📝 Word Recognition System (Hybrid CRNN OCR)

• This project is a deep learning–based word recognition (OCR) system that combines multiple CRNN models into a hybrid inference pipeline to improve robustness across diverse visual conditions.
• Version v0.0.3 introduces a confidence-aware hybrid OCR strategy, integrating:
  • A baseline CRNN model (grayscale, variable-width)
  • A transfer learning CRNN model (VGG16-based, fixed-size RGB)
• The system dynamically selects the most reliable prediction at inference time, resulting in higher real-world accuracy without retraining.
• The application is deployed using a Flask web interface and trained on the Synth90k synthetic word dataset.

---

✅ What’s New (v0.0.3)

• ✅ Hybrid OCR pipeline (multi-model inference)
• ✅ Confidence-based model selection
• ✅ Improved robustness to:
  • Stylized fonts
  • Color backgrounds
  • Mixed casing
  • Slight rotations
• ✅ Refined CTC confidence estimation
• ✅ No changes required to UI or Flask logic
• ✅ Backward-compatible with previous models

---

🚀 Features

• Image-based single-word recognition
• Hybrid inference using two CRNN models
• Confidence-aware decision logic
• CTC-based sequence decoding
• Supports:
  • Grayscale & RGB inputs
  • Fixed-width and variable-width pipelines
• TensorFlow .keras production models
• Flask-based web interface
• Lightweight and modular codebase

---

🧠 Model Architecture

Hybrid Inference Overview

Input Word Image
        ↓
Preprocessing
  ├── Grayscale (Baseline CRNN)
  └── RGB Fixed Size (Transfer Learning CRNN)
        ↓
CRNN Models (parallel)
        ↓
CTC Decoding + Confidence Scoring
        ↓
Best Prediction Selection
        ↓
Final Recognized Word

---

🧩 Model Details

Baseline CRNN Model (v0.0.1)

• Input: Variable width, grayscale
• CNN + BiLSTM (CRNN)
• CTC decoding
• Strong on:
  • Simple fonts
  • Clean backgrounds
  • Short words

Transfer Learning CRNN Model (v0.0.2)

• Input: 32 × 256 × 3 (RGB)
• Backbone: VGG16 (ImageNet pretrained)
• BiLSTM × 2
• Strong on:
  • Stylized fonts
  • Color backgrounds
  • Rotated or complex images

---

🧪 Hybrid Decision Strategy

At inference time:

1. The baseline CRNN predicts first
2. A CTC confidence score is computed
3. If confidence ≥ threshold → accept result
4. Otherwise → fallback to VGG16-CRNN

This approach:

• Avoids overfitting to one model
• Preserves speed for easy cases
• Improves accuracy for difficult samples

---

🔠 Character Set

• Character Set:

  a–z, A–Z
  

• Case-sensitive recognition

• No language model or dictionary constraints

---

🏋️ Dataset

• Dataset Name: Synth90k (Synthetic Word Dataset)

• Images: 100,000 word images

• Labels: Stored in labels.txt

• Format:

  00000.jpg slinking
  00001.jpg REMODELERS
  00002.jpg Chronographs
  

• The dataset is downloaded using the Kaggle API, making it suitable for Google Colab.

---

🧰 Technologies Used

• Python
• TensorFlow / Keras
• VGG16 (Transfer Learning)
• BiLSTM (CRNN)
• CTC Decoding
• KerasCV – Data augmentation
• Flask – Web server
• HTML / CSS – Frontend UI
• NumPy
• Kaggle API – Dataset download
• Google Colab – Model training

---

📁 Project Structure

├── main.py                                             # Flask application
├── utils.py                                            # Preprocessing & CTC decoding
├── model/
│   ├── baseline_crnn.keras                             # Baseline CRNN model
│   └── transfer_learning_crnn.keras                    # VGG16-based CRNN model
├── notebook/
│   ├── training_pipeline_basline.ipynb                 # Baseline Colab Pipeline
│   └── training_pipeline_transfer_learning.ipynb       # Transfer Learning Colab Pipeline
├── templates/
│   └── index.html                                      # Web UI template
├── static/
│   └── uploads/                                        # Uploaded images
├── requirements.txt                                    # Dependencies
├── README.md                                           # Project documentation
├── .gitignore
└── LICENSE                                             # MIT License

---

⚙️ Installation & Setup

1. Clone the repository:

   git clone https://github.com/Kalana-S/Word-Recognition-System.git
   cd Word-Recognition-System
   

2. Install dependencies:

   pip install -r requirements.txt
   

3. Run the Flask application:

   python main.py
   

4. Access the Web UI:

   http://127.0.0.1:5000
   

---

🖼️ How It Works (Inference)

1. Upload a single-word image
2. Image is preprocessed for both models
3. Each model predicts independently
4. CTC decoding generates text
5. Confidence-aware selection chooses best result
6. Final word is displayed with model info

---

🧭 Versioning

Version	Description
v0.0.1	Baseline CRNN + CTC OCR
v0.0.2	VGG16 transfer learning CRNN
v0.0.3	Hybrid OCR with confidence-based selection

---

🎥 App Demo (Screen Recording)

Full app workflow — UI → Input → Prediction

screen.mp4

---

🤝 Contribution

Contributions are welcome.

• Fork the repository
• Create a feature branch
• Submit a pull request

---

📜 License

This project is licensed under the MIT License
See the LICENSE file for details.