This research aims to develop an Indonesian Sign Language (BISINDO) alphabet detection system using the YOLOv5 algorithm, an efficient and fast deep learning-based object detection model. The dataset used consists of BISINDO alphabet images enriched through data augmentation techniques such as rotation, flipping, and brightness adjustment. The evaluation results show that the YOLOv5s model achieved very good performance, with an average precision of 85.2%, recall of 89.3%, F1-score of 87.2%, and mean average precision (mAP) of 87.1%. The confusion matrix also indicates the model's ability to differentiate each BISINDO alphabet with high accuracy. The training data testing showed the model successfully achieved consistent decreases in all loss components, such as a decrease in train box loss from 0.06 to 0.015, and validation loss converging towards 0.002 for object loss and class loss. The real time testing also shows that the YOLOv5-based BISINDO alphabet detection system can perform well and consistently, indicating the practical application potential of this system to facilitate communication between people with hearing/speech disabilities and the general public. Overall, this research has resulted in an accurate and realtime implementable BISINDO alphabet recognition system.
Read full abstract