Dynamic Thai Sign Language Recognition using Recurrent Neural Network

Sign Language is a communication using a hand gesture that can pose on head to waist along with a facial emotion. There are numerous articles attempting to recognize dynamic sign language using machine learning. However, the dynamic sign language is a temporal continuous data. In addition, the positions of the hands and facial emotion are components that contribute to the completeness of sign language communication. Therefore, a sign language recognition methodology development is still challenge. This research aims to develop a Thai sign language recognition approach using recurrence neural network (RNN). The MediaPipe library applies to landmark extraction consisting of the hands, face and posture using by coordinate (x, y, z) totally 1,662 keypoint for RNN input. After that, these keypoints are learned by RNN technique consisting of long short-term Memory (LSTM) 2) gated recurrent unit (GRU) and 3) bi-direction LSTM (BiLSTM). The dataset consists of 10 words of Thai sign Language totally 1,000 videos that are established by volunteers sign language interpreters and hearing impaired. The experiment result demonstrates that an accuracy of the proposed method at 99% by LSTM and GRU.