Deep models and optimizers for Indian sign language recognition

Abstract

Deep Learning has attracted the research community's attention for a long time, and still, new deep models come into the picture very frequently. It is challenging to know and select the best amongst such models available in the literature. Also, selecting optimizers and tuning optimization hyperparameters is a trivial task. Thus, in this paper, we carry out a performance analysis of two pre-trained deep models, four adaptive gradient-based optimizers, and the tuning of hyperparameters associated with them on a static Indian sign language dataset. Experimental results found InceptionResNetV2 and Adam optimizer to have the potential of being used for static sign language recognition using transfer learning technique. Inception-ResNetV2 model highly outperformed the state-of-the-art machine learning approaches and hand-crafted features with an accuracy of 94.42% and 85.65% on numerals and alphabets of Indian sign language, respectively.