---

🧾 ASL Detection – Image-Based Sign Language Recognition

Recognizing American Sign Language gestures from images using Convolutional Neural Networks (CNNs), data augmentation, and real-time webcam prediction.

---

📌 Table of Contents

• Overview
• Business Problem
• Dataset
• Tools & Technologies
• Project Structure
• Data Cleaning & Preparation
• Exploratory Data Analysis (EDA)
• Research Questions & Key Findings
• Dashboard
• How to Run This Project
• Final Recommendations
• Author & Contact

---

Overview

This project detects ASL gestures from image inputs using a deep learning pipeline built in Google Colab. It leverages CNNs and real-time webcam prediction to classify hand gestures with high accuracy. The workflow includes image preprocessing, model training, evaluation, and live prediction — all within a Colab notebook.

---

Business Problem

ASL recognition can empower inclusive communication for the hearing-impaired. This project aims to:

• Automate ASL gesture detection from webcam feeds
• Reduce reliance on manual interpretation
• Enable real-time prediction in browser-based environments
• Support accessibility in digital platforms

---

Dataset

• Image dataset uploaded to /content/data/ in Colab with folders per gesture (A–Z)
• Includes training and validation sets
• Augmented using rotation, flip, zoom, and brightness tuning
• Stored temporarily in Colab runtime or mounted from Google Drive

---

Tools & Technologies

• Python (TensorFlow, OpenCV, NumPy, Matplotlib)
• Google Colab (GPU acceleration, webcam access)
• GitHub (for version control and notebook hosting)
• Google Drive (optional dataset mounting)

---

Project Structure

asl_detection/
│
├── README.md
├── train_model.ipynb                # Main training notebook
├── realTime_prediction.ipynb        # Webcam prediction notebook
│
├── asl_dataset                      # Uploaded image dataset
│   ├── A/
│   ├── B/
│   └── ...Z/
│
├── models/                          # Saved models (.h5)
│   └── asl_model.h5
├── requirements.txt
├──asl_detection.pdf                 # summery pdf

---

Data Cleaning & Preparation

• Removed blurry and mislabeled gesture images
• Resized all images to 64x64 pixels
• Applied augmentation to balance gesture classes
• Encoded labels and split into train/val sets

---

Exploratory Data Analysis (EDA)

Class Distribution:

• 26 ASL gesture classes (A–Z)
• Balanced using augmentation for underrepresented gestures

Image Quality Checks:

• Verified resolution and gesture clarity
• Removed grayscale and low-contrast images

Sample Visuals:

• Grid plots of gesture samples
• Bar chart of image count per class

---

Research Questions & Key Findings

• Model Accuracy: Achieved 92.8% accuracy on validation set
• Misclassifications: Most confusion between M vs. N and U vs. V
• Augmentation Impact: Boosted accuracy by 6.3% on minority gestures
• Confidence Scores: Average prediction confidence = 0.89
• Real-Time Prediction: Webcam latency <1.2s per frame

---

Dashboard

• Real-time webcam interface via cv2.VideoCapture() in Colab
• Displays:
  • Live gesture prediction
  • Confidence score overlay
  • Frame-by-frame classification

---

How to Run This Project

1. Open the training notebook in Colab:
   🔗 ASL_Detection_Training.ipynb

2. Upload your dataset to /content/data/ or mount Google Drive:

from google.colab import drive
drive.mount('/content/drive')

3. Train the model:

# Run all cells in ASL_Detection_Training.ipynb

4. Run real-time webcam prediction:
   🔗 ASL_RealTime_Prediction.ipynb

---

Final Recommendations

• Expand dataset with dynamic hand gestures and backgrounds
• Integrate hand tracking for better localization
• Use Grad-CAM for gesture explainability
• Collaborate with accessibility-focused organizations for real-world testing

---

Author & Contact

Sushree Bandita Das

📧 Email: sushreebanditadas01@gmail.com
🔗 LinkedIn
🔗 Portfolio

---