<![CDATA[Communication is a fundamental right that often becomes a barrier for the deaf community when interacting with the general public. This limitation restricts equal access to education and information. In Indonesia, the Indonesian Sign Language System (SIBI) is the formal standard taught in Special Schools (SLB), yet the implementation of automatic translators for dynamic signs remains suboptimal. Previous studies have often focused on static signs, which are unable to capture temporal gestures. To address this, this study compares three spatio-temporal deep learning architectures for dynamic SIBI sign recognition. The three models compared include the landmark-based Stacked Bi-LSTM using MediaPipe, as well as the visual-based (appearance-based) ResNet50+Bi-LSTM and MobileNetV2+Bi-LSTM. This study utilizes a private dataset consisting of 989 videos covering 11 sign classes. The dataset was partitioned into 60% training, 20% validation, and 20% testing data using a stratified split. The landmark-based Stacked Bi-LSTM achieved the highest accuracy of 98.48%, outperforming ResNet50+Bi-LSTM (97.98%) and MobileNetV2+Bi-LSTM (95.96%). This model also proved to be the most efficient, with 19 times fewer parameters than ResNet50 and 2 times fewer than MobileNetV2, as well as an inference time 53 times faster than ResNet50. The Stacked Bi-LSTM is demonstrated to be the optimal architecture for SIBI recognition, offering the highest accuracy and best efficiency.]]>
Copyrights © 2026
