<![CDATA[Batik Parang is a prominent traditional motif in Indonesia, characterised by repetitive diagonal patterns and subtle visual variations across regional styles, such as Solo Parang and Yogyakarta Parang, which pose challenges for automated image classification. This study addresses this challenge by introducing an optimisation-focused framework that integrates hyperparameter tuning strategies with a lightweight convolutional neural network, extending the practical use of MobileNetV2 for fine-grained cultural motif classification. A balanced dataset of 160 batik images collected from Kaggle was employed and partitioned using an 80:20 stratified split to ensure class consistency. The model was evaluated on a limited yet representative dataset reflecting realistic small-scale cultural heritage scenarios. Two hyperparameter tuning methods, Bayesian Optimisation and Particle Swarm Optimisation, were applied to optimise learning rate, batch size, and dropout rate, while two optimisers, Adam and Adagrad, were compared to analyse their effects on convergence stability and generalisation. The training process followed a two-phase strategy consisting of transfer learning and selective fine-tuning of upper MobileNetV2 layers. Experimental results indicate that Adagrad-based configurations consistently outperform Adam-based models, which exhibited class collapse and poor generalisation. The optimal configuration, combining Adagrad with Bayesian Optimisation, achieved a validation accuracy of 91% with balanced precision, recall, and F1-score across both Parang classes. These findings demonstrate that careful optimisation enhances the reliability of lightweight CNNs and support extending the proposed framework to other cultural heritage classification tasks and resource-constrained real-time applications.]]>
Copyrights © 2026
