SRCNN-cpp

Pure C++ implementation of SRCNN (Super-Resolution Convolutional Neural Network)

How to Build and Run

We provide two ways to build this project: CMake and QMake

CMake

If you just want to build the program and quickly test it, type these:

$ mkdir build && cd build && cmake .. && make && cd ..
$ SRCNN_cpp <input image path>

If you want much faster speed with OpenMP, follow these steps:

1. Set i_want_openmp from OFF to ON in CMakeLists.txt.
2. Follow the usual CMake building steps then run it.

QMake

Just run the SRCNN-cpp.pro. Also, the OpenMP option in QMake is set to ON by default.

Note

This implementation uses 9-5-5 architecture.

• 9-5-5 means f1 = 9, f2 = 5, f3 = 5.
• You can see the explanation in Section 4.3.2 in SRCNN paper.

TODO

• im2col
• validate convolution using this image
• padding feature

Note

• The weight format is CHWN. For more information, download the author's training code here.
  • The weight format conversion and storing code is in saveFilters.m.

Contributors

• Cuda-Chen Cuda Chen - creator, maintainer
• masc4ii masc4ii - OpenMP and QMake utility