A CNN-Based Information System for Balinese Dance Classification with Hyperparameter Optimization

Balinese dance classification presents challenges due to limited datasets, complex postures, and the lack of real-world implementation. Existing studies often focus on model development while overlooking deployment aspects. This research proposes a lightweight Convolutional Neural Network (CNN) designed for Balinese dance classification and compares it with MobileNetV2, ResNet50, and VGG16 using consistent training settings. Data augmentation was applied to enhance generalization, and training epochs were optimized based on model convergence. The proposed CNN achieved a validation accuracy of 99.00%, with a precision of 92.55%, recall of 89.88%, and F1-score of 91.1%, using only 590 thousand trainable parameters and the fastest inference time of 476 milliseconds. Although others pretrained model, MobileNetV2 slightly outperformed in some metrics, the proposed model offered a better tradeoff between performance and efficiency. The trained model was deployed in a web-based application, demonstrating practical usability. This work supports the preservation of Balinese dance through accessible and efficient AI integration.