IMPLEMENTATION OF MODEL-BLIND VIDEO DENOISING VIA FRAME-TO-FRAME TRAINING

• Author : EHRET Thibaud ehret.thibaud@gmail.com
• Licence : AGPL v3

OVERVIEW

This code is provided to reproduce the results from "Model-blind Video Denoising Via Frame-to-frame Training , T. Ehret, A. Davy, J.M. Morel, G. Facciolo, P. Arias, CVPR 2019". Please cite it if you use this code as part of your research.

The sequences used for the article can be found in https://github.com/cmla/derf-hd-dataset.

USAGE

List all available options:
python blind_denoising.py --help

There are 4 mandatory input arguments:

• --input the path to input frames (C type)
• --flow the path to optical flow (C type), must be readable by readFlowFile.py
• --output the path to output frames (C type)
• --ref the path to reference frames (C type), it can be set to the input frames if the true reference is not available (only used to compute the PSNR)

There are 5 optional input arguments:

• --first the index of the first frame to be processed, default is 1
• --last the index of the last frame to be processed, default is 300
• --output_psnr the path to file where the PSNRs are going to be written, default is plot_psnr.txt
• --output_network the path to file where the fine-tuned network will be saved, default is final.pth
• --iter the number of backpropagation to done, default is 20
• --network the path to the network (only change if you know what you are doing), default is set to DnCNN 25

The input sequence provided should already be a degraded grayscale sequence (the code can read grayscale png, jpeg and tiff files). The optical flow should be computed before running the denoising code (and preferably on the degraded sequence).

OPTICAL FLOW

The code used to compute the optical flow for the CVPR paper is provided in the tvl1flow folder. It's a modified version of "Javier Sánchez Pérez, Enric Meinhardt-Llopis, and Gabriele Facciolo, TV-L1 Optical Flow Estimation, Image Processing On Line, 3 (2013), pp. 137–150."

The code is compilable on Unix/Linux and hopefully on Mac OS (not tested!).

Compilation: requires the cmake and make programs.

Compile the source code using make.

UNIX/LINUX/MAC:

$ mkdir build; cd build
$ cmake ..
$ make

Binaries will be created in the build/ folder.

NOTE: By default, the code is compiled with OpenMP multithreaded parallelization enabled (if your system supports it).

A script tvl1flow.sh is also provided. The command to run this script is (assuming it is in the same folder as the tvl1flow binary):

$ ./tvl1flow.sh inputPath first last outputPath

where the mandatory inputs are:

• inputPath the path to the input frames (using the C standard), for example 'frame%03d.png'
• first the index of the first frame, for example '1'
• last the index of the last frame, for example '100'
• outputPath the path to the output flow, for example 'tvl1_%03d.flo'