Comparative Evaluation of Hybrid Deep Learning Models for BISINDO Alphabetic Video

摘要

Accurately recognizing BISINDO (Bahasa Isyarat Indonesia) alphabetic gestures from video data presents significant challenges due to variations in pose, lighting conditions, and background noise. While previous studies have explored individual deep learning models, such as CNNs or LSTMs, the comparative evaluation of hybrid architectures that combine spatial and temporal feature extraction remains limited, especially in video-based sign language recognition. This study proposes a hybrid deep learning approach that integrates ResNet50 for spatial feature extraction with CNN and LSTM architectures for temporal sequence modeling, aiming to enhance the robustness and accuracy of BISINDO gesture classification. Compared to conventional CNN-based models, our hybrid architecture demonstrated an improvement of 3.4% in F1-score and over 17.6% in precision on challenging gestures. This systematic comparative evaluation reveals the superior capability of hybrid models to generalize in complex environments. Sample videos used in this study contain various backgrounds and signer styles to reflect real-world conditions. The findings contribute to developing more reliable sign language recognition systems and provide insights for future research and practical applications in real-time gesture recognition.