Вы находитесь на странице: 1из 4

A NEW WAVELET-BASED METHOD FOR CONTRASTlEDGE ENHANCEMENT

Jinhui Qin and Muhmoud R. El-Sukku

Computer Science Department, University of Western Ontario


London, Ontario, Canada
fjhqin, elsakka}@csd.uwo.ca
ABSTRACT
Contrast enhancement is usually achieved by histogram
equalizing image pixel gray-levels in the spatial domain to
redistribute them uniformly. Meanwhile, edge enhancement
attempts to emphasize the fine details in the original image. But
in the spatial domain it is hard to selectively enhance details at
different scales. Moreover, in the spatial domain, applying
contrast and edge enhancement techniques in different orders
may yield different enhancement results. To overcome the
above spatial domain enhancement issues, a new wavelet-based
image enhancement method is proposed. The proposed method
histogram-equalizes the approximation-coefficiend.
At the same
time, it high-boost filters the detail-coefficients at selected
resolution levels separately. The experiments show that utilizing
the proposed method can achieve robust contrast and edge
enhancement. Moreover, the computation cost in the wavelet
domain is less than that in the spatial domain. This is especially
b-ue when considering that currently most images are already
wavelet-compressed (the current JPEG 2000 standard is a
wavelet based scheme).

INTRODUCTION
In this research we focus on contrast and edge enhancement
issues [1][2]. Contrast enhancement, in general, is achieved by
1.

utilizing the entire brightness range in a given image. The


histogram of an image usually provides a way to determine
which particular gray scale transformation is required to
enhance the image contrast. Histogram equalization (HEQ) is
one ofthc most useful contrast enhancement schemes. When an
image's histogram is equalized, image pixel values are mapped
to uniformly distributed pixel values, as much as possible.
However, afier applying the HEQ technique to an image in the
spatial domain, multiple gray-levels in the image could be
merged into one level. This results in a loss of small details afier
performing the HEQ procedure.[3][4].
Edge enhancement attempts to emphasize edges, or details, in a
given image. Mask convolution is a commonly used technique
for enhancing edges [2][5]. But in the spatial domain, it is hard
to selectively enhance details at different scales.
Moreover, in the spatial domain, applying contrast and edge
enhancement techniques in different orders may yield different
enhancement results.
Ideally, if we can decompose an image into several components
in multiple resolution levels, where low-pass and high-pass
information are kept separately, then we can enhance the image
contrast without disturbing any details. At the same time, we
can also emphasize image detail, at a desired resolution level,
without disturbing the rest of the image information. Finally, by
adding the enhanced components together, we should get a more

0-7803-7750-8/03/%17.00 02003 IEEE

impressive result. Note that, in this case the enhancement order


'
is irrelevant, since we are dealing with separate components.
The wavelet transform framework provides an opportunity to
achieve these tasks. It provides multiple resolution
representations of a given image, each of which highlights
scale-specific image features [6][7]. Since features in those
wavelet-transformed components remain localized in space,
many spatial domain image enhancement techniques can be
adopted for the wavelet domain.
In this research, a new wavelet-based image enhancement
method is proposed [SI. The proposed enhancement method
consists of two parts. One part is to enhance the image contrast.
The contrast enhancement is achieved by histogram equalizing
the
wavelet
approximation-co~~cienis, which
are
corresponding to the low-pass information of a given image.
The other part is to enhance the image details at a desired
resolution level. This enhancement is achieved by high-boost
filtering (HBF) [2] the wavelet detail-coeflcients, which are
corresponding to the detail information of a given image.

2.

LITERATURE REVIEWS

In literature, there are many attempts lo enhance images in the


wavelet domain, rather than in the spatial domain, to benefit
from the multiple-resolution nature of the wavelet domain.
Chen et al. [9] globally shiRed the intensity value in the
approximation-coefficients to achieve contrast enhancement.
But they didn't provide an efficient way to decide the size of the
shifiing step. They also established a zero-crossing tree, which
consists of zero-crossings of each companent in multipleresolution levels, to represent multiple-resolution edges that are
used to suppress noise.
Fu et al. [3] analyzed the drawback of the HEQ procedure in the
spatial domain. Then they proposed a wavelet-based contrast
enhancement method. In their method, afier performing the
HEQ procedure in the spatial domain, the output image was
transformed into the wavelet domain. Then all approximationcoefficients were squared. They claimed that the proposed
process could compensate for the information that was lost
during the HEQ process.
Reeves et al. [41 investigated a wavelet transform domain filter,
based on the LLMMSE filter [IO] to suppress noise and enhance
edges. They also applied global HEQ to the wavelet
approximation-coefficients at the coarsest decomposition level
to enhance contrast. But further investigation was required in
order to understanding how the selection of the approximationcoefficients' range and histogram bin values affects the
reconstructed image.
Xu et al. [I I] used a wavelet phase filter [I21 at finer scales in
the wavelet domain to reduce noise, and a semi-sotl wavelet

111 - 397

shrinkage technique [I31 at coarse scales in the wavelet domain


to further reduce noise. But the proposed method still could not
automatically adjust its parameters to achieve optimal result.
There are some other image enhancement attempts. For
example, Gong et al. [I41 rationally enlarged coefficients on
multiple scales in the wavelet domain. Xu et al. [I51 altered the
amplitude o f coefficients in the wavelet domain. And Peng et al.
[I61 used a non-linear enhancement operator on coefficients at
multi-scale in the wavelet domain. However. these attempts
aimed at improving only image contrast.

3. PROPOSED WORK
The proposed enhancement method [a] consists of two parts.
One part is to enhance the image contrast, which i s achieved by
histogram equalizing the wavelet approximation-coefficients.
The other part i s to enhance the image details at a desired
resolution level, which is achieved by high-boost filtering the
wavelet detail-coefficients. Fig. I shows the framework of the
proposed scheme.
It is worth mentioning that the implementation of the HEQ
method in the wavelet domain is a little hit different from that in
the spatial domain. This i s because wavelet transformed
coefficients are floating point numbers. Moreover, some of them
are possibly negative. Choosing an appropriate range o f
coefficient values and setting of bins thresholds for HEQ in the
wavelet domain is not an easy task.
We suggest a new way to implement HEQ in the wavelet
domain. First, all coeficients are casted into integer values. The
histogram is then built based on those integer values.just like in
the spatial domain. Second, to equalize this wavelet domain
histogram, a procedure which is simillar to the spatial domain
HEQ is used. Finally, after calculating the difference of each
casted integer value' before and aAer equalization, the final
mapped value of each original coefficient is obtained by adding
the difference to each original coefficient.
Note that, if the original image does not use most of the
available dynamic range, the transformed coefficients will not
use most o f their dynamic range either. Therefore the coefficient
value range over which the HEQ is performed in the wavelet
doamin could he expanded beyond the current range to achieve
a better contrast enhancement result. But the expanded range
should not exceed the boundary values of this component,
otherwise, some artifacts may he introduced to the resulting
image.

4. EXPERIMENT RESULTS
In the proposed method, there are various parameters in each of
the three major functions (wavelet transform, HEQ and HBF in
the wavelet domain functions) that need to he adjusted.
Different parameter selections will affect the enhancement
result. The final parameter selections are based on the following
research results.
4.1. Experimental setup
In this research, the proposed method is tested on many images,
which have various characteristics. In order to accommodate the
size constraints o f this paper, the Lenna image has been chosen
to demonstrate our results. The Lenna image contains a nice

mixture of detail, flat regions, shading, and texture that do a


good job o f testing various image enhancement algorithms.

4.2. Wavelet Transform Function


The purpose o f this function is to perform a forwardinverse
wavelet transform on a given image. After applying a forward
wavelet transform, the given image is decomposed into several
components i o multi-resolution. Note that, using different
wavelet filter Sets andlor different number of transform-levels
will result in differenl decomposition results.
While any wavelet-filters can he used with the proposed
scheme. we restricted our experimentation to the one-level Haar
wavelet-filter, to limit the scope of this research and to focus on
the proposed method only. We left testing the proposed method
on various wavelet filter sets and multiple wavelet levels as a
future work.
4.3. H E Q Function
When applying HEQ in the spatial domain, some small details
may be eliminated after HEQ. A I D signal example has been
given in Fig2 to show the lost information after applying the
HEQ procedure in the spatial domain. For example, compare
signal A (the original I D signal) and signal B (the signal A after
HEQ in the spatial domain) at location IO. In the wavelet
domain, the dztail-information and approximation information
are kept separately as shown in signal C and D, respectively.
Note that applying HEQ on the approximation-coefficients D
only, to get E, does not disturb any details in C. Therefore. the
reconstructed signal F, from E and C, preserves all details in
signal A. Hence, performing HEQ in the wavelet domain can
prevent detail information from being lost.
The purpose of contrast enhancement is to redistribute pixel
values uniformly, as much as possible. Fig.3 - Fig.8 show the
Lena image and i t s histogram hefore and after performing HEQ
in the spatial and the wavelet domains. We can obseNe that:
The original image in Fig.3 does not utilize the entire
dynamic range of pixel values and the pixel values are not
uniformly distributed.
The image in Fig.4 utilizes the entire dynamic range.
However there are some gray-levels have been eliminated.
The image in Fig.5 utilizes the entire dynamic range of
pixel values. Moreover, pixel values are uniformly utilized
over the entire dynamic range. From this observation (and
the other runs that are performed on other images, but are
not included in this paper), we can conclude that the
proposed wavelet-based HEQ method outperforms the
spatial domain HEQ method.

4.4. H B F Function
The purpose o f HBF i s l o enhance high frequency information
of a given image. The proposed method treats each detail
coefficient hand as a two-dimensional image (called imagehand) and applies HBF on these image-hands in the wavelet
domain (by subtracting the low-pass filtered image-hand from
the scaled original image-hand, where the scaling factorA is not
less than I):
Highboost= ( A ) Original Lowpass
The factor A determines the degree of high frequency emphasis.
During the course of this research, a 7 x 7 Gaussian filter with a
standard deviation equal 1.0 is used as the low-pass filter in the
~

111 - 398

N levels inverse
WaYelet trmsfom

Detail-coefficients

::p,
U

Fig1 . Framework of the proposed method

........
"........
..................

...................
.......

10
8

4
0

...... ......

-D

.."..

IO

20

30

40

-20

IO

..................

20

30

40

0'

10

20

30

40

Fig.2. A: the original ID signal, B: the signal A after HEQ in the spatial domain, C: the detailed-coefficients of A in the Haar wavelet
domain, D: the approximated-coefficients of A in the Haar wavelet domain. E: the approximated-coefficients of A after HEQ in the
Haar wavelet domain, F: the original signal A,after performing HEQ in the wavelet domain (reconstructed from C and E).

Fig.3. Original Lena image

Fig.4. After HEQ in the spatial domain

Fig.5. After HEQ in the wavelet domain

3
m

300

I8oc

2 a -

200

1 2oc

loo(.

100

600
I

50

'Io0 150

200

2%

Fig.6. Histogram of the image in Fig.3.

50

100

150

200

250

Fig.7. Histogram of the image in Fig 4.

HBF function. The influence of the factor A is observed through


experiments As the A value is increased, a stronger edge
emphasized result is achieved. However, we should know that
coefficients of each component have upper and lower bounds.
This means that if the A value is increased beyond a certain

50

100

150

200

250

Fig.8. Histogram of the image in Fig.5.

value, the enhanced pixel values may jump outside the allowed
boundary range. In this case, these coefficient values will be
cast lo the nearby boundary (a saturation issue) The casting
process may introduce artifacts.

111 - 399

Fig.9. Original Lena image

Fig.10. The image aRer HEQ in the


spatial domain

4.5. Result of using the proposed method


Fig.9 shows the original Lena image. Fig.10 shows the result
after using HEQ in the spatial domain. The result of using the
proposed method (i.e., using HEQ and HBF in the wavelet
domain) is shown in Fie.1 I .

[5]
[6]

3.

CONCLUSIONS

In this paper, a new wavelet-based method for contrastledge


enhancement is proposed. Experimental results show that the
histogram of a given image after performing the histogram
equalization procedure in the wavelet domain is redistributed
more uniformly than that after performing the histogram
equalization procedure in the spatial domain.
Since contrast and edge enhancement procedures are applied
separately in the wavelet domain, they do not affect each other,
and the order of applying both of them becomes irrelevant.
Moreover, in the spatial domain, the enhancement process is
performed on the full size of an image. While in the wavelet
domain, the enhancement process is performed on some of the
decomposed components, each of which is at most a quarter the
size of the original image. Therefore, computation cost in the
wavelet domain is less than that in the spatial domain. This is
especially true when considering that currently most images are
already wavelet-compressed (the current JPEG 2000 standard is
a wavelet based scheme).

6.

171
[SI

191

ACKNOWLEDGEMENT

This work was partially support by the National Science and


Engineering Council (NSERC) of Canada. This support is
greatly appreciated.

[I]
(21
[3]
[41

7. REFERENCES
R. Gonzalez, %nage Enhancement and Restoration,
Handbook of Parrern Recognition and Image Processing,
Academic Press,New York, pp. 191-213, 1986.
R. Gonzalez and R. Woods, Digifal Image Processing,
Addison-Wesley Publishing, 1992.
J. Fu, H. Lien and S. Wong, Wavelet-based HEQ of
gastric sonogram images, Computerized Medical
Imaging and Graphics, Vo1.24; pp. 5p-68, 2000.
T. Reeves and M. Jemigan, Multiscale-based Image
enhancement. Canadian Conference on E/ectrica/ and
Computer Engineering. Vol. 2. pp.50&505, 1997.

111 - 400

Fig. I I . The image after HEQ and


HBF in the wavelet domain

Rosenfeld and A. Kak, &ita/ Picture Processing, 2d


ed., Academic Press. New York, 1982.
C. BURUS,R. Gopinath and H. Guo, Introduction to
Wovelets and Wavelet Transformations: A Primer,
Prentice-Hall. New Jersev.
.. 1998.
E. Aboufadel and S . Schlicker, Discovering Wavelets,
John Wiley & Sons, New York. 1999.
Jinhui Pin, A new wavelet-based method for
contrastledge enhancement,
M.Sc. Dissertation.
Computer Science, University of Western Ontario,
London. Ontario, Canada, 2001.
L. Chen, C. Chen and K. Parker, Adaptive feature
enhancement form mammographic images with wavelet
multi-resolution analysis, Jaurml of Electronic
Imaging, Vol. 6, No. 4, pp. 467478, October 1997.
D. Kuan. A. Sawchuk, T. Strand and P. Chavel,
Adaptive noise smoothing filter for images with signaldependent noise, IEEE Transactions on Pattern
Analysis and Machine Intelligence, Vol. PAMI-7, No. 2,
pp. 165-177, March 1985.
K. Xu, X. Zheng and X. Cheng, A novel method for
image enhancement of medical images based on
wavelet, Acta Electronica Sinica, Vol. 27, No. 9_ pp.
121-123, 1999.
E. Olsen and B. Lin, A wavelet phase filter for emission
tomography, The Proceedings of the SPIE, Vol. 2491,
pp. 829439, 1995.
A. Bruce and II. Gao, Waveshrink Shrinkage function
and thresholds, The Proceedings ofthe SPIE,Vol. 2569.
PD.
. . 27&281, 1996.
W. Gong and Y. Wang, Contrast enhancement of
infrared image via wavelet transform, Journal of
National Universiry ofDefense Technology. Vol. 22, No.
6, pp. I 17-1 19,2000.
B. Xu, C. Fu and J. Ma, Image enhancement method
based on wavelet transform, Proceedings of rhe SPIE.
Vol. 4044, pp. 150-157,2000.
B. Peng , W. Fu and C. Yang, Contrast enhancement
of radiographs using shiR invariant wavelet transform.
Wuhan University Journal of Natural Sciences, Vo1.5,
No.1, P59-P62, 2000.