Deep learning-based sign language recognition system using both manual and non-manual components fusion

Sign language is regularly adopted by speech-impaired or deaf individuals to convey information; however, it necessitates substantial exertion to acquire either complete knowledge or skill. Sign language recognition (SLR) has the intention to close the gap between the users and the non-users of sign language by identifying signs from video speeches. This is a fundamental but arduous task as sign language is carried out with complex and often fast hand gestures and motions, facial expressions and impressionable body postures. Nevertheless, non-manual features are currently being examined since numerous signs have identical manual components but vary in non-manual components. To this end, we suggest a novel manual and non-manual SLR system (MNM-SLR) using a convolutional neural network (CNN) to get the benefits of multi-cue information towards a significant recognition rate. Specifically, we suggest a model for a deep convolutional, long short-term memory network that simultaneously exploits the non-manual features, which is summarized by utilizing the head pose, as well as a model of the embedded dynamics of manual features. Contrary to other frequent works that focused on depth cameras, multiple camera visuals and electrical gloves, we employed the use of RGB, which allows individuals to communicate with a deaf person through their personal devices. As a result, our framework achieves a high recognition rate with an accuracy of 90.12% on the SIGNUM dataset and 94.87% on RWTH-PHOENIX-Weather 2014 dataset.