yangyan92 / ADMM-CSNet

Matlab code for ADMM-CSNet

ADMM-CSNet

Generic-ADMM-CSNet

---

These are testing and training codes for Generic-ADMM-CSNet in "ADMM-CSNet: A Deep Learning Approach for Image Compressive Sensing" (TPAMI 2019)

If you use thses codes, please cite our paper:

[1] Yan Yang, Jian Sun, Huibin Li, Zongben Xu. ADMM-CSNet: A Deep Learning Approach for Image Compressive Sensing (TPAMI 2019).

http://gr.xjtu.edu.cn/web/jiansun/publications

All rights are reserved by the authors.

Yan Yang -2019/04/10. For more detail or traning data, feel free to contact: [e-mail]

---

Data link:

https://pan.baidu.com/s/1nvf07g_OmMAnFAbhG1orIQ passwards：sdsq

Usage:

1. Three folders.

   1). 'Generic-ADMM-CSNet-ComplexMRI' are testing and training codes to reconstruct complex-valued MR images with 1D Cartesian masks and 2D random masks.
   2). 'Generic-ADMM-CSNet-RealMRI' are testing and training codes to reconstruct real-valued MR images with the Pseudo radial mask.
   3). 'Generic-ADMM-CSNet-Image' are testing and training codes to reconstruct natural images with the randomly permuted coded diffraction operators and Walsh-Hadamard operators.
   

   Please do not add these three folders into the path at the same time, because they contain the functions with the same name.

2. For testing the trained network for a single image.
   ('./Generic-ADMM-CSNet-ComplexMRI/main_ADMM_CSNet_test.m')
   ('./Generic-ADMM-CSNet-RealMRI/main_ADMM_CSNet_test.m')
   ('./Generic-ADMM-CSNet-Image/main_ADMM_CSNet_test.m')
   

   1). Load trained network with different stages in main_ADMM_CSNet_test.m.
   If you apply ADMM-CSNet to reconstruct other MR or natural images, it is best to re-train the models.
   

    E.g., The model './net/NET-1D-Cartesian-0.2-complex-S10.mat' is the ADMM-CSNet with 10 stages trained from 100 complex-valued MR images using 1D Cartesian mask with 20% sampling rate.
          The model './net/NET-Pseudo-radial-0.2-real-S11.mat' is the network with 11 stages trained from 100 real-valued MR images using Pseudo radial mask with 20% sampling rate.
          The model './net/Net-Diffraction-0.05-S10.mat' is the network with 10 stages trained from 100 real-valued natural images using coded diffraction operator with 5% sampling rate.
   

   2). Load test image in main_ADMM_CSNet_test.m
   The images in './data/Brain_complex_data', './data/Brain_real_data', './data/Image' are fully-sampled images.
   

   3). Load sampling mask or operator with different sampling ratios in main_ADMM_CSNet_test.m
   

    E.g., The mask './mask/1D-Cartesian-0.2.mat' is a 1D Cartesian mask with 20% sampling rate.
          The mask './mask/D-0.1.mat' is a coded diffraction operator with 10% sampling rate. 
   

   4). Network testing setting (network structure or training setting) is in 'config.m '.
   

   5). To test our ADMM-CSNet, run 'main_ADMM_CSNet_test.m'
   

3. For testing the trained network for our testing dataset.
   ('./Generic-ADMM-CSNet-ComplexMRI/AverageTesting.m')
   ('./Generic-ADMM-CSNet-RealMRI/AverageTesting.m')
   ('./Generic-ADMM-CSNet-Image/AverageTesting.m')
   

   1). Load trained network with different stages in AverageTesting.m.
   If you apply ADMM-CSNet to reconstruct other MR or natural images, it is best to re-train the models.
   

    E.g., The model './net/NET-1D-Cartesian-0.2-complex-S10.mat' is the ADMM-CSNet with 10 stages trained from 100 complex-valued MR images using 1D Cartesian mask with 20% sampling rate.
          The model './net/NET-Pseudo-radial-0.2-real-S11.mat' is the network with 11 stages trained from 100 real-valued MR images using Pseudo radial mask with 20% sampling rate.
          The model './net/Net-Diffraction-0.05-S10.mat' is the network with 10 stages trained from 100 real-valued natural images using coded diffraction operator with 5% sampling rate.
   

   2). Set the data_dir of testing dataset ans load the correspongding mask in AverageTesting.m
   

    E.g., data_dir = './data/DATA-1D-Cartesian-0.2-complex-brain/test/' is the testing dataset including 100 complex-valued brain MR image with 20% 1D-Cartesian mask.  
          data_dir = './data/Testingdata/Sdata10/D/D_0.1_1/' is the first testing dataset including 10 standard image with 10% coded diffraction operator.  
          data_dir = './data/DATA-Pseudo-radial-0.2-real-brain/test/'is the testing dataset including 50 real-valued brain MR image with 20% Pseudo radial mask.  
   

   3). Network testing setting (network structure or training setting) is in 'config.m '.
   

   4). To test our ADMM-CSNet, run 'AverageTesting.m'
   

4. For re-training the ADMM-CSNets
   

   1). Set the data_dir of training dataset ans load the correspongding mask in L_BFGSnetTrain.m.
   

   2). Modify the network setting and trainging setting in 'config.m '.
   

   3). To train ADMM-CSNet by L-BFGS algorithm, run ' L_BFGSnetTrain.m' .
   

   4). After training, the trained network and the training error are saved in './Train_output'.
   

---