DeepDeblur: Text Image Recovery from Blur to Sharp
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.

DeepDeblur: Text Image Recovery from Blur to Sharp

This work focuses on recovering the blurry text image.


Based on the deep neural network, a new short connection scheme is used. Trained by the pixel regression, higher visual quality of the image can be recovered by the network from the blurry one.

  • Sequential Highway Connection (SHC) structure


  • Loss curve comparing with the ResNet Structure:


  • Visualization



Created by:

Jianhan Mei at Nanyang Technological University

Xiang Chen at Darmstadt University of Technology

Ziming Wu at The Hong Kong University of Science and Technology

Required Environment

  • Ubuntu 16.04

  • Python 2/3, in case you need the sufficient scientific computing packages, we recommend you to install anaconda.

  • Tensorflow >= 1.5.0

  • Keras >= 2.2.0

  • Optional: if you need GPUs acceleration, please install CUDA that the version requires >= 9.0

Running The Demo

  • Text Image Dataset Generation

    Check the matlab script 'Matlab/RunProcess.m': The path of the text images should contain raw sharp text images. You can build your own dataset by convert PDF files into raw image files and save them to the text image path in "Matlab/RunProcess.m".

    Then run the matlab script 'Matlab/RunProcess.m', which helps to build the training dataset.

  • Training

    Check the training data and model saving paths in "", for which the training data should be consistent with the previous step. Then run the following script:

  • Testing

    Check the testing model and data paths in "". Then run the following script:



This work is released under the MIT License (refer to the LICENSE file for details).