See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/262290979

Edgemap-based Wiener filtering for preserving image fine details and edges

Conference Paper · July 2011

CITATIONS READS

3 133

2 authors:

Suhaila Sari Tetsuya Shimamura

Universiti Tun Hussein Onn Malaysia Saitama University
30 PUBLICATIONS 60 CITATIONS 267 PUBLICATIONS 1,351 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Channel estimation and equalization View project

Computer graphics View project

All content following this page was uploaded by Tetsuya Shimamura on 12 February 2015.

The user has requested enhancement of the downloaded file.

 Recent Researches in System Science

Edgemap-Based Wiener Filtering for Preserving

Image Fine Details and Edges
SUHAILA SARI and TETSUYA SHIMAMURA
Graduate School of Science and Engineering
Saitama University
Shimo-Okubo 255, Sakura-ku, Saitama, 338-8570
JAPAN
{suhaila, shima}@sie.ics.saitama-u.ac.jp

Abstract: - In this paper, we present a denoising technique that is capable for preserving the fine details and
edges in the restored image more effectively in blind condition. We also introduce a new edge detection method
to detect edges effectively in noisy environments. First, the noisy image is denoised by using different weights
of Wiener filtering to generate two restored images; one with highly reduced noise, and the other with
preserved fine details and edges. The noise and image power spectra required for the frequency domain Wiener
filter are estimated with different threshold setting. Then, an edgemap image is generated directly from the
noisy image. The two Wiener filtered images are utilized for the smooth and non-smooth regions based on the
constructed edgemap to produce the final restored image. Simulation results show that the proposed method
outperforms or is comparable to other Wiener filter-based denoising methods and the state-of-the-art denosing
methods, especially in higher noise environments.

Key-Words: - Edgemap, Edge detection, Image denoising, Power spectrum estimation, Wiener filter

1 Introduction important role to accomplish accurate denoising.

Although image denoising has been researched quite The ATV derives the idea of the Total Variation
extensively, developing a denoising method that (TV) [7]. The ATV reduces the total variation of the
could remove noise effectively without eliminating image adaptively. It employs strong denoising in the
the image fine details and edges is still a challenging smooth regions and weak denoising in the non-
task. Until recent years, many denoising methods smooth regions. The NLM measures the similarity
have been proposed. Some recent non-linear of the grey level between two pixels and compares
methods, such as the adaptive Total Variation (ATV) the geometrical configuration adapted to the local
[1] and the non-local means (NLM) [2], suggest and non-local geometry of the whole image. The
employing different denoising approaches for the methods such as the LBWDWF, ATV and NLM are
smooth and non-smooth regions. Conversely, linear reported to have superior performance in noise
methods such as the Wiener filter [3] balance the removal and preservation of strong edges. They,
tradeoff between inverse filtering and noise however, share a common drawback: that is, the fine
smoothing by eliminating additive noise while details and edges of the original image are not well
inverting blurring. preserved in the restored image, especially in higher
The Wiener filter is the best-known technique for noise environments.
the linear image denoising. It has been implemented To overcome this problem, a frequency domain
for image denoising in several transform domains, Wiener filter-based denoising has been proposed in
for example the spatial domain [4] and the [8]. We refer this method to as the frequency domain
frequency domain [5]. Recently, the wavelet-based Wiener filter (FDWF). It introduces a noise and
denoising methods have dominated the latest image power spectra estimation method for the
research trend in image processing. The implementation of the Wiener filter in blind
implementation of the Wiener filter in the wavelet condition. The FDWF provides the preservation of
domain has also been introduced, for example the the fine details and edges, but a certain level of
lifting-based wavelet domain Wiener filter noise still remains in the restored image.
(LBWDWF) [6]. However, to utilize the Wiener In this paper, we propose a denoising technique
filter in practical cases, where the information of the that is based on the FDWF. The image restored by
original image and the noise level are unknown using the FDWF with a lower threshold value is
(blind condition), noise estimation plays an utilized in the non-smooth regions in the final
restored image. Conversely, the image restored by

ISBN: 978-1-61804-023-7 264

 Recent Researches in System Science

using the FDWF with a higher threshold value is the image’s sub-blocks the that represent
employed in the smooth regions in the final restored the median and minimum values of the entire
image. A new edge detection method is employed to The minimum value, and the
distinguish between the smooth and non-smooth
regions effectively in the presence of noise. The maximum value, are substituted in the
edge detection is performed in four directions and following global threshold value as
the results are combined to construct an edgemap.
The final restored image is constructed by assigning
the smooth and non-smooth regions based on the
edgemap. Simulation results verify a significant
reduction of the noise level in the smooth regions where denotes a division ratio of , and
relative to that of the FDWF. corresponds to the threshold used for the power
The paper is organized as follows. We begin with spectrum estimation.
the introduction of the FDWF in Section 2, and then
The utilization of in (4) is attributed to
describe the proposed denoising method in Section
3. In Section 4, we discuss the simulation results the fact that the median represents where most of
and the performance comparison of our method. In the power spectrum are concentrated. If the division
Section 5, we draw concluding remarks. of the high and low frequencies considers the main
power concentration, the threshold value will be
more robust to the variation of the power spectrum
characteristics in different images. The threshold
2 FDWF
value is set to be slightly higher than to avoid
Our procedure for image denoising utilizes the
the probability of the noise power spectrum being
FDWF proposed in [8]. We assume that the image is
incorrectly thresholded into the image power
corrupted by independent additive zero-mean
Gaussian white noise. A noisy image, p(u,v), spectrum. In (4), is utilized to adjust the threshold
range.
corrupted by the noise, n(u,v), can be expressed as
The noise power spectrum is estimated from both
p(u,v) = h(u,v) + n(u,v) (1) high and low frequency regions, since we consider
that the noise occupies both regions. In high
where h(u,v) represents the original image. The frequency region, the image power spectrum,
FDWF employs a threshold process to estimate the and the noise power spectrum,
image and noise power spectra. The assumption is in the corresponding sub-block are approximated by
that in general, the noise power spectrum usually
occupies high frequencies, and conversely the image if
power spectrum is commonly concentrated at low then
frequencies.
First, we transform the noisy image h(u,v) to the
In low frequency region, and are
frequency domain, ɦ(s,t), by using the FFT. The
estimated as
power spectrum of ɦ is obtained by
if
ɦ (2)
then
and the logarithmic power spectrum of average
is given by

(3) where the average [] represents the averaging of

the sub-blocks in the four corners being at the
We perform the estimation for the power spectra of highest frequencies. These four blocks are assumed
the image and noise block-by-block. and to be occupied only by noise. Fig.1 demonstrates the
are divided into non-overlapping sub- described four corners in the case of k=8.
blocks. and correspond to the Finally, we perform the Wiener filtering
(i,j)th block of and respectively. Next, operation. The Wiener filter, is obtained by
we compute the average of the logarithmic power
spectrum in each sub-block, which is
denoted as , respectively. We find from all

ISBN: 978-1-61804-023-7 265

 Recent Researches in System Science

In [8], the parameters k and are set to be 32 and

5, respectively. This is due to that when =5, it
provides a robust and optimal restoration result for
the FDWF in different image characteristics and
noise levels. When is fixed to be lower (for
example =5), the threshold range will be larger.
This setting will include higher frequencies in the
estimated image power spectrum. This will
contribute to the preservation of the image fine Fig.1 Average of four corners (highest frequencies).
details and edges that usually occupy higher
frequencies. Fig.2 shows the close-up view of the
restored Cameraman image and its logarithmic
power spectrum (corrupted by the white noise with
the standard deviation, σ, of 25) by using the FDWF
with =5. From Fig. 2 we can observe that the
FWDF has preserved fine details and edges
successfully, but has not effectively eliminated the (a) (b)
noise in the restored image. Fig.2 Restored Cameraman and its logarithmic
power spectrum (σ=25) by using FDWF ( =5).

3 Proposed Algorithm
The method proposed in this paper reduces the noise
in the image restored by using the FDWF with
different parameters in the smooth and non-smooth
regions. From our investigation, narrower threshold
range setting in the FDWF reduces noise level, and
larger threshold range setting preserves the fine
details and edges. We set out to improve the FDWF
restoration performance by utilizing the advantage
of both threshold settings. Fig.3 shows a block
diagram of our method.

3.1 Image Restoration for Non-Smooth

Regions
The image restored by the FDWF with =5 is
inverse-transformed to the spatial domain and
represented as hereafter. The is
utilized for the non-smooth regions in the final Fig.3 Block diagram of proposed method.
restored image, since it preserves the fine details
and edges effectively.

3.2 Image Restoration for Smooth Regions

Conversely, an image with highly reduced noise,
which is restored by using the FDWF with higher
setting, is employed for the smooth regions in the (a) (b)
final restored image. If is set to be higher, the Fig.4 Restored Cameraman and its logarithmic
threshold range of will be narrower. This will power spectrum (σ=25) by using FDWF ( =10).
allow the FDWF to threshold only the concentrated
part of the low frequencies, which is assumed to be selected by visual effects evaluation. From the
occupied only by the image power spectrum. investigation, is set to 10. From Fig.4 we can
First, the most suitable setting that provides a observe that the noise level has been effectively
restored image with considerably low noise level is reduced since noise is mostly cut out. On the other

ISBN: 978-1-61804-023-7 266

 Recent Researches in System Science

hand, the image is blurred since the edges in higher

frequencies are cut out as well. Obvious ringing
effects can be observed in the restored image, which
is caused by the sharp cutoff of the threshold
range’s edges in the frequency domain When the
image is inverted into the spatial domain, it
generates decreasing oscillations as it progresses
(a) (b)
outward from the center. Narrower threshold range Fig.5 Result for restored Cameraman (σ=25) by
results in stronger oscillations. using FDWF ( =10) multiplied with Gaussian
To overcome the ringing effects, we multiply the lowpass filter.
output power spectrum filtered by the Wiener filter
with a Gaussian lowpass filter, to soften the
threshold edges. The restored image is inverse-
transformed to the spatial domain and denoted as
In Fig.5, the ringing effects as seen in
Fig.4 are successfully suppressed. From our
investigation, the best parameter settings for the
Gaussian lowpass filter are fixed to the size of 128 (a) horizontal (b) vertical c ˚
128 with the standard deviation of 10.

3.3 Edgemap Construction

Next, the decomposition of the smooth and non-
smooth regions is performed pixel-by-pixel based
on an edgemap. The edgemap is constructed by d − ˚ (e) edgemap(u,v) (f) noisy Lenna
executing edge detection directly from the noisy Fig.6. Edgemap constructions in four directions and
image in the spatial domain. Although there is no combined edgemap for noisy Lenna (σ=25)
pre-filtering process, the edges have been
successfully obtained by using the proposed edge non-smooth region belongs to a line along the
detection method. Another advantage of this edges or additive noise, we consider the
technique is that it requires minimum parameter corresponding pixel and its two neighbouring
setting, which is only the sub-block size for the pixels. If either difference of the (u,v)th pixel and
image division. The process of the edgemap its neighbouring pixels is larger than R, this pixel is
construction is as follows. considered as a part of a line along the edges. If
First, the noisy image is divided into k k non- both differences are lower than R, the
overlapping sub-blocks (k=32), each of which corresponding pixel is assumed isolated from other
consists of 8 8 pixels. Next, the pixel value range, non-smooth region’s pixel and considered as noise.
of each block (i,j: 2 … k) is calculated The edge detection along a direction can provide a
as more precise distinction between edges and noise
in noisy environments because isolated pixels can
- 2 be detected easier. The edge detection is performed
in four directions; (a) horizontal, (b)
where and are the maximum and vertical, (c) ˚ and d − ˚
minimum values of the pixels in the corresponding .
sub-block, respectively. Then, the that
represents the minimum value of the
entire denoted as R, is determined by or

min
or
This sub-block is assumed to be homogeneous and
represents the smooth region. Thus, any pixel
larger than R is considered as the non-smooth
region. In order to determine whether a pixel in the

ISBN: 978-1-61804-023-7 267

 Recent Researches in System Science

or human visual perception [9]. We also analyze the

visual effects of the images. Each denoising method
2 processes the test images with the same parameters
setting (no tuning of parameters was performed for
different noise level or image type).
or We have compared our method to two Wiener
filters in different domains: the LBWDWF [6] and
FDWF [8] in blind condition. In this paper, the
As can be observed from Fig. 6 ((a)-(d)), the edges LBWDWF with 4 vanishing moments (db4) lifting-
have been successfully detected from different based wavelet transform utilizes 3 decomposition
directions, regardless of the high noise level (see levels and 3 3 filtering window. The FDWF
Fig. 6(f)). However, the edge detection from only a employs the setting as in Section 2. From Fig. 7(b)-
single direction is not sufficient to represent all the (d) and Tables 1-3, our method clearly outperforms
edges for the entire image. Thus, we suggest the other two methods over the entire range of noise
combining the edge detections of all four directions levels. Our method is capable to reduce more noise
for better result. The combined edgemap, compared to the FDWF and provides better
, is constructed by preservation of the original image features than that
of the LBWDWF.
= Finally, we have also compared our method to
or or or two state-of-the-art methods: the ATV [1] and NLM
[2]. They are reported to have significant
performance in preserving details while eliminating
Fig.6(e) shows the edge detection of the the noise. Both denoising methods are performed in
Note that the provides ideal condition, where the noise variances are
relatively better edge detection along the lines in known. Conversely, our method estimates the noise
comparison with that of the single direction edge employing the approach as in Section 3. We
detection. perform the ATV based on the MATLAB code
(default setting) as in [1]. In [2], the NLM suggested
using the search window size of 21 21 and
3.4 Edgemap-Based Image Restoration similarity window of 7 7 for images with the size
The final restored image, , is constructed based of 512 512. However, the parameters setting as in
on the edgemap as [2] results in over-smoothed restored images for
small resolution test image (256 256). Thus, we
. (15) set the search window and similarity window for the
NLM to 5 5 and 2 2, respectively for better
denoising results. From Tables 1-3, note that our
If the is equal to 0, then the (u,v)th method is better than the ATV in most cases. Our
pixel of the final restored image is assumed method is also better than the NLM in many images
as the smooth region, and assigned with the (u,v)th that are highly rich in both fine details and edges. In
pixel value of Otherwise, is higher noise level, as can be seen in Fig. 7(d)-(f),
assigned with those of the ATV and NLM result in strong noise removal.
However, they eliminate the fine details and edges
at the same time. Our method reduces considerable
4 Results and Discussion amount of noise and furthermore preserves fine
We have tested our method on nine grayscale test details and edges better than the NLM and ATV.
images (256 256) from the SIDBA database. All The investigation proves that our method with
images are contaminated with additive Gaussian proposed noise estimation technique provides
white noise (σ=5, 10, and 25). The Airplane, Girl, comparable performance with the state-of-the-arts
Lenna, Woman and Boat images represent smooth methods performed in ideal condition. Table 4
natural images. The Barbara, Building, Lighthouse shows the execution time for each denoising method
and Text images represent natural images that are implemented in MATLAB computed on a 1.4 GHz
highly rich in both fine details and edges. We Intel Core 2 Duo CPU. The execution time of our
investigate the performances by using the mean method is slightly higher than that of the
measure of structural similarity (MSSIM) [9] since LBWDWF, but considerably lower if compared to
it is an image quality metric that well matches the that of the ATV and NLM.

ISBN: 978-1-61804-023-7 268

 Recent Researches in System Science

5 Conclusion Table 3.Performance comparison in MSSIM (σ=25)

From the study, it is found that our method is fast LBW FD Propo - ATV NLM
and more suitable for denoising images that are rich Test images DWF WF sed (ideal) (ideal)
in both fine details and edges, especially in higher (blind) (blind) (blind)
noise environments. The proposed edge detection Barbara 0.66 0.82 0.85 0.41 0.84
technique is also capable to detect noise directly Building 0.61 0.84 0.84 0.56 0.76
from the noisy image in any noise level, which has Lighthouse 0.66 0.72 0.79 0.75 0.79
lead to an accurate decomposition of the smooth and Text 0.67 0.82 0.84 0.72 0.80
non-smooth regions. Airplane 0.78 0.70 0.78 0.58 0.86
Girl 0.80 0.77 0.81 0.55 0.82
Lenna 0.81 0.75 0.84 0.5 0.86
Woman 0.79 0.75 0.83 0.47 0.84
Boat 0.76 0.77 0.82 0.52 0.84
Table 4. Average time execution (s)
LBW FD Propo ATV NLM
(a) Original (b) LBWDWF (c) FDWF Noise, σ
DWF WF sed
5 0.60 0.31 0.93 23.33 67.31
10 0.56 0.31 0.92 66.17 68.60
25 0.57 0.30 0.93 751.49 67.82

(d) Proposed (e) ATV (f) NLM References:

Fig. 7. Comparison of restored Lighthouse [1] G. Gilboa, N. Sochen, and Y. Zeevi, Texture
preserving variational denoising using an
(high noise, σ 25). adaptive fidelity term, Proc. IEEE Workshop
Variational and Level Set Methods in Computer
Table 1.Performance comparison in MSSIM (σ 5) Vision,pp.137-144,2003. http://visl.technion. ac.
il/~gilboa/PDE-filt/demo.adap.tv.m.
Test LBW FD Propo ATV NLM [2] A. Buades, B. Coll, and J.M. Morel, A non-
images DWF WF -sed (ideal) (ideal) local algorithm for image denoising, Proc.
(blind) (blind) (blind) IEEE Int. Conf. Computer Vision and Pattern
Recognition,vol.2,pp.60-65, 2005.
Barbara 0.67 0.95 0.96 0.94 0.97 [3] N. Wiener, Extrapolation, Interpolation, and
Building 0.62 0.96 0.96 0.96 0.96 Smoothing of Stationary Time Series, Wiley,
Lighthouse 0.68 0.91 0.94 0.90 0.93 1949.
[4] J. S. Lee, Digital image enhancement and noise
Text 0.68 0.95 0.96 0.95 0.93 filtering by use of local statistics, IEEE Trans.
Airplane 0.81 0.89 0.92 0.95 0.94 Pattern Anal. Machine Intell., vol. PAMI-2,
Girl 0.82 0.91 0.92 0.91 0.93 no.2,pp.165-168, 1980.
[5] H. Furuya, S. Eda, and T. Shimamura, Image
Lenna 0.83 0.91 0.94 0.96 0.96 restoration via Wiener filtering in the frequency
Woman 0.81 0.91 0.94 0.95 0.95 domain, WSEAS Trans. Signal Process.,
vol.5,no.2, pp.63-73, 2009.
Boat 0.79 0.93 0.94 0.93 0.96 [6] E. Ercelebi and S. Koc, Lifting-based wavelet
Table 2.Performance comparison in MSSIM (σ 10) domain adaptive Wiener filter for image
LBW FD Propo ATV NLM enhancement, IEE Proc. Vis. Image and Signal
Test Processing,vol.153,no.1,pp.31-36, 2006.
DWF WF -sed (ideal) (ideal) [7] L. Rudin, S. Osher, and E. Fatemi, Nonlinear
images
(blind) (blind) (blind) total variation based noise removal algorithms,
Barbara 0.66 0.88 0.90 0.90 0.93 Physica D,vol.60,pp.259- 268,1992.
[8] S. Suhaila, and T. Shimamura, Power spectrum
Building 0.62 0.89 0.91 0.90 0.91 estimation method for image denoising by
Lighthouse 0.67 0.80 0.86 0.90 0.89 frequency domain Wiener filter, Proc. IEEE Int.
Text 0.68 0.88 0.89 0.90 0.88 Conf. on Computer and Automation
Engineering, vol. 3,pp.608-612, 2010.
Airplane 0.80 0.77 0.85 0.92 0.91 [9] Z. Wang, A. C. Bovik, H.R. Sheikh, and E. P.
Girl 0.81 0.83 0.87 0.81 0.89 Simoncelli, Image quality assessment: from
error visibility to structural similarity, IEEE
Lenna 0.82 0.82 0.88 0.92 0.92 Trans. Image Process.,vol.13,no.4,pp.600-612,
Woman 0.80 0.82 0.87 0.90 0.91 2004.
Boat 0.78 0.84 0.88 0.86 0.92

ISBN: 978-1-61804-023-7 269

View publication stats