/Jalali-Lab-Implementation-of-RAISR

Implementation of RAISR (Rapid and Accurate Image Super Resolution) algorithm in Python 3.x by Jalali Laboratory at UCLA. The implementation presented here achieved performance results that are comparable to that presented in Google's research paper (with less than ± 0.1 dB in PSNR). Just-in-time (JIT) compilation employing JIT numba is used to speed up the Python code. A very parallelized Python code employing multi-processing capabilities is used to speed up the testing process. The code has been tested on GNU/Linux and Mac OS X 10.13.2 platforms.

Primary LanguagePythonGNU General Public License v3.0GPL-3.0

Jalali-Lab Implementation of RAISR Algorithm

Contents

  1. Introduction
  2. Dependencies
  3. Training
  4. Testing
  5. Quantitative Comparison
  6. Sample Result
  7. Citations
  8. Acknowledgments
  9. License

Introduction

RAISR (Rapid and Accurate Image Super Resolution) is an image processing algorithm published by Google Research in 2016. With sufficient training data, consisting of low and high resolution image pairs, RAISR algorithm tries to learn a set of filters which can be applied to an input image that is not in the training set, to produce a higher resolution version of it. The source code released here is the Jalali-Lab implementation of the RAISR algorithm written in Python 3.x by Sifeng He, under the guidance of Prof. Bahram Jalali. The implementation presented here achieved performance results that are comparable to that presented in Google's research paper (with less than ± 0.1 dB in PSNR). Just-in-time (JIT) compilation employing JIT numba is used to speed up the Python code. A very parallelized Python code employing multi-processing capabilities is used to speed up the testing process. The code has been tested on GNU/Linux and Mac OS X 10.13.2 platforms.

Dependencies

All the dependent Python modules needed for the using our RAISR implementation can be installed using pip package manager and are the following:

  • NumPy
  • Numba
  • Python Imaging Library (PIL)
  • scipy
  • os
  • pickle
  • skimage

Additionally, a requirements.txt file is provided for installing all the dependencies by running the following command in a shell:

pip install -U -r requirements.txt

Training

All the training images are to be stored in the trainingData directory, before executing the Train.py script. In the training script, modify the upscaling factor, R, appropriately. The training outputs three files, filter, Qfactor_str, and Qfactor_coh, which are needed in the testing phase. The training data used in this implementation is the BSD 200 (Martin et al. ICCV 2001). A pre-trained filter with upscaling factors x2, x3, and x4 are available for testing in the Filter directory. To execute training, simply run the following command in your shell:

python Train.py

Testing

All the test images are to be stored in the testData directory, before executing the Test.py script. In the test script, change the upscaling factor, R, and filters. The result image will be saved in results directory. The test data used for the development of the algorithm were Set 5 (Bevilacqua et al. BMVC 2012) and Set 14 (Zeyde et al. LNCS 2010). To execute testing, simply run the following command in your shell:

python Test.py

Quantitative Comparison

Upscale Factor Google RAISR Our implementation trained
on BSD 200 images		 Our implementation trained
on COCO (500 images)
2x 35.913 35.855 35.878
3x 32.061 31.981 32.019
4x 29.689 29.717 29.801
Table 1. Comparison of Results on Set 5 Images
Upscale Factor Google RAISR Our implementation trained
on BSD 200 images		 Our implementation trained
on COCO (500 images)
2x 31.980 31.790 31.816
3x 28.764 31.714 28.729
4x 26.912 26.901 26.935
Table 2. Comparison of Results Results on Set 14 Images

Google's RAISR implementation was trained on 10000 advertising banner images. Our implementation of RAISR was trained on BSD 200 (200 images) and on 500 images from COCO dataset. The training time for our implementation on a 3.4 GHz 6-core Xeon desktop computer is about 20 minutes for BSD 200 dataset and an hour for the 500 images from COCO dataset. The average test time for one single image in test dataset (Set 5 and Set 14) is less than 1 second.

Sample Result

A high resolution image after downscaling it by 2 was used as the input low resolution image to both bicubic interpolation and our implementation of RAISR algorithm. A comparison of the resulted output image for this testcase with bicubic interpolation and our implementation of RAISR is shown in the figure above.

Citations

Please cite the following publications if you plan to use the code or the results for your research:

  1. Sifeng He, Bahram Jalali. "Brain MRI Image Super Resolution using Phase Stretch Transform and Transfer Learning", arXiv preprint arXiv:1807.11643, (2018). URL

Bibtex:

@article{he2018brain,
  title={Brain MRI Image Super Resolution using Phase Stretch Transform and Transfer Learning},
  author={He, Sifeng and Jalali, Bahram},
  journal={arXiv preprint arXiv:1807.11643},
  year={2018}
}

Acknowledgments

Special thanks to Dr. Cejo Konuparamban Lonappan for his support and guidance.

License

The code is released under GNU General Public License 3.0. Please refer to the LICENSE file for details.