deblur

This is an unofficial Python implementation of the 2014 IEEE paper "Deblurring Text Images via L0 Regularized Intensity and Gradient Prior" by J. Pan, Z. Hu, Z. Su, and M. Yang

We also provide extra functionality to deblur images with a Kaggle Blur Dataset

Project info

This is implemented as the final project for CS1430 at Brown.

Contributing members:

Alan Gu,

Edward Xing,

Luca Fonstad,

Paul Zhiyuan Zhou

Setting up

python3 -m venv env
source env/bin/activate
pip3 install -r requirements.txt

Running

deblur images

python3 -m deblur.motion_deblur [--data ieee] [--save] [--regularizatio n] [--threshold]
# use either `ieee` or `kaggle` for the --data option to run different datasets to deblur
# use --save to save the deblurring results to disk
# use --regularization to change the regularization used for artifact removal
# use --threshold only on grayscale images to enhance results

run python3 -m deblur.motion_deblur --help for more information

evaluate the PSNR of deblurring results (this may take quite a while)

python3 -m deblur.eval

Results

We achieve comparable PSNR performance as the IEEE paper:

Furthermore, we also propose a image pixel intensity threshold method that achieves better PSNR results for grayscale text images: use the --threshold option to run the motion_deblur.py to try it out!

File structure

data/: the data directory

assets/: the assets directory holding images for README

resources/: data storage for artifacts removal functionality

deblur/: the source code directory, in which we have:

conjugate_gradient.py: our implementation of the conjugate gradient algorithm as an optimization scheme

eval.py: the script to evaluate results

get_data.py: data preprocessing and parsing

motion_deblur.py: the main magic and algorithms to deblur images

params.py: storage for hyperparameters

visualize.py: visualization methods