Visual Enhancement using Multiple Observations

PhD Thesis Proposal Defence


Title: "Visual Enhancement using Multiple Observations"

by

Mr. Jia CHEN


Abstract:

Due to the fundamental limitations of camera, the captured photographs can 
be defective. Camera blur and noise are two major types of defect. 
Enhancement of image and video is an important topic in computer vision 
and graphics, because it can serve as either a pre-processing step for 
other algorithms, or a post-processing step for directly enhanceing the 
output for users.

In this proposal, I will explore pertinent issues in visual enhancement, 
especially for blur and noise, and propose viable solutions to address the 
problem. While most traditional works mainly use a single image for the 
purpose of enhancement, I will focus on the usage of multiple 
observations. While more observations give us more information, many 
research challenges make the problem difficult. First of all, it is 
important to construct and collect multiple observations properly. Second, 
the observations should be appropriately utilized in a computational 
framework. A unified multi-observation enhancement approach is presented 
in this proposal, where we emphasize the importance of invariance in 
linking multiple observations. We will also derive specific optimization 
procedures integrating prior knowledge with these observations.

I will first analyze the image deblurring problem using two observations. 
Given that the two inputs are taken from the same static scene, the 
invariance is the common clear image. Since the two inputs have different 
motion blur artifacts, their frequency responses are complementary to each 
other. A feedback algorithm is proposed which effectively combines two 
independent input blurred images. This approach introduces image prior and 
motion prior in the context of multiple observations. The visual quality 
of enhancement can be significantly improved compared to approaches using 
single images.

The second part of this proposal will focus on the denoising problem. 
Removing noise from image is a topic that has been studied for decades. 
However, there are limitations from most previous automatic approaches 
where they usually take the image itself as the processing target.  We 
show that even with a single input image, an auxiliary observation, namely 
the noise layer, can be constructed. With this new noise layer, the 
artifacts of automatic denoising algorithms can be easily visualized and 
optimization can be performed on both image layer and the noise layer. We 
propose an interactive system based on this representation, and high 
quality image noise separation results can be achieved.

The denoising system will be extended from image to video where multiple 
frames are available as observations. The central idea this proposal has 
exploited is how to set up the connections between these observations. 
Classical methods of finding inter-frame correspondences use optical flow 
which gives pixel-wise motion field. An extended motion field, which is 
called probabilistic motion field will be introduced to setup soft 
temporal correspondences. The corresponding pixels will be placed inside a 
spatio-temporal Markov Random Field where the denoised frames are 
optimized from multiple observations.


Date:     		Monday, 22 June 2009

Time:                   1:30pm-3:30pm

Venue:                  Room 4483
 			lifts 25-26

Committee Members:      Dr. Chi-Keung Tang (Supervisor)
 			Dr. Huamin Qu (Chairperson)
 			Prof. Long Quan
 			Dr. Chiew-Lan Tai


**** ALL are Welcome ****