Sharable and unshareable within class multi view deep metric latent feature learning for video-based sign language recognition.

Suneetha, M.; Prasad, M. V. D.; Kishore, P. V. V.

Mobile based sign language recognition (SLR) is challenging in real time due to camera shudder and the signer movements for capturing continuous video data for recognition. Even though there are many state-of-the-art methods for SLR, they have ignored view sensitivity and its effects on the accuracy of the system. This work proposes a novel multi view deep metric feature learning (MVslDML) model for building a view sensitive environment into SLR, which is being investigated profoundly in human action recognition. The MVslDMLNet is an end-to-end trainable convolutional neural network where the features extracted from multiple views are learned based on the sharable and unshareable latent features within class multi view data through metric learning. Experiments performed on our multi view sign language and four benchmark action video datasets indicate a higher accuracy for the proposed framework.

SIGN language; CONVOLUTIONAL neural networks; IMAGE recognition (Computer vision); VIDEOS; HUMAN activity recognition; ARTIFICIAL neural networks

Multimedia Tools & Applications, 2022, Vol 81, Issue 19, p27247

1380-7501

Academic Journal

10.1007/s11042-022-12646-0