A Lightweight Width Scaled Architecture Integrating Nested Feature Fusion for Bengali Sign Numerals Recognition

Sign language recognition is required for communication between the deaf-mute and normal populations. However, learning and communicating in sign language can be challenging for anyone who attempts it. The challenge of sign language recognition for different languages has been the subject of study for decades, but many systems that have been developed as a consequence have not proven to be particularly successful for Bengali yet. Bengali sign language is considered to be among the most challenging sign languages on account of its extensive range of words, character sets, and expressive methods. This paper introduces a lightweight deep neural network architecture that incorporates parallel convolutional operations for scaling the width and a nested ensemble approach. For enhancing the network’s speed, the architecture has featured additional dropout layers and ReLU activation function. The BdSL47 dataset, which is the most recently available Bengali sign language dataset, was incorporated to evaluate the performance of the proposed architecture. The experimental findings demonstrated that the proposed approach recognized Bengali Sign numerals with an accuracy of 99.25%, Precision of 99.28%, Recall of 99.25%, F1-score of 99.25%, and AUC score of 0.999. Our proposed nested feature fusion integrated width scaled lightweight architecture performed the best with fewer epochs and less computational time.