Unnatural L
0
Sparse Representation for Natural Image Deblurring
Li Xu Shicheng Zheng Jiaya Jia
The Chinese University of Hong Kong
http://www.cse.cuhk.edu.hk/leojia/projects/l0deblur/
Abstract
We show in this paper that the suc cess of previous maxi-
mum a posterior (MAP) based blur removal methods partly
stems from their respective intermediate steps, which im-
plicitly or explicitly create an unnatural representation con-
taining salient image structures. We propose a generalized
and mathematica lly sound L
0
sparse expression, together
with a new effective method, for motion deblurring. Our
system does n ot require extra filtering during optimization
and demonstrates fast energy decreasing, mak ing a small
number of iterations enough fo r convergence. It also pro-
vides a unified fra m ework for both uniform and non-unifo rm
motion deblurring. We extensively validate our method and
show comparison with other approaches with respect to
convergence speed, running time, and result quality.
1. Introduction
Single-imag e motion d eblurring, a.k.a. blind deconvolu-
tion, was extensively studied in these a f ew years and has
achieved great success with a few milestone solutions. Be-
cause naive maximum a posterior (MAP) inference could
fail on natural images, state-of-the-art m e thods either max-
imize marginalize d distributions [5, 17, 18, 6] or propose
novel techniques in MAP to effectively avoid tr ivial delta
kernel estimates [1 1, 20, 3, 25, 4, 10].
The set of effective techniques that reinvigorate MAP
[20, 3, 25] can produce high-quality results in seconds and
were broadly adopted in a number of applications. Th ey
also form basic steps fo r sp atially-variant deblurring. The
particularly useful techniques include adaption of the en-
ergy function du ring optimization [20], explicit sharp edge
pursuit [11, 13, 19, 3, 25, 4, 10], edge selection [25], and
employment of normalized sparsity measure [16]. These
methods achieve efficient inference with their distinct f or-
mulation and o ptimization steps, discussed below.
1.1. Analysis
Prior MAP-based approaches can be roughly categoriz e d
into two groups, i.e., methods with ex plicit edge prediction
(a) inp ut (b) Shan et al. [20]
(c) Krishnan et al. [16] (d) Cho and Lee [ 3]
(e) our ˜x (f) final restored image
Figure 1. Intermediate unnatural i mage representation exists in
many state-of-the-art approaches.
[11, 13, 19, 3, 25, 9, 23], which are referred to as semi-blind
schemes, and those [20, 16] implicitly incorporating spe-
cial regularization to remove detrimental structures in early
stages an d gradually enrich image details in iterations.
We fo und representative approaches in these two re spec-
tive groups share th e commonness in the middle of the pro-
cedure to generate one or multip le spec ia l maps only con-
1