROnPlateau. The experiment produced 81.25% validation accuracy, while the independent test set produced 78.75% accuracy, 0.8170 macro precision, 0.7875 macro recall, 0.7843 macro F1-score, and 1.1609 test loss. These findings indicate that EfficientNetB0 is a feasible transfer learning baseline for batik motif classification on a small dataset, with only a modest drop when evaluated on unseen data.