Classification of Batik Keraton motifs using ResNet-50 fine-tuning architecture
Fine-tuning a ResNet-50 to classify four Keraton batik motif classes, reaching 97.8% test accuracy. Published in AITI volume 23.
AITI · Vol. 23, No. 1 · pp. 46–60 · 2026
- 97.8%
- test accuracy
- 96.8%
- validation accuracy
- 4
- motif classes
- ResNet-50
- fine-tuned
Abstract
Batik Keraton motifs carry distinct visual patterns that are difficult to separate with hand-crafted features, especially when the dataset is small and class balance is uneven. This study evaluates transfer learning as a practical approach for classifying four Keraton motif classes.
Method
A ResNet-50 backbone pre-trained on ImageNet was fine-tuned on a curated dataset of Keraton batik images. The training pipeline used augmentation and gradual unfreezing to adapt the deeper layers without overfitting the limited samples. Performance was measured on a held-out test split alongside validation accuracy.
Results
The fine-tuned model reached 97.8% test accuracy and 96.8% validation accuracy across the four motif classes. The gap between the two suggests the model generalised rather than memorising the training set, though the small number of classes remains a limitation.
Takeaways
Fine-tuning a strong vision backbone is a realistic baseline for cultural heritage image classification. The paper documents the preprocessing and training choices so the result can be reproduced on a larger motif set.