Sign Language Interpretation in Live Streams Using MediaPipe with Hybrid Deep Learning Model

This work explores the development and comparative analysis of deep learning models for hand gesture recognition using a labeled dataset of hand images. The primary objective is to enhance recognition accuracy by leveraging Media-Pipe for feature extraction alongside various neural network architectures. The proposed work employs Media-Pipe for precise land mark detection, followed by a CNN for feature extraction and bidirectional-long short-term memory (Bi-LSTM) for temporal dynamics. It is trained and validated on augmented data with performance metrics such as accuracy and loss tracked over 100 epochs. This study achieved the highest validation accuracy of 97.24% and has the capability to recognize the alphabet, which has similar hand gestures.