🧾 End-to-End Vietnamese Text Recognition (OCR)

---

🧠 Overview

This project presents an end-to-end OCR system specifically designed for Vietnamese scene text recognition. It processes the MC_OCR dataset and implements a two-stage architecture:

• Text Detection using PaddleOCR
• Text Recognition using a custom CNN + Transformer model built in PyTorch

---

✨ Key Features

• ✅ Two-Stage OCR Pipeline: separates detection from recognition for modularity and accuracy
• 🧠 Custom Recognition Model: CNN Encoder + Transformer Decoder
• 📦 Vietnamese Dataset Support: parses and trains on MC_OCR (polygon-based labels)
• 📊 Evaluation: uses Character Error Rate (CER) & Sequence Accuracy
• 🎯 Visual Output: draws bounding boxes and predicted text on original image

---

🧱 Model Architecture

🔹 Encoder – CNN

• Inspired by CRNN
• Converts text-line images into sequential visual features

🔹 Decoder – Transformer

• Uses self-attention and cross-attention
• Autoregressively generates characters one by one

---

📋 Steps

1. Open the notebook in Google Colab

2. Connect to GPU runtime

3. Mount Drive and install required libraries

4. Run the notebook sequentially:

   • 📊 Preprocessing: parses MC_OCR CSV with polygon annotations
   • 🧠 Model: defines CNN + Transformer OCRModel
   • 🔁 Training: saves best model based on validation CER
   • 🧪 Evaluation: computes final CER & Sequence Accuracy
   • 🖼️ Inference: visualizes prediction on sample images

---

📊 Evaluation Results

---