A Machine Learning Approaches for Barakhadi Recognition of Devnagari Sign Language

Abstract

Hearing impairments affect millions of individuals worldwide, including approximately 63 million people in India who depend on sign language for communication. However, a lack of awareness and understanding among the general population generates significant communication challenges. Recognizing Devnagari Sign Language (DSL), widely used in India, remains an underexplored area despite advancements in Sign Language Recognition (SLR). Key challenges in SLR include variations in lighting, diverse hand gestures, and complex backgrounds, which hinder effective real-time translation. This study introduces a system for recognizing the complete Barakhadi of DSL, comprising 408 characters. A robust dataset of static hand signs, captured in varied environmental conditions, is utilized for training. Multiple classifiers such as SVM, FFNN and Random Forest are leveraged and compared to identify the DSL Barakhadi characters. Customized RNN-LSTM model evaluated by employing diverse optimization strategies. The system's performance is measured using precision, recall, and F1-score metrics. The proposed RNN-LSTM model, optimized using the Adam algorithm, achieves an outstanding accuracy of 99.62%, demonstrating its effectiveness in DSL recognition.