Devi Ajeng Efrilianda, Yahya Nur Ifriza, Satrio Fajar Utomo
The deaf community in Indonesia frequently uses Indonesian Sign Language (BISINDO), a visual language that combines hand gestures and facial emotions. However, the general public's ignorance of BISINDO continues to hinder communicated between the hearing and the deaf communities. This barrier not only hinders social interaction but also affect access to education, service, and employment opportunities for individuals with hearing impairments. To solve this problem, this paper suggests a classification method based on computer vision that recognizes the BISINDO alphabet using machine learning techniques. A custom dataset was developed, consisting of 2,600 images representing 26 alphabet gestures (A-Z), captured under controlled conditions using a mobile camera. Instead of using raw images alone, the system extracts 2D hand landmark coordinates using the MediaPipe Hands framework to enhance input quality and robustness. A Convolutional Neural Network (CNN) was trained on the extracted landmarks to classify each gesture. Furthermore, Bayesian Optimization was employed to fine-tune the model's hyperparameter and maximize accuracy. According to the results, the improved CNN model out baseline models that employed either raw pictures or untuned parameters, achieving a test accuracy of 99.61 %. These findings suggest that the integration of hand landmark extraction, CNN-based modeling, and Bayesian Optimization is highly effective for accurate and efficient BISINDO alphabet recognition. © 2026 IEEE.
Universitas Negeri, Department of Computer Science, Semarang, Indonesia
