Abstract:
The present paper is a research project report creating a real-time Bangla Sign Language
(BdSL) recognition system in a two-phase approach. The initial phase concentrated on
developing a state-of-the-art baseline with our proposed hybrid Conv1D-BiLSTM
architecture including the Attention mechanism with an accuracy of 97.3% on the
published 60-word BDSLw60 v2 dataset, which is higher than the previously published
benchmarks. The second stage of research was devoted to the important issue of
vocabulary development, and a new set of signs of 39 signs, which had been collected by
hand, was presented. In order to address the dramatic accuracy drop experienced with
first fine-tuning, we designed and adopted a professional quality multi-stage fine-tuning
policy that includes class weights and gradual unfreezing. This sophisticated procedure
was very practical and the final and strong model reached 93 percent total accuracy in
the whole vocabulary (99-word). This last model is implemented on a real-time Flask
web application, which translates signs directly to text, showing an end-to-end cycle
between a validation of the baseline to a scalable and real-world application.
