complete overview of the project

September 4, 2023

Overview

SENYAS is a real-time Filipino Sign Language (FSL) recognition system developed using MediaPipe and CNN-LSTM architectures. This application aims to bridge communication gaps between the Filipino deaf community and non-signers by recognizing hand gestures in FSL and converting them to text.

Key Features

  • Real-time Recognition: Utilizes computer vision and deep learning to convert FSL gestures from video input into text output.
  • MediaPipe: Extracts essential landmarks for hand and pose gestures, simplifying input complexity.
  • CNN-LSTM Architecture: Combines Convolutional Neural Networks (CNNs) for spatial feature extraction and Long Short-Term Memory (LSTM) networks for temporal dependencies in hand gestures.
  • Performance: The best-performing model, CNN-MP-LSTM, achieved 98% training accuracy and a 0.92 score on key performance metrics such as precision, recall, and F1-score.

Model Comparisons

The study evaluated five different CNN-LSTM models, comparing their performance on:

  1. Prediction Speed: CNN-MP-LSTM averaged 377.08 microseconds per prediction.
  2. Accuracy: Achieved weighted accuracy of 86% in real-world scenarios.
  3. Metrics: Precision, recall, and F1-score consistently exceeded 0.90.

Dataset and Training

  • A custom dataset was developed, including 10 dynamic and 5 static FSL gestures.
  • Data augmentation techniques, such as horizontal flipping and zooming, expanded the dataset for training.
  • Landmarks were extracted using MediaPipe, creating numpy arrays for model input.

Application Deployment

  • Web-based: Hosted on Vercel, making the application accessible via web browsers.
  • TensorFlow.js Integration: Allows for real-time gesture recognition on mobile and desktop devices.

Conclusion

SENYAS offers a viable alternative to human interpreters, fostering inclusivity for the Filipino deaf community by enabling real-time sign language recognition.