Deep Learning based Fourier Spectrum Sampling Strategy for Improving Performance of Imaging

EasyChair Preprint 1476

Abstract

Single-pixel imaging is a framework that reconstructs the target information only with a single bucket detector. The principle of the single-pixel imaging is correlating the measurements of a single bucket detector and the corresponding 2D light field distributions modulated by an optical field device in the scene. Single-pixel imaging has a good prospect in various imaging applications. To improve the imaging quality and speed, the compressed sensing and the basis scan strategies are demonstrated at the current stage. Based on the Fourier single-pixel imaging, the representative one of the basis scan strategies, and the deep learning, we propose a deep learning based Fourier spectrum sampling strategy for Fourier single-pixel imaging. Our goal is to predict and acquire the significant Fourier coefficients instead of the traditional sampling strategies to recover higher quality image under the same measurements. The simulation results demonstrate that the reconstruction image of the proposed strategy outperforms others. Applications to high quality and speed imaging could benefit from our strategy.

Keyphrases: Convolutional Neural Network, Fourier single-pixel imaging, deep learning, single-pixel imaging

@booklet{EasyChair:1476,
  author    = {Kaiyu Zhang and Jie Cao and Lei Yan and Fanghua Zhang and Qun Hao},
  title     = {Deep Learning based Fourier Spectrum Sampling Strategy for Improving Performance of Imaging},
  howpublished = {EasyChair Preprint 1476},
  year      = {EasyChair, 2019}}