Decemb er 10, 2010 / Vol. 8, No. 12 / CHINESE OPTICS LETTERS 1127
Modified level set method with Canny operator for image
noise removal
Pengcheng Wen (文文文鹏鹏鹏程程程)
1∗
, Xiangjun Wang (王王王向向向军军军)
1
, and Hong Wei (卫卫卫 红红红)
2
1
State Key Lab oratory of Precision Measuring Technology and Instruments, Tianjin University, Tianjin 300072, China
2
Scho ol of Systems Engineering, University of Reading, Whitenights, PO Box 217, Berkshire UK
∗
E-mail: victorlionwen@yaho o.com.cn
Received May 10, 2010
The level set metho d is commonly used to address image noise removal. Existing studies concentrate mainly
on determining the speed function of the evolution equation. Based on the idea of a Canny operator, this
letter introduces a new method of controlling the level set evolution, in which the edge strength is taken
into account in choosing curvature flows for the speed function and the normal to edge direction is used
to orient the diffusion of the moving interface. The addition of an energy term to penalize the irregularity
allows for better preservation of local edge information. In contrast with previous Canny-based level set
metho ds that usually adopt a two-stage framework, the prop osed algorithm can execute all the above
op erations in one pro cess during noise removal.
OCIS codes: 110.4280, 100.2000.
doi: 10.3788/COL20100812.1127.
The use of the level set method
[1,2]
has become popu-
lar due to its flexibility and capability in modeling com-
plex structures. The basic idea of this geometric ap-
proach is to evolve a higher dimensional implicit func-
tion, whose zero level set always represents the position of
the propagating front according to a partial differential
equation. It executes natural topological changes and
allows efficient numerical solutions. Several previous
studies
[3−5]
have referred to the application of the level
set method to the issue of image noise removal. Mal-
ladi et al. have driven the image evolution under flows
controlled by the min/max and mean curvature
[3]
. This
algorithm is less sensitive to the nature of noise and is
applicable to both salt-and-pepper grey-scale noise and
full-image continuous noise. However, this algorithm pe-
nalizes high curvature value regardless of the curve reg-
ularity and could cause edge blurring. Gil et al. have
presented a novel geometric flow that included a function
measuring the degree of local irregularity in the curve
[4]
.
It achieved nontrivial steady states representing a smooth
model of level curves in a noisy image. Li et al. have
defined a region-scalable fitting energy to be incorpo-
rated into a variational level set formulation with a regu-
larization term
[5]
, from which a curve evolution equation
is derived for energy minimization. This region-based
model can cope with intensity inhomogeneity, preserve
regularity, and avoid expensive reinitialization.
All the above approaches concentrated on the selection,
modification, and regularization of the speed function, a
term of level set formulation determining the diffusion
of the moving interface during the evolution. Actually,
there also exist other methods to detect edge informa-
tion that ensure contrast preservation in the process of
noise removal
[6,7]
using the level set method. For exam-
ple, combining the advantages of an edge detector and
controlling the evolution direction are both promising
choices. In this letter, a modified Canny operator-based
level set method for image noise removal is presented.
Canny operator
[8]
is robust to noise and is probably the
most widely used edge detector. Unless the preconditions
are particularly suitable, it is difficult to find an edge
detector, which performs significantly better than the
Canny operator. Researchers have recently introduced
the Canny operator into the level set method. Heydar-
ian et al. have developed a semi-automated method to
determine the desired object boundary in magnetic reso-
nance (MR) and computed tomography (CT) images
[9]
.
Rough object boundaries can be obtained manually by
applying the Canny operator. Then, using the output of
the genetic algorithm to fix the level set method param-
eters, accurate object boundaries can be detected auto-
matically. Xia et al. have presented an optimal initial-
ization scheme to improve the segmentation performance
of the Chan-Vese model
[10]
. They first used the Canny
operator to compute rough edges. By connecting edge
points iteratively according to a local cost function, the
final closed object contours have been generated using
a morphological filter to remove noise and redundant
edges. Qin et al. have used the maximum magnitude
of edge gradient, which is the result of Canny processing,
to replace the curve plane in the level set formulation and
continuously evolve the moving interface for the precise
segmentation of medical images
[11]
. This algorithm es-
sentially produced rough initial contours with the Canny
operator and then segmented the images using the level
set method.
The two-stage framework, in which the Canny oper-
ator works for the initial conditions and the level set
method works for the final results, has been applied in
the previous studies mentioned above. This letter, how-
ever, adopts a different approach, in which the normal
to edge direction is used to orient the level set evolu-
tion based on the idea of the Canny operator. This re-
sulted in more preserved local edge information due to
the penalized energy function. The key point is that the
abovementioned operations can be executed in one pro-
cess during the image noise removal.
The level set method is based on partial differential
equation. Its main idea is that the moving interface is
viewed as a zero level set in one higher dimensional space.
1671-7694/2010/121127-04
c
° 2010 Chinese Optics Letters