[Bayesian Convolutional Neural Networks for Compressed Sensing Restoration]

Introduction

Bayesian Convolutional Neural Networks (BCNNs) is a new Compressed Sensing (CS) restoration algorithm that combining Convolutional Neural Networks (CNNs) and Bayesian inference method. In this paper, we show significant improvements in reconstruction results over classical Structured Compressed Sensing (SCS) algorithms and restoration methods based on neural networks, such as ReconNet, DR2Net, and LDAMP. The code provided here helps one to reproduce some of the results presented in the paper.

Citation (BibTex):

If you are using this code, please cite the following paper.

@artical{BCNNs,
author = {Xinjie Lan and Xin Guo and Kenneth E. Barner},
title = {Bayesian Convolutional Neural Networks for Compressed Sensing Restoration},
booktitle = {arVix:1811.04356},
month = {Nov.},
year = {2018}
}

System Requirements:

This software has been tested on Matlab R2018a.

Reconstructing images from CS measurement using BCNNs:

In the 'restoration' folder, there are four per-trained BCNN models. Each models have different parameters of BCNN. You can find the detailed information of these models in the experimental section of our paper.

Run demo_bcnns_cs.m from /restoration folder to use these models to implement CS reconstruction. The default image dimension is 64 by 64, CS measurement ratio is 0.25, and not add any noise. You can reproduce our simulations if you reset parameters as in our paper (e.g. image dimension changes to 128 by 128, measurement ratio is 0.1, SNR = 8 etc).

In our paper, we compared BCNN model with two classical Structured Compressed Sensing methods and three neural network algorithms as follows.

BCS (L. Yu, C. Wei, and G. Zheng, “Adaptive bayesian estimation with cluster structured sparsity,” Signal Proc. Letters, vol. 22, pp. 2309–2313, 2015.) https://sites.google.com/site/link2yulei/publications
TV (C. Li, “An efficient algorithm for total variation regularization with applications to the single pixel camera and compressive sensing,” Master’s thesis, Rice University, Houston, Texas, 2009.) https://www.caam.rice.edu/~optimization/L1/TVAL3/
ReconNet (K. Kulkarni, S. Lohit, P. Turaga, R. Kerviche, and A. Ashok, “Reconnet: Non-iterative reconstruction of images from compressively sensed measurements,” in IEEE Conf. on CVPR, June 2016.) https://github.com/KuldeepKulkarni/ReconNet
DR2Net (H. Yao, F. Dai, D. Zhang, Y. Ma, S. Zhang, and Y. Zhang, “Dr2-net: Deep residual reconstruction network for image compressive sensing,” arXiv preprint arXiv:1702.05743, 2017.) https://github.com/coldrainyht/caffe_dr2/tree/master/DR2
LDAMP (C. A. Metzler, A. Mousavi, and R. G. Baraniuk, “Learned d-amp:principled neural network based compressive image recovery,” arxivpreprint arXiv:1704.06625, 2017.) https://github.com/ricedsp/D-AMP_Toolbox

Training models:

Run demo_learning.m from /learning folder to train BCNN model using an image dataset from /+image_patches/training91 folder. Since ReconNet and DR2Net use this image dataset to train their models, we also use this dataset in order to make a fair comparison.

Contact:

Xinjie Lan, (lxjbit@udel.edu)

Acknowledgements:

Our code is generated based on the Field of Experts code.

csyifangli/BCNN