GitHub - kristiantorres/CNN-RPGD: Code for 'CNN-Based Projected Gradient Descent for Consistent Image'.

kristiantorres / CNN-RPGD Public

forked from harshit-gupta-cor/CNN-RPGD

Notifications You must be signed in to change notification settings
Fork 0
Star 1

Code for 'CNN-Based Projected Gradient Descent for Consistent Image'.

ieeexplore.ieee.org/abstract/document/8353870

1 star 11 forks Branches Tags Activity

Star

Notifications

Branches Tags

Name		Name	Last commit message	Last commit date
Latest commit History 11 Commits
ProjectorTrainingFunctions		ProjectorTrainingFunctions
RPGDfunctions		RPGDfunctions
.DS_Store		.DS_Store
.gitattributes		.gitattributes
RPGDCTMeasurementModel.m		RPGDCTMeasurementModel.m
Readme.txt		Readme.txt
TrainingCTMeasurementModel.m		TrainingCTMeasurementModel.m
addpathsPT.m		addpathsPT.m
addpathsRPGD.m		addpathsRPGD.m
cnn_unet_init.m		cnn_unet_init.m
dev_fun.m		dev_fun.m
getBatch_test.m		getBatch_test.m
loadCandStep.m		loadCandStep.m
loadNettesting.m		loadNettesting.m
loadNettraining.m		loadNettraining.m
vl_euclideanloss.m		vl_euclideanloss.m
vl_nnupconv.m		vl_nnupconv.m
vl_simplenn_fbpconvnet.m		vl_simplenn_fbpconvnet.m
vl_simplenn_fbpconvnet_eval.m		vl_simplenn_fbpconvnet_eval.m

Repository files navigation

==================================
New announcement: For a python/pytorch based version please refer to https://github.com/PhanHuyThong/Image-Reconstruction-by-CNN-based-PGD/

In that repo the pytorch code is used in training the projector network. This trained network can then be plugged in RPGD scheme which has both python and matlab support. For the matlab case onnx is used to translate the learnt network from pytorch to matlab. The parameters for the projector training and RPGD can be passed on using config file. The forward model that was used in [1] was the radon/iradon of matlab. However, similar forward model can be found in python.
I highly recommend using the code from that repo.
==================================


This repo is in matconvnet.




[1] H. Gupta, K. H. Jin, H.Q.Nguyen, M.T. McCann, and M. Unser, 'CNN-Based Projected Gradient Descent for Consistent Image Reconstruction', IEEE TMI, 2018. https://ieeexplore.ieee.org/abstract/document/8353870

[2]  K. H. Jin, M.T. McCann, E. Froustey, and M. Unser, 'Deep CNN for Inverse problem in Imaging', IEEE Transactions on Image Processing, 2017. http://ieeexplore.ieee.org/abstract/document/7949028/

Readme

1. Before launching CNN-RPGD, kindly install the MatConvNet (http://www.vlfeat.org/matconvnet/) (For the GPU, it needs a different compilation.)
2. Modify addpathsRPGD and addpathsPT based on the realtive directory paths on your machine. TrainingCTMeasurementModel, RPGDCTMeasurementModel should be edited to change the default parameters of RPGD.
3. The codes TrainingCTMeasurementModel.m is based on GPU computation. Kindly modify it to use it on CPU.
4. Use TrainingCTMeasurementModel for training the projector. After training, run RPGDCTMeasurementModel to get the result on the test data.

*note : these codes successfully ran on Matlab 2016a with GPU TITAN X (architecture : Maxwell)
contact : Harshit Gupta ([email protected]),