6TH INTERNATIONAL CONFERENCE ON APPLIED ENGINEERING AND NATURAL SCIENCES ICAENS 2024, Konya, Türkiye, 25 Eylül - 26 Ekim 2024, ss.560-565
This study addresses the recognition of American
Sign Language (ASL) using a real-time system built on the MUD dataset, which
includes 36 sign classes representing letters and digits. The system employs
complex Zernike moments for feature extraction and utilizes a complex-valued
deep neural network to classify hand signs captured via webcam. In the study,
the Mediapipe library, which identifies key points on the hand, was used to
detect the hand and to ensure consistent feature extraction, the hand was framed
in a square. Initial results demonstrate a recognition accuracy between 90% and
95%, though challenges remain, particularly in distinguishing similar signs.
The study emphasizes the impact of background and lighting on performance,
suggesting that further optimization and training with lower-resolution images
could enhance speed and efficiency, ultimately supporting mobile applications.
The findings indicate that effective real-time sign language recognition
systems require robust models and careful environmental control.