Penerapan Teknik Augmentasi Geometrik dan Edge Detection pada Pre-Processing Dataset Bahasa Isyarat Indonesia (BISINDO) untuk Mengatasi Variasi Gerakan Real-Time dalam Aplikasi ElCue

Abstract

Advancements in communication technology, such as video calls, have significantly improved social interactions. However, they remain inaccessible to individuals with hearing impairments. Indonesian Sign Language (BISINDO), the primary communication method for the Deaf community, presents challenges for automated recognition, particularly due to variations in hand gestures, differences in viewing angles, and inconsistent lighting conditions. One approach to addressing these challenges is by implementing pre-processing techniques to enhance data diversity and improve model robustness in gesture recognition during video calls. This research applies pre-processing techniques using Geometric Augmentation (rotation, translation, scaling, shearing, and flipping) and Edge Detection with Gaussian Blur, Adaptive Thresholding, and Otsu’s Thresholding to increase dataset variation while simultaneously enhancing the clarity of hand contours and boundaries. The results show an expansion of the dataset to 380,800 images, preserving the fundamental structure of hand gestures. A Convolutional Neural Network (CNN) model trained on the pre-processed dataset achieved a validation accuracy of 83.77% and a test accuracy of 83.88%, confirming that the applied pre-processing methods significantly enhance classification accuracy. Consequently, the integration of Edge Detection and Geometric Augmentation in BISINDO pre-processing plays a crucial role in improving model performance and holds strong potential for implementation in real-time sign language recognition systems within the ElCue video call application.