Traditional Batik Pattern Recognition with MobileNetV2 and Sampling-Based Hyperparameter Optimization

Abstract

Batik holds significant cultural value in Indonesia, reflecting the nation's historical and artistic heritage through its intricate patterns. Preserving these designs is essential for maintaining cultural identity and supporting artistic and economic communities. With the advancement of technology, deep learning has emerged as an effective approach for recognizing and classifying batik patterns. Convolutional Neural Networks (CNNs), particularly MobileNetV2, are widely recognized for their efficiency and accuracy in image classification. However, the performance of deep learning models is highly influenced by hyperparameter selection, which remains a challenging task. This study investigates the effectiveness of MobileNetV2 in classifying traditional Indonesian batik motifs, including Kawung, Mega Mendung, Parang, and Truntum, by applying different hyperparameter optimization methods such as Treestructured Parzen Estimator (TPE), Gaussian Process Sampler (GPS), Grid Search, and Random Search. The experimental results show that TPE achieved the best overall performance with a test accuracy of 91.94% and an F1 score of 92.09%. GPS and Grid Search obtained identical test accuracy of 90.83% with F1 scores of 90.89% and 90.87%, respectively, while Random Search produced the lowest performance with an accuracy of 88.61% and F1 score of 88.61%. These findings highlight the importance of structured hyperparameter optimization, particularly TPE, in enhancing the robustness of CNN-based batik classification. The results provide valuable insights for the development of automated batik pattern recognition systems that support cultural heritage preservation and related image classification applications.