Text Sharpener

This repository contains the implementation of a Convolutional Neural Network (CNN) model to denoise and deblur text images, improving their readability and clarity. The model extends the U-Net architecture and is trained on synthetic blurry images, optimizing a combined objective of reconstruction and edge loss.

Table of Contents

1. Overview
2. Dataset
3. Architecture
4. Results
5. Usage

Overview

Blurry text in images poses significant challenges for readability and information extraction, often caused by noise, motion blur, or poor lighting conditions. This project aims to address these issues by developing a CNN-based solution that significantly enhances text image quality.

Key features include:

• Development of a modified U-Net architecture tailored for text deblurring tasks
• Implemented a combined objective of reconstruction and edge loss
• Improved Peak Signal-to-Noise Ratio (PSNR) on the testing set by 62%, from 16.97 dB to 27.21 dB.

Dataset

The dataset was sourced from the textbook Pattern Recognition and Machine Learning (2006) by Christopher Bishop. Motion blur was randomly applied to generate input and target pairs.

Architecture

The model builds upon the U-Net architecture with modifications to enhance performance:

• Residual connections added to double convolution blocks for better gradient flow.
• Average pooling used instead of max pooling in downsampling layers to retain context.
• Padding applied to convolutional layers to prevent spatial size reduction.
• Global skip connection directly from input to output for residual learning.

Results

Improved from 16.97 dB to 27.21 dB on the test set.

Usage

Data Preparation

src/data/collect.py: Collect text images from PDFs.

src/data/blur.py: Generate synthetic blurry images.

Models

UNetV2.py: Best-performing model located in src/model. Other attempted models like BasicNet, BetterNet, and SharpNet are also included.

Training

src/notebooks/train.ipynb: Notebook to train models. Ran on Google Colab with T4 GPU.

Evaluation

src/notebooks/evaluation.ipynb: Notebook to evaluate and visualize model performance.

Pretrained Model

Pretrained UNetV2 model available in src/ckpt/unetv2_1000. Test it with:

import torch
from src.model.UNetV2 import UNetV2

ckpt = torch.load("src/ckpt/unetv2_1000", map_location=torch.device('cpu'))
model = UNetV2(in_channels=1, out_channels=1, features=[64, 128])
model.load_state_dict(ckpt["model_state_dict"])
model.eval()

(back to top)