Skip to content

Feature Request: Improve ASL Prediction Accuracy Using Data Augmentation Techniques #11

@jyothi1235

Description

@jyothi1235

Problem Statement

The current ASL recognition model performs gesture detection effectively in controlled environments, but prediction accuracy may decrease under real-world conditions such as:

  • Different lighting environments
  • Hand orientation variations
  • Camera distance changes
  • Background noise
  • Partial hand visibility
    This can lead to unstable or incorrect gesture predictions during real-time usage, reducing the reliability and usability of the system.

Proposed Solution

Introduce a robust data augmentation pipeline during model training to improve the generalization capability of the ASL recognition model. The goal is to expose the model to diverse variations of hand gestures so it can perform more accurately across different real-world scenarios.

Key Features

Advanced Data Augmentation

Apply augmentation techniques such as:

  • Rotation
  • Zoom
  • Horizontal shifting
  • Brightness adjustment
  • Flipping
  • Noise injection

Improved Generalization

Train the model on more diverse samples to reduce overfitting and improve real-time prediction performance.

Better Real-Time Stability

Reduce sudden prediction fluctuations caused by:

  • Hand movement
  • Lighting changes
  • Webcam quality variations

Training Performance Comparison

Add evaluation metrics to compare:

  • Accuracy before augmentation
  • Accuracy after augmentation
  • Validation loss improvements

Scalable Training Pipeline

Create a modular augmentation pipeline that can easily support future datasets and model upgrades.

Technical Suggestions

  • Use TensorFlow/Keras ImageDataGenerator or tf.image augmentation pipeline
  • Add preprocessing normalization
  • Introduce batch-wise augmentation during training
  • Compare model performance using confusion matrix and accuracy graphs
  • Maintain compatibility with the existing ASL recognition architecture

Benefits

  • Improves prediction accuracy
  • Enhances robustness in real-world usage
  • Reduces model overfitting
  • Creates smoother real-time gesture detection
  • Makes the project more production-ready

Future Scope

  • Dynamic gesture augmentation
  • Synthetic gesture generation
  • Adaptive training for different users
  • Low-light ASL recognition optimization
  • Mobile-device optimized inference

Contribution Interest

I would love to contribute to this feature under GSSoC'26 by working on:

  • Data augmentation pipeline
  • Model retraining and evaluation
  • Accuracy comparison experiments
  • Training visualization improvements
  • Real-time prediction stability enhancements

Looking forward to contributing to this impactful accessibility-focused project 🚀

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions