Fast gradient vector flow computation based on augmented Lagrangian method
Dongwei Ren
a
, Wangmeng Zuo
a,
⇑
, Xiaofei Zhao
a
, Zhouchen Lin
b
, David Zhang
a,c
a
Biocomputing Research Centre, School of Computer Science and Technology, Harbin Institute of Technology, Harbin, 150001, China
b
Key Laboratory of Machine Perception (MOE), School of EECS, Peking University, Beijing, 100871, China
c
Biometrics Research Centre, Department of Computing, The Hong Kong Polytechnic University, Kowloon, Hong Kong
article info
Article history:
Received 9 March 2012
Available online 4 October 2012
Communicated by G. Borgefors
Keywords:
Gradient vector flow
Convex optimization
Augmented Lagrangian method
Fast Fourier transform
Multiresolution method
abstract
Gradient vector flow (GVF) and generalized GVF (GGVF) have been widely applied in many image pro-
cessing applications. The high cost of GVF/GGVF computation, however, has restricted their potential
applications on images with large size. Motivated by progress in fast image restoration algorithms, we
reformulate the GVF/GGVF computation problem using the convex optimization model with equality
constraint, and solve it using the inexact augmented Lagrangian method (IALM). With fast Fourier trans-
form (FFT), we provide two novel simple and efficient algorithms for GVF/GGVF computation, respec-
tively. To further improve the computational efficiency, the multiresolution approach is adopted to
perform the GVF/GGVF computation in a coarse-to-fine manner. Experimental results show that the pro-
posed methods can improve the computational speed of the original GVF/GGVF by one or two order of
magnitude, and are more efficient than the state-of-the-art methods for GVF/GGVF computation.
Ó 2012 Elsevier B.V. All rights reserved.
1. Introduction
Gradient vector flow (GVF) field (Xu and Prince, 1998a) was
first introduced as a new external force to address the two key
problems in parametric active contour model (ACM) (Kass et al.,
1987): the small capture range of the external forces and difficul-
ties of progressing into boundary concavities. Xu and Prince
(1998b) further proposed a generalized GVF (GGVF) external field
to improve the convergence to long thin boundary indentations
by incorporating two spatially varying weights. Besides parametric
ACM, GVF had also been adopted by Paragios et al. (2004) in geo-
metric ACM for image segmentation.
Moreover, the applications of GVF can be extended to other im-
age processing tasks, e.g., tracking, denoising, restoration, and skel-
etonization. In (Ray and Acton, 2004), by embedding the motion
direction in the GVF energy functional using a regularized Heavi-
side function, Ray and Acton proposed a motion GVF external force
for tracking rolling leukocyte. In (Yu and Chua, 2006), Yu and Chua
introduced GVF field in the field of image restoration to reformu-
late several popular anisotropic diffusion models, e.g., shock filter,
mean curvature flow, and Perona–Malik model, to obtain better
robustness against noise and spurious edges with improved high-
order derivative estimation. In (He et al., 2008), GVF was used as
new intensity diffusion direction for better color photo denoising.
In (Ghita and Whelan, 2010), Ghita and Whelan proposed a new
GVF field formulation for adaptive denoising of mixed noise. In
(Hassouna and Farag, 2009), Hassouna and Farag incorporated
GVF in the variational skeleton model by introducing modified
GVF medial function to improve the accuracy and robustness of
the curve skeleton method.
Despite its success and popularity, the GVF method requires a
high computational cost, which has restricted their potential
applications to images with large sizes. One possible solution is
to develop alternative external forces which can be efficiently com-
puted. For example, Li and Acton (2007) proposed a vector field
convolution (VFC) external force. Another solution is to design effi-
cient numerical schemes for fast GVF computation. In (Ntalianis
et al., 2001), Ntalianis et al. proposed a multiresolution implemen-
tation of GVF computation for video object segmentation. In (Bou-
kerroui, 2009), Boukerroui compared several efficient numerical
schemes for GVF computation, and showed that the alternating
direction explicit scheme (ADES) may be a suitable alternative to
the multigrid method. Recently, Han et al. (2007) proposed a mul-
tigrid GVF/GGVF (MGVF/MGGVF) algorithm, which can signifi-
cantly improve the computational efficiency. To the best of our
knowledge, MGVF/MGGVF are the most efficient schemes for
GVF/GGVF computation.
In this paper, we propose a novel fast GVF/GGVF computation
scheme based on the augmented Lagrangian method (ALM). We
reformulate GVF as a constrained convex optimization problem,
and use an efficient optimization scheme, i.e., ALM (Afonso et al.,
2010; Wang et al., 2008) with variable splitting method (Liu
et al., 2010), for fast GVF and GGVF computation. Part of the paper
had been presented in (Li et al., 2011). In this paper, we further
proposed different variable splitting methods for GVF and GGVF
0167-8655/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.patrec.2012.09.017
⇑
Corresponding author. Tel.: +86 451 86412871.
E-mail address: cswmzuo@gmail.com (W. Zuo).
Pattern Recognition Letters 34 (2013) 219–225
Contents lists available at SciVerse ScienceDirect
Pattern Recognition Letters
journal homepage: www.elsevier.com/locate/patrec