Joint Demosaicing and Denoising of Noisy Bayer Images with ADMM

Hanlin Tan, Xiangrong Zeng, Shiming Lai, Yu Liu and Maojun Zhang

College of Information System and Management, National University of Defense Technology, China

This project is a demo for our ICIP 2017 paper Joint Demosaicing and Denoising of noisy Bayer Images with ADMM. The project compare three algorithms, DeepJoint[2], FlexISP[3] and the proposed ADMM algorithm, on two datasets: Kodak and McMaster[1].

Note that we slightly modified I/O code (mainly redirecting input and output paths) of FlexISP and DeepJoint(Demosaicnet) to perform comparisons. However, the kernel part of those algorithms, including all parameters, remains the same as their authors provided.

Folder Structure

Folder Name	Explanation
data	Datasets Kodak and McMater[1] are here.
demo_joint_isp	Code of the proposed ADMM method.
demosaicnet	DeepJoint[2] code.
flexisp_demosaicking_demo	FlexISP[3] code.
res	Results.
utils	Code needed for PSNR computation and Bayer mask computation.
wrappers	Wrappers for comparative algorithms

Usage

In summary, useful entrance files are:

File Name	Explanation
compareAlgorithms.m	The entrance file. Run this file with Matlab and the results are stored in compareResults.mat in res/{dataset} folder.
analyze.m	Produce figures and tables of results. This file is called automatically by compareAlgorithms.m. Results will be saved in res/{dataset} folder.
combineResultImages.m	Reproduce figures of visual results using data in res folder.

Simply run compareAlgorithms.m, which will generate a result mat file in res/{dataset} folder.

• You can control the behavior of the code by modifying structure variable conf.

• Note DeepJoint method depends on caffe. You can turn off evaluation of that algorithm by setting conf.enableDeepJoint to false in the control section of compareAlgorithms.m if you do not have caffe installed.

%% control
conf.debug = false;
conf.enableADMM = true;      % enable admm method
conf.enableFlexISP = true;   % enable flex isp method
conf.enableDeepJoint = false; % enable deep joint method
conf.cpuorgpu = 'gpu';    % 'gpu' or 'cpu', wheter to use gpu in deep joint
conf.tmpdir = '~/tmp/';   % a directory to store temporary output

It will call analyze(dataset, sigma, nAlgos) to produce result images, where dataset and sigma are in accordance with those in compareAlgorithms.m. For exmaple,

analyze('Kodak', 0, 3)

Optionally, run combineResultImages(dataset) to combine result images of different noise levels into single view, which we used in our paper. For exmaple,

combineResultImages('Kodak')

References

If you use the code in your work, please cite our paper:

@inproceedings{tan2017joint, 
  title={Joint Demosaicing and Denoising of noisy Bayer Images with ADMM},
  author={Tan, Hanlin and Zeng, Xiangrong and Lai, Shiming and Liu, Yu and Zhang, Maojun},
  booktitle={Image Processing (ICIP), 2017 IEEE International Conference on},
  pages={2951--2955},
  year={2017},
  organization={IEEE}
}

The comparison code directly use the dataset and code from the following papers:

[1] Zhang, Lei, et al. "Color demosaicking by local directional interpolation and nonlocal adaptive thresholding." Journal of Electronic imaging 20.2 (2011): 023016-023016.

[2] Gharbi, Michaël, et al. "Deep joint demosaicking and denoising." ACM Transactions on Graphics (TOG) 35.6 (2016): 191.

[3] Heide, Felix, et al. "Flexisp: A flexible camera image processing framework." ACM Transactions on Graphics (TOG) 33.6 (2014): 231.