Feature-Oriented Coupled Bidirectional Flow for

Image Denoising and Edge Sharpening

Shujun Fu1,2*, Qiuqi Ruan2, Yuliang Geng3, Wenqia Wang1

1. School of Mathematics and System Science, Shandong University, Jinan, 250100, China
2. Institute of Information Science, Beijing Jiaotong University, Beijing, 100044, China
3. Institute of Computer Science, Beijing Jiaotong University, Beijing, 100044, China
*E-mail: shujunfu@163.com

Abstract—In this paper, a new type of diffusion process that
simultaneously denoises and sharpens images is considered. We
presents a feature-oriented coupled bidirectional flow process,
where, according to image features such as edges, textures, and
fine parts, it can switch from a forward diffusion to a backward
(inverse) one along the normal directions to the isophote lines
(edges), while a forward diffusion is performed along the tangent
directions. To eliminate the conflict between the backward and the
forward force, we split them into a coupled scheme. In order to
enhance image features the nonlinear diffusion coefficients are
locally adjusted according to the directional derivatives of the
image. Experimental results demonstrate that our algorithm can
substantially enhance features on denoising smoother areas of the
image.
Index Terms—bidirectional diffusion, directional derivative,
edge sharpening, image denoising, image enhancement.
1. INTRODUCTION
Image denoising and enhancement are important opera- tions in
image processing and computer vision. The success of many
applications depends often on the results of these operations,
where a challenge is to preserve and enhance important fea-
tures during the denoising process. For images, for example, an
edge is one of the most universal and crucial features.
A widely used method for noise elimination is the Gaus-
sian filter, where an image is smoothed out by the convolution
of it with a Gaussian kernel. Nonetheless, the Gaussian opera-
tor is isotropic and therefore smoothes the image in all direc-
tions blurring sharp boundaries. Denoising via linear filters
normally does not perform satisfactorily since both noise and
edges contain high frequencies. Therefore, any practical de-
noising model has to be non-linear.
The scale–space approach and other partial differential
equations (PDEs) techniques have been extensively applied
over the last decade in signal and image processing [1]-[4].
Actually the diffusion process (or heat equation), widely used
in this context, is equivalent to a smoothing process with a
Gaussian kernel. A major drawback of the linear scale-space
framework is its uniform filtering of local image features and
noise. This problem was addressed by Perona and Malik (P-M)
[5], who proposed a non-linear anisotropic diffusion process,
where diffusion can take place with a variable diffusion in or-
der to preserve singularities, such as step jumps or edges and
thin lines in images. Afterwards many techniques have been
introduceed to improve the idea of anisotropic diffusion filters
[6]–[11].
However, above diffusion processes are mainly based on a
forward diffusion form. In this paper, by adding an inverse one
to the diffusion formalisms, we present a unified bidirectional
flow (BDF) process, and apply it to image denoising and edge
sharpening. To eliminate the conflict between the backward
and the forward force, we split BDF into the coupled bidirec-
tional flow (CBDF) scheme. At the same time, to enhance im-
age features we design properly the diffusion coefficients using
the first and second directional derivatives of the image. Fi-
nally we implement the scheme and test it on real images.
2. A UNIFIED BIDIRECTIONAL FLOW
An important contribution for edge preservation on denoising
images was introduced by P-M, which substituted the heat
equation by an anisotropic diffusion equation. The grey levels
of an image u ( x, y, t ): Ω × [ 0, +∞ ) → R , are diffused according
to:
 ∂u ( x, y, t ) = div ( g ( ∇u ( x, y, t ) )∇u ( x, y, t ))
. (1)
∂t
The scalar diffusivity g ( ∇u ) , chosen as a non-increasing
function, governs the behaviour of the diffusion process. A
typical choice for the diffusivity function is:
2 isotropic diffusion is practiced in homogeneous domains, while
g ( ∇u ) = 1 (1 + ( ∇u K ) ) ,
with K some gradient threshold. By formally developing the
divergence term, (1) can be put in terms of second order de-
G
rivatives taken in the directions of the gradient vectors ( n ) and
G
in the orthogonal tangent ones ( t ):
∂u = ( K 2 ( K 2 − ∇u 2 ) ( K 2 + ∇u 2 ) 2 )u
∂t nn
2
+ ( K 2 ( K 2 + ∇u ))utt .
This expression allows an easier interpretation of the original
equation: (1) acts like a low pass filter diffusing along the edge
directions and, selectively, for diffusion functions as (2), can
preserve edges without diffusing across edges or even enhance
them provided that their gradient value is greater than K . Re-
sults obtained with the P-M process paved the way for a variety
of PDE-based methods that were applied to various problems
in low-level vision.
Another PDE-based enhancement process were pro- posed
by L. Alvarez and L. Mazorra [12], which couples an anisot-
ropic diffusion with the shock filter (we call it ADSF) of S. J.
Osher and L. I. Rudin [13], yielding an equation of the form:
∂u = − sign(G ∗ u ) ∇u + cu ,
σ (4)
∂t nn tt
where c is a positive constant, G σ is a Gaussian of standard
deviation σ . The first term on the right side creates results
approaching piecewise constant regions separated by shocks at
the zero-crossings of the smoothed second derivative of the
K
image along n . The second term is an anisotropic diffusion
K
along the level-set lines t .
Noticing the expression:
sign(s) = s s , s ≠ 0,
now we can define a unified bidirectional flow (BDF) equation
covering both (3) and (4):
∂u = α ( −c (u , u , u )u ) + β (c (u , u , u )u )
tt ,
∂t n n nn tt nn t n nn tt
where α , β are the backward and forward flow control coeffi-
cients, cn ( s ), ct ( s ) are diffusion coefficients of their arguments,
which should be properly designed to enhance features of the
image such as edges, textures and fine parts on denoising
smoother areas of the image.
Aiming at different domains in an image, we hope that an
(2)
an anisotropic diffusion is done in domains of edges and local
details, which diffuses inversely along the normal direction of
edges, and diffuses normally along the tangent one. Therefore,
we need two opposing forces of diffusion, acting simultane-
ously on the image: one is a backward force, to sharpen edges
K
along n , and the other is a forward one, used for reducing
(3) noise and suppressing jaggies and oscillations to smooth con-
K
tours along t (see Fig.1).
Fig.1 Edge sharpening process (the solid line), compared with
original edge (the broken line).
3. FEATURE-ORIENTED COUPLED BIDIRECTIONAL
FLOW
We divide an image into three parts by its smoothed gradient
magnitude: big gradients (such as boundaries of different ob-
jects), middle gradients (such as textures, corners and fine parts)
and small gradients (such as smoother segments inside differ-
ent areas). For big gradients, we enhance the image by (4). For
middle gradients, however, in (4) indicating edges by the
zero-crossing is a binary decision process, by which, unfortu-
nately, the obtained result is a false piecewise constant image
whose texture and fine details are lost (see Fig.2). We substitute
sign ( s ) by a hyperbolic tangent function th (s ) , controlling
(5) softly the variety of gray levels of the image beside the edge
center. For small gradients, we enhance the image by the iso-
tropic diffusion.
Based on preceding discussion, on implementing itera-
tively (6) we find that the backward and forward force will
(6)
cancel mutually in a single formula. Here we split (6) into two
formulas and propose the following coupled bidirectional flow
(CBDF) scheme by iterating with time steps: We compare CBDF with different methods: Gaussian
smoothing, P-M, ADSF, and BDF. From Fig.3 we can see that
⎧ v 0 = u 0 , w = Gσ ∗ u ;
⎪ n +1 n n n Gaussian smoothing results in much blurred edges. P-M de-
⎪ v = u + ∆t Ln (u ) ;
⎪ u n +1 = v n +1 + ∆t Lnt (v n +1) ; noises the image well, specially in the smoother segments, but
⎪ it still produces blurred edges, whose ability to sharpen edges
⎪ ⎧ α (−cnsign (wnn )) , wn >th1 (7)
⎨ ⎪ is not very great. Though ADSF sharpenes edges very well, it
⎪ Ln (u ) = ⎨ α (−cn th( l wnn )) , th2< wn ≤ th1 ; cannot reduce noise well only by a single directional diffusion
⎪ ⎪ ∆u , else
⎩
⎪ in the smoother areas, and a binary decision process results in a
⎪ ⎧ β (ct vtt ) , wn > th2
⎪ Lt ( v ) = ⎨ 0 , else .
false piecewise constant image which looks unnatural with a
⎩ ⎩
discontinueous transition between two different areas (see the
with Neumann boundary condition, time step ∆t , original im- face, the camera and the trousers in Fig.3). Both BDF and
0
age u . By computing iteratively in the order of u → v →
0 0
CBDF denoise very well in the smoother areas, and they
v1 → u1 → v 2 → u 2 → " , we finally obtain the enhanced im- sharpen edges as powerfully as ADSF does. However, CBDF
age after some steps. produces sharper edges than BDF by using a splitting scheme.
In addition, we find that overshoots or ringing artifacts Moreover, employing the fuzzy hyperbolic tangent function to
appear on edges in the experiments. Thus, to decide diffusion control the flow yields a natural transition between two differ-
speed cn only by the gradient information does not work. Be- ent areas. In a word, it is obvious that the best visual quality is
cause we see that the backward flow means “flowing away obtained using CBDF, which enhances most features of the
from the local mean”, which manifests itself as increasing of image and produces pleasing sharper edges on denoising effec-
unn at overshoot pixels more and more largely with iteration tively.
times, we add the second derivative information to suppress
5. CONCLUSIONS
this plague:
2
cn = un (1 + l1unn ). (8) This paper presents a feature-oriented coupled bidirectional
flow process to enhance images, by which we not only can
At the same time, we can see that utt at the angle is much
denoise images well, but also can effectively sharpen edges.
bigger than that on the edge in value along the tangent direc-
Enhancing image features such as edges, corners and fine parts
tion, here we add the second derivative information to ct to
with a natural transition between two different areas, this
prevent over smoothness to corners and small details:
method produces better visual results than some relative diffu-
ct = 1 (1 + l2utt2 ) . (9) sion equations.

4. EXPERIMENTAL RESULTS

We used the explicit Euler method with the central difference

scheme. A number of images have been used to test our scheme
(7)-(9). In Fig.2, we take a Cameraman image for example to
evaluate the performance of our scheme. In Fig.3, a noisy
blurred cameraman image, which is corrupted by a Gaussian
white noise and a Gaussian blur, is enhanced using different
methods, where we adopt the following parameters: Gaussian
smoothing, 5 × 5 mask, σ = 2 ; P-M, K = 9 ; ADSF, c = 2 ;
BDF and CBDF, [ l , l1 , l2 ]=[200, 7.7 × 10-4, 2.3 × 10-4],
[ α , β ]=[1, 1], [th1, th2]=[30, 2.5]. All parameters have been
optimized to obtain the best results. Fig.2 Test image: the Cameraman image.
 (a) (b)

(c) (d)

(e) (f)
Fig.3 CBDF processing of Cameraman image, compared with others (from top left to bottom right): (a) noisy blurred image,
(b) Gaussian smoothing, results by (c) P-M, (d) ADSF, (e) BDF, and (f) CBDF.

ACKNOWLEDGMENTS istry of Railways, China (No. TDXX0510), the Technological

Innovation Fund of Excellent Doctorial Candidate of Beijing
This work is supported by the Key Laboratory Project of In- Jiaotong University, China (No. 48007), and the National
formation Science & Engineering of Railway of National Min- Natural Science Fund, China (No. 60472033).
 REFERENCES
[1] G. Aubert , P. Kornprobst, Mathematical Problems in Image Processing:
Partial Differential Equations and the Calculus of Variations, Applied
Mathematical Sciences volume 147, Springer-Verlag, 2001.
[2] G. Sapiro, Geometric Partial Differential Equations and Image Process-
ing, Cambridge University Press, Cambridge, UK, 2001.
[3] D. Tschumperle, “PDE’s Based Regularization of Multivalued Images
and Applications”, phD thesis, Universite de Nice-Sophia Antipolis,
France, 2002.
[4] D. Tschumperle, R. Deriche, “Vector-Valued Image Reg- ularization with
PDE’s: A Common Framework for Different Applications”, Proc. IEEE
CVPR, 2003.
[5] P. Perona, J. Malik, “Scale-space and edge detection using anisotropic
diffusion”, IEEE Trans. Pattern Anal. Machine Intell, 12(7): 629–639,
1990.
[6] F. Catte, P. L. Lions, J. M. Morel, and T. Coll, “Image selective smooth-
ing and edge detection by nonlinear diffusion”, SIAM J. Numer. Anal.,
vol. 29, no. 1, pp. 182–193, 1992.
[7] L. Alvarez, P. L. Lions, and J. M. Morel, “Image selective smoothing and
edge detection by nonlinear diffusion”, SIAM J. Numer. Anal., vol. 29,
no. 3, pp. 845–866, 1992.
[8] L. Rudin, S. Osher, and E. Fatemi, “Nonlinear total variation based noise
removal algorithms”, Phys. D, vol. 60, pp. 259–268, 1992.
[9] Y. L. You and M. Kaveh, “Fouth-order partial differential equations for
noise removal”, IEEE Trans. Image Processing, vol. 9, pp. 1723–1730,
Oct. 2000.
[10] P. Kornprobst, R. Deriche and G. Aubert, “Image coupling, restoration
and enhancement via PDE’s”, Proceedings of the International Confer-
ence on Image Processing, Santa Barbara Santa-Barbara (USA), pp.
458-461, 1997.
[11] J. Weickert, “Coherence-enhancing diffusion of color imag- es”, Image
Vis. Comput., vol. 17, pp. 199–210, 1999.
[12] L. Alvarez and L. Mazorra, “Signal and image restoration using shock
filters and anisotropic diffusion”, SIAM J. Numer. Anal., 31(2): 590–605,
1994.
[13] S. J. Osher and L. I. Rudin, “Feature-oriented image enhancement using
shock filters”, SIAM J. Numer. Anal., vol. 27, pp. 919–940, 1990.