Image Deblurring: A Modern Approach

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


PhD Thesis Defence


Title: "Image Deblurring: A Modern Approach"

By

Mr. Lu Yuan


Abstract

The Recovery of a sharp version of the blurred image is a challenging 
problem in digital imaging. Previous works have achieved dramatic 
progress, yet the heavy ill-posedness of the problem leads to the results 
still far from perfect. In my thesis, I will explore valuable information 
or latent priors from observations to provide a better condition for image 
deblurring.

The first idea is to take advantage of additional correlated images. By 
combining information between two degraded images – blurred/noisy image 
pair, we can estimate a very accurate blur kernel and restore a 
high-quality original image, which cannot be obtained by simple single 
image denoising or single image deblurring. The idea further pushes me to 
develop a more general framework for image deblurring with a sequence of 
images, which do not limit to blurred/noisy image pair, and even can be 
multiple blurred images. Our approach is based on the assumption that 
these blurred images with different blurs which are derived from the same 
original image and different blurs will result in the loss of different 
frequency components during imaging. By integrating these complementary 
information together, we can see these additional correlated images can 
further eliminate ambiguous solutions in both kernel estimation and image 
restoration. Furthermore, a prominent problem in image deblurring with 
multiple images is how well image pairs can be aligned. We then proposed a 
fully automatic alignment approach for multiple images using the 
sparseness prior of blur kernels. Thus our methods are very practical and 
effective for achieving satisfactory photos in dim light conditions using 
off-of-shelf hand-held camera.

The second idea is to make use of the prior of sharp image structure. In 
our image deblurring with blurred/noisy image pair, the noisy image as the 
guide image provides sharp large-scale edges for accurate kernel 
estimation and high-quality image restoration. I will show this insight 
can be furthermore applied to single image deblurring. As we observed, the 
reconstructed image usually contains unpleasant artifacts, i.e. ringing, 
due to the ill-posedness of image deconvolution even if the blur kernel is 
known. To suppress ringing artifacts and preserve recovered signals, we 
require the guide image to tell where edges and texture regions are, and 
where flat regions are. Thus, we develop an inter-scale and intra-scale 
deconvolution framework to progressively recover such a guide image, which 
is then used to adaptively suppress artifacts in texture regions and flat 
regions. Our progressive deconvolution approach can produce very promising 
results not only in synthetic experiments, but also in various types of 
real cases. Our results show our approach outperforms other 
state-of-the-art techniques and have wide applications in scientific and 
daily areas


Date:			Monday, 24 August 2009

Time:			2:00pm-4:00pm

Venue:			Room 3501
 			Lifts 25-26

Chairman:		Prof. Tongxi Yu (MECH)

Committee Members:	Prof. Long Quan (Supervisor)
 			Prof. Chiew-Lan Tai
 			Prof. Chi-Keung Tang
 			Prof. Bing Zeng (ECE)
 			Prof. Tien-Tsin Wong (Comp. Sci. & Engg., CUHK)


**** ALL are Welcome ****