Missing channel reconstruction for sloan digital sky survey images using linear models and generative adversarial networks

The Sloan Digital Sky Survey (SDSS) dataset, one of the largest astronomical surveys, suffers from noise and missing information in some of the image channels. This thesis implements two methods—the linear model and the deep learning model—on 12,730 SDSS images for missing channel re- construction. Specifically, for the linear model, linear regression and patch- based regression are examined. For the deep learning model, the generative adversarial networks (GANs) with U-Net are deployed in the experiment. Several preprocessing techniques including normalization and cropping are done before feeding the images into the model. The results indicate that both methods can generate satisfactory results. In addition, there is a trade- off between training speed and the accuracy. Specifically, the training of the linear model is much faster than that of the GAN model, while the L1 loss of the GAN model can achieve average 15.29 per image, which is much smaller than the L1 loss of the linear model.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2019

Author dc:creator

• Cheng, Yuan

Contributors dc:contributor

• Zhao, Zhizhen

• Image Reconstruction, Deep Learning, Sloan Digital Sky Survey, Generative Adversarial Networks, Linear Regression

Statement dc:rights

• Copyright 2019 Yuan Cheng

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/105266

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/105266