Klasifikasi Jenis Tari Bali Menggunakan Hyperparameter Tuning CNN dan Transfer Learning ResNet18

Abstract

Balinese dance is a cultural heritage that carries deep philosophical and historical values. In the field of computer vision, image classification of Balinese dance poses a unique challenge due to similarities in movement patterns, costumes, and backgrounds. This research compares two approaches to Balinese dance image classification: a Convolutional Neural Network (CNN) model enhanced with hyperparameter tuning via grid search, and a transfer learning model based on ResNet18. The dataset consists of seven dance classes, each with approximately 240 to 254 images, which are balanced to ensure fair evaluation. The CNN model's hyperparameters, including learning rate, dropout rate, batch size, and optimizer, were optimized using grid search, achieving a top training accuracy of 96.51% and validation accuracy of 72.30%. Meanwhile, the ResNet18 model, leveraging transfer learning from ImageNet, outperformed with a training accuracy of perfect 100% and a validation accuracy of 96.79%. The experimental results suggest that transfer learning significantly boosts performance compared to CNNs trained from scratch, even when carefully tuned. These findings highlight the practical advantage of leveraging pre-trained models in cultural heritage preservation tasks through computer vision.