CompletedPythonTensorFlowMediaPipe+2 more
ASL Recognition
Real-time bidirectional ASL alphabet recognition using BiLSTM-Attention network with confidence-gated decoding.
Timeline
2 months
Role
ML Researcher
Team
Solo
Status
CompletedTechnology Stack
Python
TensorFlow
MediaPipe
BiLSTM
OpenCV
Key Challenges
- Real-time inference
- Dataset curation
- Attention mechanism
- Confidence-gated decoding
Key Learnings
- BiLSTM architectures
- Attention networks
- MediaPipe integration
- Research paper writing
ASL Alphabet Recognition System
Overview
A real-time ASL fingerspelling recognition pipeline using hand landmarks and a BiLSTM + Attention model with confidence-gated decoding, published as a TechRxiv preprint.
Research Highlights
- Published: TechRxiv preprint (2026)
- Validation Accuracy: 99.4%
- Top-3 Accuracy: 100%
- Real-time Performance: ~30 FPS
Dataset
- Custom dataset of ~2,500 sequences
- Over 75,000 frames across 26 ASL classes
- Curated and annotated for research quality
Technical Architecture
- Feature Extraction: MediaPipe for 63-dimensional hand landmarks
- Model: 1.36M parameter BiLSTM-Attention network
- Decoding: Confidence-gated decoding for robust inference
Tech Stack
- Python for development
- TensorFlow for deep learning
- MediaPipe for hand tracking
- BiLSTM for sequence modeling
- OpenCV for video processing
Key Contributions
- Built real-time inference pipeline from webcam video
- Emphasized robustness for real-world deployment
- Created comprehensive research documentation