Recognition of Bangla Sign Language Characters and Digits Using CNN

Sign language is an elemental way of communication in the mute and hearing-impaired community. Bangla Sign Language (BSL) is the conversation medium for Bengali deaf people. In this paper, a BSL dataset containing 2660 images is presented where 2160 belong to 36 classes of letters and 500 belong to 10 digit classes. A deep learning based system can be a great replacement for human interpreters and motion sensing devices to ensure a cheaper, rapid and accurate BSL recognition. Convolution Neural Network (CNN) can effectively extract features from images and generalize unseen images. Transfer learning involving InceptionV3, VGG19, InceptionResnetV2 and ResNetl52V2 pretrained architectures has been conducted in the experiments. Insignificant image preprocessing was required to train these models and minimal overfitting could be ensured by thoroughly tuning the hyperparameters. The validation accuracy for the best model (ResNetl52V2) is 98.5% when trained on 12k augmented images.