Indian Sign Language Recognition using Prototypical Networks 👐

A state-of-the-art few-shot learning solution for Indian Sign Language (ISL) recognition that achieves 99.6% accuracy with only 15 samples per class. Bridges communication gaps for 63M+ hearing-impaired Indians through AI-powered gesture recognition.

Key Features ✨

• 🖐️ 5-Shot Learning: Recognizes new signs with just 5-15 examples using Prototypical Networks
• 🖼️ Advanced Preprocessing: Hybrid edge detection combining adaptive thresholding + Otsu's method
• 📊 Multi-Model Comparison: Benchmarks against CNN (99.98%), LSTM (98.19%), and Attention-CNN (99.67%)
• 🔄 Cross-Dataset Validation: Tested on two distinct ISL datasets from Kaggle

Datasets 📦

1. ISL Alphabet & Digits (35 classes)

https://www.kaggle.com/datasets/vaishnaviasonawane/indian-sign-language-dataset

2. ISL Words & Numbers (50 classes)

https://www.kaggle.com/datasets/princegupta0353/indian-sign-language-image-dataset

Models

1. CNN
2. CNN- with Attention
3. LSTM
4. SVM
5. Few-Shot Models
6. Prototypical Network
7. Matching Network

Results

Deep learning and Machine Learning models

Model	Accuracy	Training Data Required
CNN	99.98%	960 images/class
LSTM	98.19%	960 images/class
Attention-CNN	99.67%	960 images/class
SVM	99.90%	960 images/class

Few Shot Models

Model	Model Configuration	Accuracy	Training Data Required
Prototypical Network	5-Way 5-Shot (Dataset 1)	99.60%	15 images/class
Prototypical Network	5-Way 1-Shot (Dataset 1)	98.80%	11 images/class
Prototypical Network	5-Way 5-Shot (Dataset 2)	95.72%	15 images/class
Prototypical Network	5-Way 1-Shot (Dataset 2)	87.48%	11 images/class
Matching Network	5-Way 5-Shot (Dataset 1)	97.7 %	15 images/class