A SURVEY OF LEARNING-BASED IMAGE DENOISING

PhD Qualifying Examination


Title: "A SURVEY OF LEARNING-BASED IMAGE DENOISING"

by

Mr. Chenyang QI


Abstract:

DSLR Cameras and mobile phones convert light intensity into raw digital images 
through optical lenses and sensor circuits. Taking the raw image as input, 
image processing pipelines (ISP) produces the final RGB image for 
visualization. Image noise can be very pronounced in low-light images (e.g., 
microscope or capturing in the dark night), which makes image denoising an 
indispensable step in the ISP. In this literature review, we provide a 
comprehensive review of image denoising and noise modeling. Specifically, we 
first introduce the physics background of camera sensors and the image 
processing pipeline, which provides the common property of noise. Then, we 
review the neural network architecture in recent learning-based image 
denoising, including building blocks(e.g., convolution, attention) and 
intra-block architecture (e.g., multi-stage, multi-scale). The advantages and 
disadvantages of each category are also analyzed. Moreover, to provide a 
high-quality large-scale training data set, we discuss recent noise synthesis 
methods, including physics-based, GAN-based, and Flow-based generators. In the 
end, we discuss several potential research directions.


Date:  			Friday, 17 June 2022

Time:                  	2:00pm - 4:00pm

Zoom Meeting:
https://hkust.zoom.us/j/92158471956?pwd=aXV2dVAwOHlrSVdjNGszWU9ZcHlnZz09

Committee Members:	Dr. Qifeng Chen (Supervisor)
 			Dr. Dan Xu (Chairperson)
 			Dr. Xiaomeng Li
 			Dr. Yingcong Chen (AI Thrust)


**** ALL are Welcome ****