Image enhancement using a contrast measure in the compressed domain

Abstract

An image enhancement algorithm for images compressed using the JPEG standard is presented. The algorithm is based on a contrast measure defined within the discrete cosine transform (DCT) domain. The advantages of the psychophysically motivated algorithm are 1) the algorithm does not affect the compressibility of the original image because it enhances the images in the decompression stage and 2) the approach is characterized by low computational complexity. The proposed algorithm is applicable to any DCT-based image compression standard, such as JPEG, MPEG 2, and H. 261.

Full text

Page 1

IEEE SIGNAL PROCESSING LETTERS, VOL. 10, NO. 10, OCTOBER 2003 289
Image Enhancement Using a Contrast Measure
in the Compressed Domain
Jinshan Tang, Senior Member, IEEE, Eli Peli, and Scott Acton, Senior Member, IEEE
Abstract—An image enhancement algorithm for images com-
pressed using the JPEG standard is presented. The algorithm is
based on a contrast measure defined within the discrete cosine
transform (DCT) domain. The advantages of the psychophysically
motivated algorithm are 1) the algorithm does not affect the com-
pressibility of the original image because it enhances the images
in the decompression stage and 2) the approach is characterized
by low computational complexity. The proposed algorithm is
applicable to any DCT-based image compression standard, such
as JPEG, MPEG 2, and H. 261.
Index Terms—Compressed domain, contrast measure, discrete
cosine transform (DCT), human vision system, image enhance-
ment, JPEG/MPEG.
I. INTRODUCTION
T
than the original image for a specific application or set of
objectives [1]. Many image enhancement algorithms have
been proposed. One of the most widely used algorithms is
global histogram equalization [1], which adjusts the intensity
histogram to approximate a uniform distribution. The main
disadvantage of global histogram equalization is that the global
image properties may not be appropriately applied in a local
context [2]. In fact, global histogram modification treats all
regions of the image equally and, thus, often yields poor local
performance in terms of detail preservation. Therefore, several
local image enhancement algorithms have been introduced to
improve enhancement [2]–[8]. Each of these algorithms can
be classified into two types of image enhancement methods
[4]: indirect image enhancement methods and direct image
enhancement methods. The algorithms described in [2] and
[3] belong to the class of indirect image contrast enhancement
methods, since they enhance the image without measuring the
contrast. The algorithms described in [4]–[8] are called direct
local contrast enhancement methods because they establish
a criterion of contrast measure and enhance the images by
improving the contrast measurement directly.
HE GOAL of image enhancement is to improve the
image quality so that the processed image is better
ManuscriptreceivedMarch21,2002;revisedNovember4,2002.Theworkof
S. Acton and J. Tang was supported in part by the National Science Foundation
(NSF) under Grant 0121596, and the work of E. Peli was supported in part by
the National Institutes of Health under Grant EY05957 and Grant EY12890.
The associate editor coordinating the review of this manuscript and approving
it for publication was Dr. Ramanujan Kashi.
J. Tang and S. Acton are with the Department of Electrical and Computer
Engineering, University of Virginia, Charlottesville, VA 22904 USA (e-mail:
jt6w@virginia.edu).
E. Peli is with the Schepens Eye Research Institute, Harvard Medical School,
Boston, MA 02114 USA.
Digital Object Identifier 10.1109/LSP.2003.817178
A key step in the direct image enhancement approach is
the establishment of a suitable image contrast measure. For
simple patterns, two definitions of contrast measure have been
frequently used. One is the Michelson contrast measure [9]; the
other is the Weber contrast measure (see [10]). The Michelson
contrast measure is used to measure the contrast of a periodic
pattern such as a sinusoidal grating, while the Weber contrast
measure assumes a large uniform luminance background with
a small test target. Both measures are therefore unsuitable
for measuring the contrast in complex images. A number of
contrast measures were proposed for complex images [5]–[7],
[10], [11]. A local contrast measure is proposed in [6], where
the contrast is measured using the mean gray values in two
rectangular windows centered on a given pixel. Another con-
trast measure based on a local analysis of edges is defined in
[7] and is derived from the definition in [6].
Becausehumancontrastsensitivityisafunctionofspatialfre-
quency, an image’s spatial frequency content should be consid-
eredinthedefinitionofcontrast.Thecontrastmeasureproposed
in [10] explicitly satisfies this requirement. That definition of
localbandlimitedcontrastassignsacontrastvaluetoeverypoint
intheimageforeachspatialfrequencyband.Foreachfrequency
band, contrast is defined as the ratio of the bandpass-filtered
image at that frequency to the image lowpass-filtered to an oc-
tave below the same frequency. This multiscale contrast struc-
ture has found wide applications especially in image processing
problems related to the human vision system [8], [10].
Thisletterprovidesanewcontrastmeasurethatcanbeusedto
measurethecontrastofimagesintheDCTdomain.Thecontrast
measure is defined as the ratio of high-frequency content and
low-frequency content in the bands of the DCT matrix. Like the
contrast measure defined in [10], our contrast measure also has
a multiscale structure that corresponds with the human vision
system. Based on this contrast measure, an image enhancement
algorithmfordirectapplicationtothecompresseddomainisde-
veloped. The basic idea of our algorithm is to filter the image
by manipulating the DCT coefficients according to the contrast
measure defined. The proposed algorithm has the following ad-
vantages: 1) the algorithm does not affect the compressibility
of the original image; 2) given a majority of zero-valued DCT
coefficients (after quantization), the algorithm expense is rela-
tivelylow;and3)theproposedimageenhancementalgorithmis
applicabletoanyDCT-basedimagecompressionstandard,such
as JPEG, MPEG 2, and H. 261.
II. IMAGE CONTRAST ENHANCEMENT IN JPEG DOMAIN
A. Preliminaries
A JPEG system is composed of an encoder and a decoder.
In the encoder, the image is first divided into nonoverlapping
1070-9908/03$17.00 © 2003 IEEE

Page 2

290IEEE SIGNAL PROCESSING LETTERS, VOL. 10, NO. 10, OCTOBER 2003
8
each 8
are quantized using a specified quantization table. Quantiza-
tion of the DCT coefficients is a lossy process, and in this step,
many small coefficients (usually high frequency) are quantized
to zeros. The zig-zag scan of the DCT matrix followed by en-
tropy coding makes use of this property to lower the bit rate re-
quiredtoencodethecoefficients.Inthedecoder,thecompressed
imageisdecodedandthendequantizedbypointwisemultiplica-
tion with the quantization table and inverse-DCT-transformed.
Let
be an 88 block in the original image, and the
DCT transform of it is
. The 2-DCT transformation is ex-
pressed as
8 blocks. Then, the two-dimensional DCT is computed for
8 block. Once the DCT coefficients are obtained, they
(1)
where ,
, and
if
otherwise.
(2)
The DCT inverse transformation can be expressed as
(3)
where ,
From (3), we see that each
corresponding to the
in the output DCT block are arranged left to right, and top
to bottom in order of increasing spatial frequencies in the hori-
zontal and vertical spatial dimensions, respectively.
The spatial frequency properties of the DCT coefficients pro-
vide a natural way to define a contrast measure in the DCT do-
main.Itisknownthatthehumanvisualdetectiondependsonthe
ratio between high-frequency and low-frequency content [10].
Thus, the contrast measure can be defined as the ratio of high-
and low-frequency content in the bands of the DCT matrix.
We first classify the coefficients into 15 different frequency
bands. The th band is composed of the coefficients with
. A band defined by
approximation to a circle and, thus, selects approximately equal
radial frequencies. Therefore, the image block generated using
(3) by retaining only one band can be thought of as the band-
pass version of the original image block. As the band number
increases, the frequency content of the bandpass image block
corresponds with higher frequencies and, thus, creates a primi-
tive multiscale structure. Our local contrast measure is defined
on each band with band number more than0. The contrast at the
th band () is defined as
.
represents the contribution
th waveform [12] and the coefficients
gives a diamond-shaped
(4)
where
(5)
Fig. 1.DCT output block.
is the average amplitude over a spectral band. Fig. 1 illustrates
the first and fourth bands and
(6)
Here, for the sake of simplicity, we assume that visual acuity is
isotropic.
Thedefinitionprovidesalocalcontrastmeasureforeachband
(
) in the DCT domain. The contrast measure
band is the ratio of the frequency content of the bandpass image
block obtained by the th band and the frequency content of the
lowpassimageblockthatcanbegeneratedusing(3)byretaining
allofthebandsinwhichthebandnumbersarelessthan .Thus,
the definition of our contrast measure has a multiscale structure
similar to [5] and [10] in a primitive sense.
in the th
B. Image Enhancement in JPEG Domain
There are three ways to enhance the JPEG compressed im-
ages. The first is to enhance the image before compression.
However, there are two disadvantages of this approach. One is
that enhancement will reduce the compressibility of the original
image;theotheristhatitwillaffectallthereceivers.Thesecond
way is to enhance the image after decompression. Because the
postcompressionapproachdoesnotaffectthecompressibilityof
the original image, it is often adopted. In this letter, we consider
direct enhancement in the compressed domain. The basic idea
ofthismethodistoenhancetheimagebymanipulatingtheDCT
coefficients. Compared with the image enhancement in the spa-
tial domain, this method can reduce storage requirements and
computational expense as the majority of the coefficients in the
DCT domain are zeros after quantization.
The proposed image enhancement algorithm is based on the
contrast measure proposed in Section II-A. Let the contrast
of the original block be
the contrast at a specific frequency band corresponding to
and let the contrast of the enhanced block be denoted by
. If, for example, one wishes to enhance
the contrast uniformly for all frequencies, then
, whereis
(7)

Page 3

TANG et al.: IMAGE ENHANCEMENT USING A CONTRAST MEASURE IN THE COMPRESSED DOMAIN291
(a)(b)
(c)(d)
Fig. 2.
(d) The proposed contrast-measure-based method with ? ? ????.
Enhanced images using different algorithms. (a) Decompressed JPEG image. (b) Global histogram equalization. (c) Local histogram equalization [2].
leading to
(8)
Equation (8) can be stated as
(9)
where
(10)
From (9), we can obtain the enhanced DCT coefficients
as
(11)
can be obtained by recursion. The pro-
posed algorithm can be summarized as follows.

Page 4

292IEEE SIGNAL PROCESSING LETTERS, VOL. 10, NO. 10, OCTOBER 2003
(e)
Fig.2.
image using alpha-rooting algorithm with ? ? ????.
(Continued)Enhancedimagesusingdifferentalgorithms.(e)Enhanced
Step 1. Let ? ? ?,???? ? ??? and
??? ? ?? ? ??????
(12)
Step 2. Let ?
??.
Step 3. Use (11) to obtain????? (? ? ? ? ?).
Step 4. If ??
??, use (5) and (9) to compute
?? and ???, respectively. Else, end.
Step 5. Return to Step 2.
?
? ? ? and use (10) to compute
Hereis an image enhancement control factor that is chosen
by the user. When
, the image will be enhanced. When
, the image will be softened.
III. EXPERIMENTAL RESULTS AND DISCUSSION
In the experiments provided here, a JPEG compressed image
was used to evaluate the performance of the proposed algo-
rithm. The decompressed image without enhancement is shown
in Fig. 2(a). The input image has a gray resolution of eight bits.
The size of the image is 256
256.
The enhanced images obtained by global histogram equal-
ization, local histogram equalization [2] (with a window size of
30
30), and the proposed method are shown in Fig. 2(b)–(d),
respectively. In Fig. 2(d), the value of enhancement factor
was decided by a subjective test. When compared with
the original image, the histogram equalization methods and the
proposed method produced moderatelyenhancedimages. How-
ever, in our judgment, the proposed method obtained an en-
hanced image with improved visual quality compared to both
of the histogram equalization methods. With histogram equal-
izations,someregionsappearoverlydarkenedandothersoverly
lightened [see Fig. 2(b) and (c)]. The visual quality of the en-
hanced image obtained by global histogram equalization is su-
perior to that obtained by local histogram equalization as the
local enhancement artificially overemphasizes local details.
Inadditiontothecomparisonbetweenhistogramequalization
methodsandtheproposedalgorithm,wealsocomparedthepro-
posedalgorithmwithanotherDCT-basedenhancementmethod:
the alpha-rooting algorithm [13], [14]. In this algorithm, the
magnitude of each DCT coefficient is raised to a power
is a positive real number). Let
modified DCT coefficient
is expressed as [14]
(
be the DCT coefficients; the
(13)
Fig.2(e)displaystheenhancedimagebythealpha-rootingal-
gorithm. The value of
used in our experiments is 0.98, which
wasobtainedbyasubjectivetest.Comparingtheresultantimage
from the proposed method with the alpha-rooting algorithm,
one can see that the image obtained with the contrast-measure-
based method retains more detail than the image obtained with
the alpha-rooting algorithm. The image obtained by the alpha-
rooting algorithm is darker than the original image when ob-
served on the screen; however, the printed version differs (the
difference between the screen view and the printed version is
due to the nonlinear gamma response of the monitor).
IV. CONCLUSION
In this letter, we have described an image contrast enhance-
mentalgorithmthatisbasedonacontrastmeasuredefinedinthe
DCT domain. The comparative analysis between the proposed
algorithm and two existing algorithms has shown the merit of
the contrast measure-based approach.
REFERENCES
[1] R. Gonzalez and P. Wintz, Digital Image Processing.
Addison–Wesley, 1987.
[2] R.H.SherrierandG.A.Johnson,“Regionallyadaptivehistogramequal-
izationofthechest,”IEEETrans.Med.Imaging,vol.MI-6,pp.1–7,Mar.
1987.
[3] A. Polesel, G. Ramponi, and V. J. Mathews, “Image enhancement via
adaptive unsharp masking,” IEEE Trans. Image Processing, vol. 9, pp.
505–510, Mar. 2000.
[4] H. D. Cheng and H. Xu, “A novel fuzzy logic approach to contrast en-
hancement,” Pattern Recognit., vol. 33, pp. 809–819, 2000.
[5] A. Toet, “Multiscale contrast enhancement with applications to image
fusion,” Opt. Eng., vol. 31, no. 5, 1992.
[6] W. M. Morrow, R. B. Paranjape, R. M. Rangayyan, and J. E. L. De-
sautels, “Region-based contrast enhancement of mammograms,” IEEE
Trans. Med. Imaging, vol. 11, pp. 392–406, Sept. 1992.
[7] A.BeghdadiandA.L. Negrate,“Contrastenhancementtechnique based
on local detection of edges,” Comput. Vis. Graph. Image Process., vol.
46, pp. 162–274, 1989.
[8] T. Zong, H. Lin, and T. Kao, “Adpative local contrast enhancement
method for medical images displayed on a video monitor,” Med. Eng.
Phys., vol. 22, pp. 79–87, 2000.
[9] A. A. Michelson, Studies in Optics.
1927.
[10] E. Peli, “Contrast in complex images,” J. Opt. Soc. Amer. A, vol. 7, pp.
2032–2040, 1990.
[11] S. S. Agaian, K. Panetta, and A. M. Grigoryan, “Transform-based
image enhancement algorithms with performance measure,” IEEE
Trans. Image Processing, vol. 10, pp. 367–382, Mar. 2001.
[12] V. Bhaskaran and K. Konstantinides, Image and Video Compression
Standards: Algorithms and Architecture.
Kluwer, 1995.
[13] H. C. Andrews, A. G. Tescher, and R. P. Kruger, “Image processing by
digital computer,” IEEE Spectrum, vol. 9, pp. 20–32, July 1972.
[14] S. Aghagolzadeh and O. K. Ersoy, “Transform image enhancement,”
Opt. Eng., vol. 31, pp. 614–626, Mar. 1992.
Reading, MA:
Chicago, IL: Univ. Chicago Press,
Dordrecht, The Netherlands: