<![CDATA[Recognizing wayang characters remains difficult because many classes share similar visual traits, capture conditions vary, and labeled data is scarce. This study develops a web-based wayang character recognition system that combines transfer learning with visual explanation for classroom use. Two lightweight backbones, EfficientNetV2-S and DeiT-Tiny, were fine-tuned on ten characters using horizontal flipping, RandAugment, and TrivialAugment, then evaluated through five-fold stratified cross-validation. Performance was measured using accuracy, macro-F1, and the Matthews Correlation Coefficient. EfficientNetV2-S reached a mean accuracy of 86.8%, a macro-F1 of 0.87, and an MCC of 0.85, slightly above DeiT-Tiny at 86.0%, 0.86, and 0.83. Confusion analysis identified recurring errors between visually similar pairs such as Nakula-Sadewa and Gareng-Petruk. Grad-CAM overlays were integrated into a Streamlit interface so that every prediction carries visual evidence that non-experts can inspect. Median inference latency on a mid-range CPU was 28 ms for EfficientNetV2-S and 32 ms for DeiT-Tiny, fast enough for interactive use. The system shows that accuracy, speed, and interpretability can be balanced within the constraints of cultural heritage datasets.]]>
Copyrights © 2026
