<![CDATA[Indonesian Sign Language (BISINDO) is the primary communication medium for the deaf community. However, public understanding remains limited, and recognizing similar gestures and adapting to dynamic lighting pose significant challenges. This study applies the YOLOv13-Nano architecture with HyperACE and FullPAD modules for static BISINDO alphabet detection (A–Z). The model was trained on 5,835 images from the Mendeley Dataset (70:15:15 split) using MediaPipe Hands auto-labeling, and implemented on a FastAPI dashboard with debouncing and temporal auto-spacing logics. Offline testing shows YOLOv13-Nano achieved 89.22% classification accuracy, 83.02% mAP50, 90.63% Precision, 89.62% Recall, 89.74% F1-Score, and 23.62 ms inference time. This model outperforms K-Nearest Neighbors (KNN) baselines, where KNN (Full Image) achieved 9.90% accuracy and KNN (Hand Cropped) achieved 11.23% accuracy. Limited real-time testing on the test subject yielded a Word Accuracy of 87.4% under normal conditions, 88.3% under extreme low-light, and 90.4% alphabet accuracy against complex backgrounds. These results demonstrate the potential of YOLOv13-Nano to support a responsive real-time BISINDO translation system under various real-world environmental conditions.]]>
Copyrights © 2026
