Handwriting CRNN

A deep learning project for recognizing handwritten text using a CRNN (Convolutional Recurrent Neural Network).

The goal is to build an end-to-end system that converts handwritten text images into machine-readable text. The final system will be capable of processing entire handwritten pages, detecting individual text lines, and recognizing them.

---

Project Status

🚧 Early development stage

Current progress:

• Environment setup complete
• CUDA-enabled PyTorch configured
• IAM dataset parsing and preprocessing complete
• dataset.csv generated with image paths and labels

---

Requirements

• Python 3.12
• PyTorch
• CUDA (optional but recommended for training)
• NVIDIA GPU (tested on RTX 4050)

---

Dataset

This project uses the IAM Handwriting Database.

U. Marti and H. Bunke "The IAM-database: an English sentence database for offline handwriting recognition" International Journal on Document Analysis and Recognition, 2002.

Dataset: http://www.fki.inf.unibe.ch/databases/iam-handwriting-database

The dataset is not included in this repository due to licensing restrictions.

Note: The IAM Handwriting Database is restricted to non-commercial research use. This repository only contains code and does not redistribute the dataset.

Expected Dataset Structure

data/
 └── raw/
      ├── words/
      └── words.txt

---

Data Pipeline

The IAM words dataset is processed into a structured CSV format for training.

Steps:

1. Parse words.txt
2. Filter valid samples (ok)
3. Reconstruct image file paths
4. Validate image existence
5. Generate dataset.csv

Output Format

image_path,label
data/raw/words/a01/a01-000u/a01-000u-00-00.png,A

This CSV serves as the input for the PyTorch Dataset class.

---

Setup

Create a virtual environment:

py -3.12 -m venv venv

Activate it:

venv\Scripts\Activate.ps1

Upgrade pip:

python -m pip install --upgrade pip

Install dependencies:

pip install -r requirements.txt

---

Quick Start

Verify that PyTorch and CUDA are correctly configured:

python -c "import torch; print(torch.cuda.is_available())"

Expected output:

True

---

Outputs

Processed dataset:

data/processed/dataset.csv

This file contains image paths and corresponding labels used for training.

---

Planned Architecture

The final OCR pipeline will follow this structure:

Page Image
   ↓
Line Segmentation
   ↓
CRNN Recognition
   ↓
Text Output

Recognition Model

CNN → BiLSTM → CTC Loss

• CNN extracts visual features from the image
• BiLSTM models sequential dependencies
• CTC Loss enables alignment-free sequence prediction

---

Project Structure

handwriting-crnn/
│
├── data/                # Dataset directory (not included in repo)
│   ├── raw/             # Original IAM dataset (words/, words.txt)
│   └── processed/       # Generated dataset (dataset.csv)
│
├── src/                 
│   └── dataset/         # Dataset handling
│       └── iam_dataset.py
│
├── prepare_dataset.py   # Script to generate dataset.csv from words.txt
├── test_dataset.py      # Temporary script to verify dataset loading
│
├── requirements.txt
├── requirements-min.txt
├── LICENSE
└── README.md

Roadmap

• Phase 0 — Environment Setup ✔
• Phase 1 — Dataset Preparation ✔
• Phase 2 — CRNN Model Implementation
• Phase 3 — Training Pipeline
• Phase 4 — Line Segmentation
• Phase 5 — Full Page OCR Pipeline

---

Future Work

• Implement CRNN architecture
• Build PyTorch Dataset and DataLoader
• Train model using CTC loss
• Add decoding (greedy / beam search)
• Develop line segmentation pipeline
• Create demo interface for handwritten text recognition

---

License

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