Adaptive-Neighborhood Histogram
Equalization of Color Images
Vasile Buzuloiu1,
Mihai Ciuc1,
∗
Rangaraj M. Rangayyan2,
Constantin Vertan1,
1
Laboratorul de Analiza şi Prelucrarea Imaginilor,
Universitatea “Politehnica” Bucureşti,
Bucharest, Romania
2
Department of Electrical and Computer Engineering,
University of Calgary,
Calgary, Alberta, T2N 1N4, Canada
∗
Please address all correspondence to :
Dr. R.M. Rangayyan
Professor
Department of Electrical and Computer Engineering
University of Calgary
2500 University Drive, N.W.
Calgary, Alberta, T2N 1N4, Canada.
Email : [email protected]
Tel : (403) 220-6745
Fax : (403) 282-6855
Submitted to the Journal of Electronic Imaging
March 2000
1
Abstract
Histogram equalization (HE) is one of the simplest and most effective tech-
niques for enhancing gray-level images. For color images, HE becomes a more
difficult task, due to the vectorial nature of data. We propose a new method for
color image enhancement that uses two hierarchical levels of HE: global and lo-
cal. In order to preserve the hue, equalization is only applied to intensities. For
each pixel (called the “seed” when being processed) a variable-sized, variable-
shaped neighborhood is determined to contain pixels that are “similar” to the
seed. Then, the histogram of the region is stretched to a range that is computed
with respect to the statistical parameters of the region (mean and variance) and
to the global HE function (of intensities), and only the seed pixel is given a new
intensity value. We applied the proposed color HE method to various images and
observed the results to be subjectively “pleasant to the human eye”, with em-
phasized details, preserved colors, and with the histogram of intensities close to
the ideal uniform one. The results compared favorably with those of three other
methods (histogram explosion, histogram decimation, and three-dimensional his-
togram equalization) in terms of subjective visual quality.
Keywords: Histogram equalization, Color image processing, Color image en-
hancement, Adaptive-neighborhood image processing.
2
1 Introduction
Histogram equalization (HE) is one of the simplest, commonly used, and effective techniques
for enhancing gray-level images. The hypotheses underlying HE are that the pixels are
independent random variables with identical probability density functions (PDF), and that
the image is a realization of an ergodic random field. Hence, the nearer an image histogram
is to a uniform distribution, the more informative the image is. It is known from probability
theory that for any random variable ξ characterized by the cumulative distribution function
(CDF) Fξ (x), the random variable η = Fξ (ξ) is uniformly distributed over (0, 1) [1]. This
result is used to achieve HE of gray-level images. The major drawback of this simple approach
is the effect of gray-level quantization: in the continuous case, extending the PDF range raises
no problems; in the discrete case, this technique is not able to increase the number of gray
levels (e.g., 32 values dispersed over a range of 256 levels will remain as 32 discrete levels,
albeit more distant from one another). Thus, the histogram of the output image is only
approximately uniform. The similarity between the output histogram and an ideal uniform
distribution is related to the number of distinct gray levels in the original image. However,
in practice, HE produces images with increased contrast, which may appear to be enhanced
and more “readable”.
When dealing with color images, HE becomes a much more difficult task, due to the
vectorial nature of the data: each pixel is represented by a vector with three components,
i.e., the amount of red (R), green (G), and blue (B) that compose the given color. It is well-
known that, in general, the three components are mutually correlated. Hence, attempting
to equalize the histogram of each channel separately would lead to an incorrect result by
neglecting the inter-component correlation. Therefore, techniques to equalize (or modify)
the histogram of a color image by joint processing of the three channels have been developed.
We should add at this point that HE is just one technique among many others for
image enhancement, mainly useful when enhancement is aimed at improving visibility and
detectability of image features; enhancement for preference would require specific effects that
might alter the characteristics of the image to suit individual preferences. Regardless, even
enhancement for preference could benefit from HE, with HE being followed by an appropriate
second step [1, 2, 3].
There are two main approaches to HE of color images: the first one considers processing of
color image data in the original RGB space. The “3-D histogram equalization”method pro-
posed by Trahanias and Venetsanopoulos [4] consists of three-dimensional (3-D) histogram
specification in the RGB cube, with the output histogram being uniform. The 3-D CDF
FRGB (r, g, b) of the original image, as well as an ideal, uniform CDF F̄R0 ,G0 ,B 0 (r0 , g 0 , b0 ) are
computed. Then, each pixel [R, G, B] is assigned the smallest value [R0 , G0 , B 0 ] for which
F̄R0 ,G0 ,B 0 (R0 , G0 , B 0 ) ≥ FRGB (R, G, B). Since these conditions are ambiguous, allowing more
than one solution, a method to determine R0 , G0 , and B 0 by sequentially incrementing (or
decrementing) R, G, and B is further proposed. Thus the output 3-D histogram gets as
close as possible to a uniform 3-D distribution.
More recently, Mlsna and Rodrı́guez [5] proposed a 3-D technique that exploits the full
3-D gamut, and called the method “histogram explosion”. For each point in the RGB cube
3
corresponding to an image color, a ray that starts from some central point (which is usually
chosen as the average color of the image) is defined to pass through that point. Then, all
points within a threshold distance of the ray are projected on to the ray. In this way, a 1-D
histogram along the ray is created. By modifying (equalizing) the 1-D histogram, the new
color value for the original point is determined. By this technique, color points are almost
uniformly spread in the color space. Another version of this method [6] deals with colors
represented in the Commission Internationale de l’Eclairage CIE-LUV space.
The “histogram decimation” technique [6, 7] attempts to uniformly scatter the color
points over the full 3-D gamut by means of an iterative algorithm. The algorithm starts by
setting the full 3-D color space as the current space. Each iteration consists of two steps:
in the first step, all color points within the current space are shifted such that their average
overlaps the geometric center of the space; in the second step the current color space is
divided into eight equally-sized subspaces. Each newly-created color subspace is set as the
current space for the next iteration, and so on. The algorithm stops when the subspace size
has reached its minimum value. Thus, the color points are spread to occupy the full gamut.
All of these methods, although interesting, are in general computationally expensive and
have the drawback of modifying color hues. In many cases, the latter may lead to results
that are unpleasant to a human observer, since it is known that the human visual system is
extremely sensitive to shifts in hue.
The second class of methods considers equalization in other perceptual color spaces,
mainly the Hue, Saturation, Intensity (HSI) space [8, 9, 10]. The advantage of representing
colors using HSI is that one can consider modifications of only intensity, or both intensity
and saturation, leaving the hue unmodified. Thus, the drawback of the previously presented
methods is overcome. Pitas and Kiniklis [9] present a method to jointly equalize intensity
and saturation. However, it is mentioned that modification of saturation is not advised, since
it results in unnatural output images. A recent method reports on histogram modification
in the color difference C–Y space [11, 12], where the authors propose modifications of satu-
ration and hue only based on a priori knowledge of the image content. HE of the intensity
(brightness) component only appears to be effective: it has a low computational cost (the
same as in the case of gray-level images) and yields results that are visually correct, since
it does not create colors that are not natural. An improvement to this simple technique is
proposed by Rodrı́guez and Yang [10], who investigate the effect of quantizing brightness to
a predefined number of bins before equalization.
Starting from the constraint that only the brightness component is to be equalized, we
propose in this paper a new method for HE of color images via the adaptive-neighborhood
approach [13, 14, 15, 16, 17, 18, 19, 20]. The method is designed to increase the number
of intensity levels in the image by taking into account values of pixels within a certain
neighborhood when computing the new intensity value of a pixel. The neighborhood is
determined adaptively for each pixel in the image by a region growing algorithm, rather than
forcing it to a predefined shape and size. Details of the proposed method are presented in
Section 2. Section 3 provides the results obtained by the proposed technique and comparison
with the results of other techniques. In Section 4, some conclusions are presented.
4
2 Adaptive-neighborhood histogram equalization
(ANHE) of color images
In this section, a new method for HE of color images using the adaptive-neighborhood
approach is presented. The adaptive-neighborhood paradigm has been used to filter signal-
dependent and signal-independent noise in gray-level images [13, 14, 15, 16], and has been
adapted to the case of noisy color images [17]. The paradigm was also used to achieve
HE of gray-level images [18] and enhancement of medical images [19]. Regardless of the
application, the idea is to determine for every pixel in the image (called the “seed” when being
processed) a variable-sized, variable-shaped neighborhood by a region growing procedure.
The neighborhood contains only those spatially-connected pixels that are “similar” to the
seed. The similarity between two pixels is generally assessed by comparing the Euclidean
distance between the values of the two pixels with a threshold value. Then, the pixels in the
adaptive neighborhood are used to derive the new value for the seed location. It should be
noted that the procedure is applied to every pixel in the image: each pixel becomes the seed
for region growing when being processed.
The proposed algorithm [20] combines together several ideas:
• The hue must be preserved; thus, equalization will be applied only to the brightness
component of the color image. One may argue that this principle does not always
correspond to real-world situations. For instance, in color reproduction applications
[21, 22], it is preferable to shift hues to desired values. Still, HE is a basic enhancement
operation that is assumed to yield correct results without prior information about the
image content. As mentioned above, HE can be the first step of a more elaborate chain
of enhancement steps that also control hue shifts. Thus, in order to avoid uncontrolled
hue deviation, it is preferable that HE does not alter hues in any way.
• A good way of extending the number of levels is to take into account the neighboring
values, i.e., to compute the new value as a function of a few local variables (as in the
case of filtering with local kernels) and not as a function of only the original value of
the pixel being processed.
• The neighborhood that contributes to the derivation of the new value of a pixel must
be chosen adaptively, for reasons of uniformity. The advantage of using an adaptive
neighborhood instead of a fixed one is that stationarity is guaranteed within such
neighborhoods. Only pixels that have values similar to the seed are chosen to contribute
to its output value. By such an approach, local details are emphasized.
• Generally, a well–balanced image is one with the average intensity at the middle of
the gamut; therefore, the global histogram of the intensity image is to be taken into
account when computing the new value of a pixel. In the case of images for which the
above remark does not stand true, point operators other than the global HE function
can be used (see Section 3).
5
The steps of the ANHE algorithm for color images are shown in the flowchart in Figure 1
and detailed in the following subsections.
2.1 Computation of the global HE function
In the first step, the intensity image I is computed from the three color channels R, G, and
B. Since I is equivalent to a gray-level image, its 1-D intensity histogram (or PDF) can be
computed as
−1 N −1
1 MX X
hI (i) = δ(I(m, n) − i), i = 0, . . . , L − 1,
M N m=0 n=0
where M and N are the number of rows and columns in the image, L is the number of levels
in the intensity image (usually, L = 256), and
(
1 if j = 0
δ(j) =
0 otherwise.
Based on the histogram, the global HE function FI is then determined:
i
X
FI (i) = (L − 1) × hI (j) , i = 0, . . . , L − 1.
j=0
For each value I(m, n), FI (I(m, n)) stores the new value as given by the global HE technique.
2.2 Histogram equalization using the adaptive neighborhood
As stated before, for each pixel in the intensity image I, an adaptive neighborhood (region)
is to be determined to contribute to the computation of its new value. Pixels belonging to
the region have to fulfill two conditions:
• to have the intensity value close to that of the seed;
• to be spatially connected to the seed.
2.2.1 Adaptive region growing
The region-growing procedure [13, 14] consists of evaluating the absolute difference dkl be-
tween each of the eight-connected neighbors I(k, l) of the seed and the seed I(i, j) as
dkl = |I(k, l) − I(i, j)| .
Pixels I(k, l) having dkl ≤ T , where T is a fixed, predefined threshold are included in the
region. The algorithm then proceeds by checking all eight-connected neighbors of the newly-
included pixels in the same manner and stops when either the inclusion criterion is no longer
fulfilled for any neighboring pixel, or the number of included pixels equals a predefined
6
upper-limit Nmax . The region size may be limited in order to keep the computational cost
of the algorithm within practical limits.
Although the choice of T is not critical, it is important to have its value between some
given limits depending upon the image: a large value of T would lead to regions (neighbor-
hoods) not representative of the object to which a seed belongs (uniformity is not preserved),
whereas too small values of T would result in very small regions that are inadequate to sup-
port statistical processing. Tests conducted on many images support the observation that
an acceptable range for T is T ∈ [20, 40]. If, however, for the chosen T , the region has not
grown adequately (i.e., it contains less than a predefined lower-limit of Nmin pixels), T is
increased and the region-growing procedure is repeated.
2.2.2 Modification of the region intensity histogram
After the region has been grown, its intensity histogram hreg is computed. Due to the
thresholding step in the region-growing procedure, the intensity histogram of the region will
be narrow, i.e., it will have non-zero components in a limited interval. In order to compute
the new intensity value for the seed, the intensity histogram has to be modified. It is obvious
that attempting to equalize the intensity histogram of the region over the whole gamut would
lead to incorrect results. This is because a few intensity levels within a limited range would
get scattered from black to white. Since the purpose of this step is to reinforce details,
we chose not to equalize, but to stretch the intensity histogram of the region to a range
[imin , imax ]. The values of imin and imax must be computed with respect to the data in the
intensity histogram of the region, and, as mentioned before, with further reference to the
global HE function FI .
In order to determine the limits of the range to stretch the intensity histogram of the
region, the mean µreg and standard deviation σreg of the population of pixels within the
region are computed. Then, we take:
imin = FI (µreg ) − κσreg ,
and
imax = FI (µreg ) + κσreg ,
where κ is a constant (for the entire image) to be chosen by the user. Thus, the new intensity
value of the seed pixel is computed as
I 0 (i, j) = [imin + (imax − imin ) × Freg (I(i, j))] ,
where Freg is the CDF of the region computed from the intensity histogram hreg . We stress
that only the seed intensity is updated according to the modified histogram of the region.
The choice of the values of imin and imax may be described as follows: the whole histogram
of the region is to be shifted such that the mean µreg overlaps its value given by the global
HE function FI (µreg ) and is further extended to a range that is proportional to the standard
deviation of the pixels within the region. The need for histogram shifting can be related
to the fact that we want the resulting image to be well-balanced, i.e., to have luminosities
7
globally spread from black to white. This can only be achieved by the use of the global
HE function. If, in the limit case, we impose the maximum number of pixels in the region
Nmax = 1, the outcome of our method would be the image with the equalized histogram of
intensities.
The range width is proposed to be proportional to the standard deviation σreg rather than
being fixed for the following reason: we should not amplify very small variations that cannot
be interpreted as significant details, and are most probably due to imperfections of the image
acquisition system. For instance, a blue sky area in an image has to remain uniform if there
are no clouds. On the contrary, if the nonuniformity of the region is important, i.e., there
are significant details in the region, we want to emphasize them by extending the range over
which the histogram of the region is stretched. In Figure 2, the initial and modified intensity
histograms of a region of an image are presented.
The parameter κ, which has to be chosen by the user, controls the width of the range
over which the region histogram is expanded, i.e., it controls the “amplification” of details.
For normal-contrast images, we suggest to choose κ ∈ [3, 7]; a larger value for κ is indicated
for low-contrast images. κ has to be set with respect to the threshold T used in the region-
growing procedure. Indeed, if regions are grown with a small T , then the standard deviation
of the pixel population within a region σreg would be small, and thus one should set κ to a
larger value in order to obtain a visible emphasis of details.
The ANHE procedure for color images consists of a two-level HE procedure: global and
local. After the new intensity value I 0 has been computed, the new values of the color
components R0 , G0 , and B 0 of a pixel are determined by scaling the old values R, G, and B
with the factor I 0 /I. Thus, the color hues are preserved, and only brightness is modified.
The overall procedural flowchart of ANHE for color images is shown in Figure 1.
3 Results and discussion
In this section, the results of ANHE of color images are provided and compared with those
of other HE techniques for color images.
The user of the results of HE is, in most cases, the human visual system. Thus, it is
difficult to derive objective criteria to assess the effectiveness of such an operation. Some
authors have proposed, as a measure of quality, the difference between the histogram of the
output image and an ideal 3-D uniform histogram[6, 7]. Due to the fact that our method
introduces more levels of brightness, we get histograms much closer to the ideally uniform one
than those provided by other methods. Nevertheless, our goal is an optimal expansion of the
existing values in the whole 3-D color space with constraints, rather than free, unconstrained
equalization, due to the artifacts of the latter. A comparison between the outcomes of
different HE techniques for color images can be done only subjectively, by a human observer.
Figures 3, 4, 5, and 6 present the original versions of four test images (obtained from the
ftp site ipl.rpi.edu) and their processed versions as given by ANHE and four other techniques.
Figures 7 and 8 present the original and the ANHE versions of six more images. The outcome
of the brightness-only global HE method is well contrasted and provides natural colors, but
the details are not emphasized. Equalization of each channel independently creates false
8
colors since it is performed separately within three different contexts, i.e., the HE function
on each channel is computed with no attention paid to the two other channels.
After “3-D histogram equalization”, the images tend to become too bright, and colors
faint, due to color modification along the main diagonal of the RGB cube. The “histogram
decimation” technique provided the poorest results in terms of hue modification, leading to
the appearance of many unnatural colors in the image. This is because of the scattering
of color points in the RGB cube with the aim of obtaining a uniform distribution with
no respect to the original hues. One other drawback of this technique is that colors that
are originally close may be treated in different manners, due to the fact that they could
be assigned to different subspaces; thus, they may severely differ in the resulted equalized
image.
“Histogram explosion” provided acceptable results. The main drawback of the method
is the uncontrolled color hue modification. In spite of the fact that the shifts in hue created
by “exploding” the histogram could result in colors that may be pleasant to the human eye,
some of the newly-created colors gave pictures an unnatural quality.
The images equalized with the ANHE method were very good in terms of visual appear-
ance. The ANHE features of color preservation, uniform spreading of luminosities from black
to white, and detail emphasis (see Figure 9) acted together to provide good enhancement.
Figure 10 shows the intensity histograms of all of the images presented in Figure 3. The
histogram after ANHE is the closest to a uniform one.
After analyzing the results of tests conducted with many images, we can state that the
images yielded by ANHE of color images appear to be the most pleasant to the human eye:
the images retain their natural colors and the details are very well contrasted. Unlike the
other techniques, the number of intensity levels is increased by ANHE, and consequently,
smooth variations are maintained within objects with no false contours. Moreover, local
details are slightly emphasized, which gives an overall pleasant aspect to the images. Table
1 presents the number of distinct colors in four of the test images after applying various HE
techniques. As expected, the ANHE result has the maximum number of distinct colors.
Since equalization is applied to the brightness component only, the ANHE technique
can be applied to gray-level images as well, without loss of effectiveness. In the case of
gray-level images, there are applications when other point operators (such as piece-wise
linear, logarithmic, exponential, or nonlinear operators [1]) might perform better than HE.
Another alternative to HE is histogram specification, where the aim is to obtain not a uniform
histogram, but one of a pre-defined shape. It is possible to take into consideration operators
as above when applying ANHE to color images by chosing the function FI to be not the
global HE function but some other function that maps one gray-level to another.
In terms of computational requirements, the most efficient techniques are global equaliza-
tion of brightness only and independent equalization of each channel, which is expected, due
to their simplicity. The histogram decimation procedure is also fast, whereas 3–D histogram
equalization and histogram explosion are computationally the most expensive procedures;
moreover, these three techniques have the drawback that their processing times depend upon
the number of distinct colors in the input image. The time requirement of ANHE strongly
depends upon the parameter Nmax , i.e., the maximum number of pixels allowed in a region.
9
Tests on various images have indicated that there is no noticeable improvement in the ANHE
result if Nmax is chosen over a limit of about 100 pixels (for a 256×384-pixel image). Table 2
summarizes the execution times for all of the algorithms run on a SUN ULTRA SPARC 10
workstation powered by a 440 MHz processor, with 1GB RAM. The computing time could
be reduced by optimization of the program code and use of parallel computing systems.
4 Conclusion
We have proposed a new method for histogram equalization (HE) of color images by using
an adaptive-neighborhood approach. The main idea of the algorithm is to use two levels
of equalization: global and local. The new intensity of a pixel is computed with respect
to the global HE function, and with respect to the values of pixels in a neighborhood that
is adaptively determined for each pixel individually. The results yielded by the proposed
technique are better in terms of visual quality and the number of distinct colors than those
provided by other HE techniques for color images. Since the computational requirement is
not very demanding, the proposed technique is a powerful tool for color image enhancement.
Acknowledgements
We thank the Romanian Ministry of Research and Technology, the University Research
Grants Committee (URGC) of the University of Calgary, and the Natural Sciences and
Engineering Research Council (NSERC) of Canada for supporting this project.
10
References
[1] A.K. Jain. Fundamentals of Digital Image Processing. Prentice Hall, Englewood Cliffs,
NJ, 1989.
[2] W.K.Pratt. Digital Image Processing. John Wiley & Sons, New York, 2nd edition, 1991.
[3] R.C.Gonzalez and R.E.Woods. Digital Image Processing. Addison Wesley, Reading MA,
1992.
[4] P.E.Trahanias and A.N.Venetsanopoulos. Color image enhancement through 3-D his-
togram equalization. Proc. of the 11th IAPR Intl. Conf. on Pattern Recognition, The
Hague, Netherlands, pages 545–548, 1992.
[5] P.A.Mlsna and J.J.Rodrı́guez. A multivariate contrast enhancement technique for mul-
tispectral images. IEEE Trans. on Geoscience and Remote Sensing, 33:212–216, 1995.
[6] P.A. Mlsna, Q. Zhang, and J.J. Rodrı́guez. 3-D histogram modification of color images.
Proc. IEEE Intl. Conf. on Image Processing, Lausanne, Switzerland, III:1015–1018,
1996.
[7] Q. Zhang, P.A. Mlsna, and J.J. Rodrı́guez. A recursive technique for 3-D histogram
enhancement of color images. Proc. IEEE Southwest Symp. on Image Analysis and
Interpretation, San Antonio, TX, pages 218–223, 1996.
[8] I.M. Bockstein. Color equalization method and its application to color image processing.
Journal of the Optical Society of America, 3:735–737, 1986.
[9] I. Pitas and P. Kiniklis. Multichannel techniques in color image enhancement and
modelling. IEEE Trans. on Image Processing, 5:168–171, 1996.
[10] J.J. Rodrı́guez and C.C. Yang. High-resolution histogram-modification of color images.
Graphical Models and Image Processing, 57:432–440, 1995.
[11] A.R. Weeks, L.J. Sartor, and H.R. Myler. Histogram specification of 24-bit color images
in the color difference (C–Y) color space. In E.R. Dougherty and J.T. Astola, editors,
Proc. SPIE on Nonlinear Image Processing X, volume 3646, pages 319–329. SPIE, March
1999.
[12] A.R. Weeks, L.J. Sartor, and H.R. Myler. Histogram specification of 24-bit color images
in the color difference (C–Y) color space. Journal of Electronic Imaging, 8(3):290–310,
1999.
[13] R.B. Paranjape, R.M.Rangayyan, and W.M.Morrow. Adaptive neighborhood mean and
median filtering. Journal of Electronic Imaging, 3:360–367, October 1994.
[14] R.B.Paranjape, T.F.Rabie, and R.M.Rangayyan. Image restoration by adaptive neigh-
borhood noise subtraction. Applied Optics, 33:1861–1869, May 1994.
11
[15] R.M.Rangayyan and A. Das. Filtering multiplicative noise in images using adaptive
region-based statistics. Journal of Electronic Imaging, 7:222–230, 1998.
[16] R.M. Rangayyan, M. Ciuc, and F. Faghih. Adaptive-neighborhood filtering of images
corrupted by signal-dependent noise. Applied Optics, 37(5):4477–4487, May 1998.
[17] M. Ciuc, R.M. Rangayyan, T. Zaharia, and V. Buzuloiu. Adaptive neighborhood filters
for color image filtering. In E.R. Dougherty and J.T. Astola, editors, Proc. SPIE on
Nonlinear Image Processing IX, volume 3304, pages 277–286. SPIE, May 1998.
[18] R.B. Paranjape, W.M. Morrow, and R.M. Rangayyan. Adaptive neighborhood his-
togram equalization for image enhancement. CVGIP:Graphical Models and Image Pro-
cessing, 54:259–267, 1992.
[19] W.M. Morrow, R.B. Paranjape, R.M. Rangayyan, and J.E.L.Desautels. Region-based
contrast enhancement of mammograms. IEEE Trans. on Medical Imaging, MI-11:392–
406, 1992.
[20] V. Buzuloiu, M. Ciuc, R.M. Rangayyan, L. Kij, and C. Vertan. Histogram equalization
of color images using the adaptive neighborhood approach. In E.R. Dougherty and
J.T. Astola, editors, Proc. SPIE on Nonlinear Image Processing X, volume 3646, pages
330–338. SPIE, March 1999.
[21] G. Beretta. Dynamic exposure control in color scanners. IS&T/SPIE Symposium on
Electronic Imaging Science & Technology, Technical Conference 3019, San José, CA,
pages 76–88, February 1997.
[22] Special issue: Color imaging for digital publishing. In G. Beretta, editor, SPIE/IS&T
International Technical Group Newsletter Electronic Imaging, volume 9. January 1999.
12
Image Original Independent Brightness 3-D histogram Histogram Histogram ANHE
equalization equalization equalization decimation explosion
“03” 33936 29226 28868 28684 32281 31976 41784
“08” 63126 55711 53672 51008 61086 60306 64536
“13” 59974 51416 51198 45618 57074 57792 66175
“22” 63145 59072 50103 56040 62126 62246 65834
Table 1: Number of distinct colors in images before and after histogram equalization.
Image Independent Brightness 3-D histogram Histogram Histogram ANHE
equalization equalization equalization decimation explosion (Nmax = 100)
“03” 0.08 0.2 241.9 21.1 1167.6 31.9
“08” 0.08 0.2 873.6 123 4607.8 30.6
Table 2: Execution times (in seconds) for the various histogram equalization methods stud-
ied.
13
START
Compute the image of intensities I.
Compute the global HE function F I .
Set next pixel in the I image as seed.
Add seed to region queue (RQ).
Set next pixel in RQ as current.
Inspect next neighbor of the current pixel.
Compute its absolute difference with respect to the seed d.
Yes No
d<T?
Add neighbor to RQ
All
Yes neighbors No
inspected ?
No No. of region Yes
pixels > N max ?
No All pixels in Yes
RQ inspected?
Yes No. of region No
pixels < N min ?
Reset RQ. Add seed to RQ. Compute histogram h reg,
Increase T. mean µ reg and standard
deviation σ reg.
Stretch h reg over
[ F I ( µ reg ) - κ σ reg , F I ( µ reg ) + κ σ reg ].
Update seed intensity.
Compute seed’s new color components.
Yes All pixels in No
STOP the I image
inspected ?
Figure 1: Flowchart of the ANHE method.
14
(a) (b)
(c) (d)
Figure 2: Steps for histogram modification of a region: (a) Region grown for seed pixel
(182,139) in the “03” image in Figure 3.a having intensity value 60; the region pixels are in
white, seed pixel in black. The region size is limited to Nmax = 500 pixels (b) Global HE
function. FI (63) = 47. (c) Original histogram of the region; µreg = 63, σreg = 6.2, imin = 18,
imax = 78. (d) Histogram of the region after modification.
15
(a) (b)
(c) (d)
(e) (f)
Figure 3: Results of histogram equalization: (a) The original 256×384 pixel “03” image.
(b) The image after histogram equalization on each channel independently. (c) The image
after 3-D histogram equalization. (d) The image after histogram decimation. (e) The image
after histogram explosion. (f) The image after ANHE with Nmax = 100, Nmin = 20, T = 20,
and κ = 3.
16
(a) (b)
(c) (d)
(e) (f)
Figure 4: Results of histogram equalization: (a) The original 256×384 pixel “08” image.
(b) The image after histogram equalization on each channel independently. (c) The image
after 3-D histogram equalization. (d) The image after histogram decimation. (e) The image
after histogram explosion. (f) The image after ANHE with Nmax = 100, Nmin = 20, T = 20,
and κ = 3.
17
(a) (b)
(c) (d)
(e) (f)
Figure 5: Results of histogram equalization: (a) The original 256×384 pixel “13” image.
(b) The image after histogram equalization on each channel independently. (c) The image
after 3-D histogram equalization. (d) The image after histogram decimation. (e) The image
after histogram explosion. (f) The image after ANHE with Nmax = 100, Nmin = 20, T = 20,
and κ = 3.
18
(a) (b)
(c) (d)
(e) (f)
Figure 6: Results of histogram equalization: (a) The original 256×384 pixel “22” image.
(b) The image after histogram equalization on each channel independently. (c) The image
after 3-D histogram equalization. (d) The image after histogram decimation. (e) The image
after histogram explosion. (f) The image after ANHE with Nmax = 100, Nmin = 20, T = 20,
and κ = 3.
19
(a) (b)
(c) (d)
(e) (f)
Figure 7: Results of ANHE: (a) Original “01” image (b) “01” image after ANHE (c) Original
“05” image (d) “05” image after ANHE (e) Original “06” image (f) “06” image after ANHE
20
(a) (b)
(c) (d)
(e) (f)
Figure 8: Results of ANHE: (a) Original “07” image (b) “07” image after ANHE (c) Original
“14” image (d) “14” image after ANHE (e) Original “23” image (f) “23” image after ANHE
21
(a) (b) (c)
Figure 9: Detail emphasis by ANHE: (a) 88×138-pixel portion of the original “13” image.
(b) Same portion of the image after histogram equalization of brightness only. (c) Image
after ANHE.
22
(a) (b)
(c) (d)
(e) (f)
Figure 10: Histograms of intensities of the images in Figure 3: (a) Original. (b) After
histogram equalization on each channel independently. (c) After 3-D histogram equalization.
(d) After histogram decimation. (e) After histogram explosion. (f) After ANHE.
23
