Sparse-Coding-based-Image-Super-Resolution-with-CUDA

By Hanlin Tan, NUDT

This project implements the image super resolution algorithm of the paper LASSO Approximation and Application to Image Super-resolution with CUDA Acceleration using matlab.

This project benefits from Yang's paper Image super-resolution via sparse representation and code.

Hardware Requirements

You need a Nvidia GPU with at least 3GB RAM to run CUDA version of the proposed image super-resolution algorithm.

Software Requirements

You need to intall:

matlab 2014a or later.
CUDA driver.
CUDA 8.0 SDK. Update mex_CUDA_*.xml if you use a different version.
Visual Studio 2013, for Windows OS. Update mex_CUDA_win64.xml if you use a different version.
XCode 8.0, for Mac OS X. Update mex_CUDA_maci64.xml if you use a different version.
GCC, for Linux OS.

Compile mex CUDA program

You can use pre-compiled mex files, or compile srCuda.cu using matlab command

mex -v srCuda.cu

The above command will generate

srCuda.mexw64 on Windows
srCuda.mexmaci64 on Mac OS X
srCuda.mexa64 on Linux

using compiler configurations stored in mex_CUDA_win64.xml, mex_CUDA_maci64.xml and mex_CUDA_glnxa64.xml,respectively.

Note if you install different versions of CUDA or native C++ compiler, you need to update the xml files with correct version information.

Notes for Linux

On Linux, you may need to install Matlab 2017a to run the code with CUDA 8.0. Otherwise, you may encounter an error:

Can't reload '/usr/local/MATLAB/R2016a/bin/glnxa64/libmwgpu.so'

Another choice is to install CUDA 7.5 and Matlab 2016a.

Replicate the results in the paper

To replicate the results, including figures, tables and result images in our paper:

To evaluate comparable algorithms, run compareAlgorithms.m.
To replicate the figure and tables in the paper, run analyze.m.

Train your own dictionaries

The algorithms use a pair of sparse dictionaries in Dictionary folder. You can train your own dictionaries by cd to the train folder and run Demo_Dictionary_Training.m.

Our algorithm also relies on pre-computed sparse coefficents of base vectors stored in Children_sparse_coe.mat, which can be updated by calling Children_SC(D), where D is the trained low resolution dictionary.

Reference

If you use the code provided by the project, please cite our paper:

@inproceedings{tan2017lasso,
  title={LASSO approximation and application to image super-resolution with CUDA acceleration},
  author={Tan, Hanlin and Xiao, Huaxin and Liu, Yu and Zhang, Maojun and Wang, Bin},
  booktitle={Image, Vision and Computing (ICIVC), 2017 2nd International Conference on},
  pages={483--488},
  year={2017},
  organization={IEEE}
}

linzehua/Sparse-Coding-based-Image-Super-Resolution-with-CUDA