<![CDATA[This study is motivated by the importance of automatic batik motif identification to support cultural preservation and the adoption of intelligent technologies in the creative industry. Although numerous previous studies have applied Convolutional Neural Networks (CNNs) for batik classification, most of them are limited to the use of a single architecture or do not provide a comprehensive performance comparison among modern models. As a result, a research gap remains regarding the selection of the most effective architecture for handling the high complexity of batik motifs. The main challenge lies in the wide variation and intricate patterns of batik designs, which pose significant difficulties for convolutional models in achieving accurate classification. Therefore, this study aims to comparatively evaluate the performance of three transfer learning architectures—ResNet50, MobileNetV2, and MobileNetV3-Large—in classifying 20 Indonesian batik motifs. The research methodology includes dataset collection from Kaggle, image preprocessing, data augmentation, model training for 30 epochs, and performance evaluation using precision, recall, and validation loss metrics. The experimental results indicate that MobileNetV3-Large achieves the best performance, with a training accuracy of 95.67%, a validation accuracy of 96.60%, and a validation loss of 10.17%. ResNet50 ranks second with stable performance, while MobileNetV2 exhibits the lowest accuracy due to its limited capability in capturing complex motif details. Consequently, this study contributes to addressing the identified research gap by providing a more comprehensive comparison of transfer learning architectures and demonstrates that modern Neural Architecture Search (NAS)-based models can significantly improve batik motif classification accuracy. The findings further confirm that MobileNetV3-Large is the most suitable architecture for developing efficient and accurate deep learning–based batik motif identification systems.]]>
Copyrights © 2026
