libsrcnn

Super-Resolution imaging with Convolutional Neural Network

A stand-alone library for Super-Resolution, Non-OpenCV model related in these projects:

• https://github.com/rageworx/SRCNN_OpenCV_GCC
• https://github.com/shuwang127/SRCNN_Cpp.

Introduction

This is an open source project from original of this: SRCNN_Cpp is a C++ Implementation of Image Super-Resolution using SRCNN which is proposed by Chao Dong in 2014.

• If you want to find the details of SRCNN algorithm, please read the paper:

  Chao Dong, Chen Change Loy, Kaiming He, Xiaoou Tang. Learning a Deep Convolutional Network for Image Super-Resolution, in Proceedings of European Conference on Computer Vision (ECCV), 2014

• If you want to download the training code(caffe) or test code(Matlab) for SRCNN, please open your browse and visit http://mmlab.ie.cuhk.edu.hk/projects/SRCNN.html for more details.
• And thank you very much for Chao's work in SRCNN.

WIKI

• See Wiki page.
• Includes OpenMP support for macOS.

Additional references

• Fast resizing methods for BiCubic filtering
  • Free Image Project
  • librawprocessor
  • fl_imgtk

Features

• Faster about 400% or more than OpenCV GCC version of SRCNN, with OpenMP.
  • references :
  • commit/b340b885a58605f793aa000eebc7f96c19c8e9fe#commitcomment-103507343
  • commit/b340b885a58605f793aa000eebc7f96c19c8e9fe#commitcomment-103507802
• None-OpenCV, no OpenCV required !
• Compilation availed for almost of POSIX g++.
• Simply optimized for basic OpenMP.
• Works well even without OpenMP like macOS.
• Not support M$VC, sorry MS guyz.

Sample images

• Original 100%

  

• Bicubic 150%

  

• SRCNN 150%

  

• SRCNN (Convolution Y channel) 150%

  

Supporting platforms

• Windows 32, 64 with MSYS2 + MinGW-W64
• Almost any Linux, x86_32, x86_64, arm, armhf, aarch64
• macOS ( clang, llvm )

Latest Changes

Verison 0.1.10.40

• Better speed, less memory usage by convolution I+II
• Regards to zvezdochiot@github

Previous Changes

Verison 0.1.9.35

• Fixed memory bug in float images from RGB case.
• header version flag fixed.

Verison 0.1.9.34

• Fixed don't use color space scaling with bicubic filter.
• Now supporting alpha channel.

Verison 0.1.8.30

• Precision step scaling bug fixed.

  Verison 0.1.8.28

• Precision step scaling option availed.
• included option by reason of libsrcnn trained for maximum double multiply.

  Verison 0.1.6.23

• Fixed a small bug of wrong internal copying size.

  Verison 0.1.6.22

• Fixed bug of original source (ShuWang's SRCNN).
  • Use last layer (3) to Y channel at last construction.
• Changed ProcessSRCNN() method to get optional convolutional result.

  Verison 0.1.6.20

• Fixed memory leak after convolution55.
• Changed ProcessSRCNN() method to get convolutional gray.

  Version 0.1.4.17

• Bug fixed for color space conversion.

  Verison 0.1.5.18

• Supports variable filters for interpolation.
  1. Nearest
  2. Bilinear
  3. Bicubic
  4. Lanczos-3
  5. B-Spline

License

• Follows original source GPLv2, but this project is LGPLv3.

Requirements

• Your G++.

How to build ?

• Make a symlink from Makefile.{your platform} in makefiles directory.
  • eg.) ln -s makefiles/Makefile.macos Makefile
• Then build with make.
• Testing applications may one of these,
  • make -f makefiles/Makefile.test
  • or
  • make -f makefiles/Makefiles.testmac

Dependency

• Testing application by make -f makefiles/Makefile.test may requires FLTK and fl_imgtk libraries.
• FLTK should be installed by anyway, but recommend to my below FLTK-custom with fl_imgtk.
• FLTK-custom and fl_imgtk for build test program for read and write image files.