<![CDATA[This research addresses the computational optimization of convolutional neural network (CNN) architectures for the classification of Indonesian Sign Language System (Sistem Isyarat Bahasa Indonesia, SIBI) static alphabet imagery to enhance digital communication accessibility. Utilizing a domain-specific dataset comprising 1,165 images across 26 alphabet classes, this study tackles the prominent challenges of limited sample sizes and severe class imbalance. We evaluate five state-of-the-art CNN architectures MobileNetV2, DenseNet121, Xception, InceptionV3, and ResNet50V2 under four distinct training data paradigms before and after adaptive fine-tuning. To eliminate predictive bias without pixel-level distortion, oversampling is operationalized via Latent Space SMOTE on flattened vector embeddings, combined with dynamistic runtime image augmentation. The experimental results reveal that MobileNetV2, when optimized through partial layer-freezing (locking 150 baseline layers) under the integrated augmentation and oversampling combination scenario, achieved the highest macro-classification accuracy of 98.30%. This architecture also demonstrated superior efficiency, reducing the computational training latency to 0.53 minutes. The findings underscore the strategic advantage of leveraging optimized lightweight networks like MobileNetV2 for domain-specific visual recognition tasks.]]>
Copyrights © 2026
