Skip to content

Latest commit

 

History

History
64 lines (37 loc) · 2.75 KB

an-autoencoder-based-formulation-for-compressed-sensing-reconstruction-2018-200118.md

File metadata and controls

64 lines (37 loc) · 2.75 KB

Paper:

An autoencoder based formulation for compressed sensing reconstruction

Author:

Angshul Majumdar

Year:

2018

Notes:

阅读日期:2020.01.18

泛读。这篇文章研究的是MRI 重建。作者的想法是用自编码来自适应重建图像,类似于 DL 或 TL,这种做法不需要训练数据,是在重建图像的过程中训练自编码。作者使用 ADMM 算法来求解,因为自编码器只使用了一层隐层,而且激活函数是可逆的,因此在数学上可以比较容易的求解 ADMM 的子问题。另外因为不需要 SVD 分解,求解的速度比 DL 和 TL 要快。(以及,作者在文章中说 deep learning 的方法比 CS 好,但比 DL 和 TL 要差)

DL 的优化目标函数:

TL 的优化目标函数:

自编码器的优化目标函数:

个人感觉自编码器有点像 DL 和 TL 的结合。

ADMM 算法的求解步骤:

  1. least squares problem

  1. least squares problem

  1. least squares problem

  1. $l_1$-regularized least squares problem,using iterative soft thresholding algorithm

Bibtex:

@article{majumdar2018autoencoder,
  title={An autoencoder based formulation for compressed sensing reconstruction},
  author={Majumdar, Angshul},
  journal={Magnetic resonance imaging},
  volume={52},
  pages={62--68},
  year={2018},
  publisher={Elsevier}
}