Image Contrast Enhancement based on Histogram by jasonndcosta


A MODIFIED UNSHARP-MASKING TECHNIQUE FOR IMAGE CONTRAST, A NEW APPROACH FOR VERY DARK VIDEO DENOISING AND ENHANCEMENT, A PDE Approach to Super-resolution with, A Unified Histogram and Laplacian Based for Image, An Adaptive Image Enhancement Technique, An Improved Retinex Image Enhancement Technique, Automatic Exact Histogram Specification for, Bayesian Foreground and Shadow Detection in, Color Image Enhancement and Denoising Using an Optimized Filternet, Content Based Image Retrieval Using, Contrast Enhancement for Ziehl-Neelsen Tissue, DETAIL WARPING BASED VIDEO SUPER-RESOLUTION USING IMAGE GUIDES, Gray-level Image Enhancement By Particle Swarm, Image Contrast Enhancement based on Histogram, Image Enhancement and Segmentation Using Dark, Image Enhancement Technique Based On Improved, Image Quality Improvement fo r Electrophoretic Displays by, Image Reconstruction Using Particle Filters and, IMPROVED IDENTIFICATION OF IRIS AND EYELASH FEATURES, Improving Colour Image Segmentation on Acute, K�R ANALİZ Y�NTEMLERİ İLE İMGE İYİLEŞTİRME, NATURAL RENDERING OF COLOR IMAGE BASED ON RETINEX, Power-Constrained Contrast Enhancement, Research on Road Image Fusion Enhancement, Shadow Detection and Compensation in High Resolution Satellite Image Based, Smoothing Cephalographs Using Modified, Three-Dimensional Computational Integral Imaging Reconstruction by Use of, Towards integrating level-3 Features with

More Info
									                                                             2010 3rd International Congress on Image and Signal Processing (CISP2010)

  Image Contrast Enhancement based on Histogram
   Smoothing and Continuous Intensity Relocation
                      N.M. Kwok1 , Xiuping Jia2 , D. Wang1, S.Y. Chen3, Q.P. Ha4 and Gu Fang5
             School of Mechanical and Manufacturing Engineering, The University of New South Wales, Australia
              School of Engineering and Information Technology, The University of New South Wales, Australia
                            College of Computer Science, Zhejiang University of Technology, China
            School of Electrical, Mechanical and Mechatronic Systems, University of Technology Sydney, Australia
                                 School of Engineering, University of Western Sydney, Australia

   Abstract—Image contrast enhancement is a fundamental and           an evolutionary computation technique, the genetic algorithm
important pre-processing stage in the application of image            was applied to enhance image contrast [12]. On the other hand,
processing techniques. Among revenues of possible approaches,         image fusion approaches [13] had also been used in enhancing
histogram equalization is a popular and attractive candidate
method to produce resultant images of increased contrast. How-        image contrasts. In [14], image enhancement was tackled from
ever, images obtained from canonical histogram equalization           the point of view of noise filtering and edge boosting where
frequently produce artefacts and give rises to uncomfortable          the method was applied in colour images.
viewing particularly in homogeneous regions. In this work, the           Although the above-mentioned methods had shown promis-
problem is tackled using the histogram matching concept where         ing results in a variety of problem domains, histogram equal-
the intensity histogram of the input image is matched to its
smoothed version for contrast enhancement. Furthermore, pixel         ization still remains as one of the most popular technique
intensities are randomly perturbed for a remedy of artefacts.         for its implementation simplicity [15] and satisfactory perfor-
The resultant image intensities are thus distributed over the         mance in general. In a canonical implementation, the resultant
available range and an increased image contrast is derived.           image has a histogram resembling a linear transformation
Satisfactory results are obtained from experiments using a            or stretching from the original image histogram. In [16],
collection of outdoor scenery pictures to verify the effectiveness
of the proposed approach.                                             spatial relationships between neighbouring pixels were taken
                                                                      into consideration. Similarly, a local histogram equalization
                      I. I NTRODUCTION                                scheme was proposed in [17]. In [18], the input image were
   Image processing techniques have found a lot of applica-           sub-divided, equalized independently, and finally fused to
tions in a very wide domain of areas. For example, in [1],            produce a contrast-enhanced image. This approach was further
ancient paintings were restored by contrast enhancement and           developed in [19] where the original image is divided into
pattern synthesis. Vehicle licence plates could be located for        overlapping sub-blocks and equalized according to the pixel
transport automation using image processing methods [2]. In           characteristics within the block. In [20], the image histogram
the construction sector [3], surface defects are able to be au-       is matched to a distribution determined from a windowed and
tomatically detected. Images from cephalic radiography could          filtered version of the original histogram. Manipulations on
be enhanced for better diagnosis of illnesses [4]. Autonomous         the histograms were also frequently suggested by researchers.
vehicle navigation could be guided [5] with the use of vision.        These include specific considerations in minimizing the mean
The quality of remote sensing data could be improved using            brightness error between the input and output images [21]. In
image processing techniques [6]. In manufacturing [7], three-         [22], the maximum entropy or information content criterion
dimensional model construction could be facilitated by the use        was invoked in contrast enhancement.
of properly structured illumination in image acquisition.                In this work, the causes of the occurrence of artefacts in
   There are several designated demands from image process-           linear histogram equalized images are first demonstrated and
ing procedures. They include noise filtering, blur removal,            investigated. A histogram matching to a smoothed version
contrast enhancement, feature identification, and object recog-        of the original image histogram is proposed that removes
nition. Among these desirable outcomes, contrast enhancement          the artefacts. A further refinement is adopted with a per-
is one of the fundamental processes that critically affects the       turbed relocation of pixel intensities such that intensities are
quality of subsequent operations such as object recognition. In       continuously distributed across the available intensity range
the context of contrast enhancement, there are also a number          and increases the information content conveyed in the image.
of possible approaches. In [8], a morphological filter was used        Colour images of outdoor scenes are used to verify the
to sharpened images. The contrast could also be improved by           suggested approach. The results are assessed using the entropy
making use of the curvelet transform [9]. Similarly, a homo-          measure and quantified by objective viewing.
morphic filter was used in [10] to enhance the image contrast.            The rest of this paper is structured as follows. In Section II,
In the field of soft computing [11], the image contrast could          the conventional histogram equalization method is reviewed
be increased by a fuzzy intensification process. Alternatively,        and the generation of artefacts is revealed. The proposed

      978-1-4244-6516-3/10/$26.00 ©2010 IEEE                         688
alternatives using smoothed histogram matching and contin-                    by referring to the nth element in C d and override by, say, the
uous intensity relocation are presented in Section III and                    j th intensity. That is
IV respectively. In Section V, experiments carried out are
described and results are discussed. A conclusion is drawn                                            i     j
                                                                                                     Vuv → Vuv , for ci = cd .
                                                                                                                           j                                                         (6)
in Section VI.
                                                                                 The process aforementioned is here referred to as linear
                                                                              histogram equalization. This method is easy to implement
   Histogram equalization is a technique frequently applied                   but there are also limitations in its performance, particularly,
in the enhancement of image contrasts. Historically, this                     artefacts are created and gives rise to uncomfortable objective
procedure was used mostly in gray-level images. However,                      viewing.
this method is extendable to colour images when they are                         An example image of an outdoor scene illustrating the
converted to, operated on their intensity equivalents, and                    artefact phenomenon is depicted in Fig. 1 together with a
reverted to the colour space.                                                 histogram of the intensity space after the RGB to HSV
   Given the input or original colour image represented by                    conversion. It is easily observed that in homogeneous regions
            P = {Puv },        Puv = {Ruv Guv Buv },                    (1)   (the sky portion at top-left of Fig. 2(a)), the image is over-
                                                                              equalized and results in an artefact. A high magnitude of
where u, v are pixel coordinates in the width and height                      the so-called salt-and-pepper noise is observed. This makes
dimensions. Since the RGB space contains three colour related                 objective viewing uncomfortable. The cause of this drawback
signals, it is rather involved to operate on the three signal                 can be revealed from an inspection of the corresponding
spaces simultaneously for enhancement. Furthermore, since                     histogram drawn in Fig. 1(b). The original histogram is smooth
the human visual system is more sensitive to intensity variation              and has an elevated peak at the high intensity range, the sky
when accessing the image contrast, the image is converted                     in the picture. The result from linear histogram equalization
to the intensity domain before applying enhancement. For                      shown in Fig. 2(b), however, shows a breakout of the high
example, the image is converted to the hue-saturation-value                   intensity region in the histogram into a few isolated intensity
(HSV) format.                                                                 levels. It is evident that artefacts are produced due to an
                                                                              excessive extension of a concentration of partial histogram
             [Huv Suv Vuv ]T = Γ[Ruv Guv Buv ]T ,                       (2)
                                                                              into non-continuous portions in the intensity space.
where Γ is the format transformation matrix. The H com-                          In seeking a remedy of the artefact phenomenon, we sug-
ponent represents the colour tone, S denotes saturation, and                  gest to allocate the pixel intensities continuously across the
V corresponds to the image intensity. The canonical or linear                 available range. In order to achieve this goal, the smoothed his-
histogram equalization process proceeds as follows.                           togram matching and perturbed intensity relocation approaches
   A histogram is obtained from intensities Vuv , giving                      are proposed and described in the following sections.

          H = {hi },           hi = N,    i = 1, · · · , L,             (3)

where hi is the number of pixels in the image that have the                                                           12000

ith intensity level, N = u × v is the total number of pixels.                                                         10000

The number of levels is taken as L = 256 corresponding to                                                              8000

electronic displays in multiples of 8-bits (28 = 256) and the                                                          6000

maximum intensity is normalized to unity.                                                                              4000

   In principle, the image contrast would be enhanced if                                                                  0
                                                                                                                           0   0.1   0.2   0.3   0.4   0.5   0.6   0.7   0.8   0.9    1
the available intensity range can be used completely and
                                                                                               (a)                                               (b)
uniformly. A uniform histogram is therefore used where the
number of pixels that fall inside each intensity level are equal.                             Fig. 1.   Original image (a), histogram (b).
That is, the desired histogram is

        Hd = {hd },
               j          hd = N L−1 ,
                           j               j = 1, · · · , L.            (4)                                           14000


  To perform enhancement, two cumulative histograms are                                                               10000

constructed from the input and desired histograms respectively.                                                        8000

We have                                                                                                                6000

                     i                                      j
                                         {cd },   cd             hd .
  C = {ci }, ci =         hk ; and C =     j       j   =          k     (5)                                               0
                                                                                                                           0   0.1   0.2   0.3   0.4   0.5   0.6   0.7   0.8   0.9    1
                    k=1                                    k=1
                                                                                               (a)                                               (b)
   For a pixel, whose original intensity i at the nth position
in the cumulative histogram C, its equalized value is obtained                  Fig. 2.   Result from linear histogram equalization (a), histogram (b).

           III. S MOOTHED H ISTOGRAM M ATCHING                                                                        IV. P ERTURBED I NTENSITY R ELOCATION
   The linear histogram equalization procedure, as illustrated                                                 Consider the case with the linear histogram equalization
in the last section, attempts to match the original histogram                                               procedure. A bin in the histogram, in general, may contain
to a uniform distribution and artefacts occur mostly in ho-                                                 a count of pixels of more than one. The intensities of the
mogeneous regions in the image. In these regions, almost all                                                associated pixels are assigned to another value. In order to
pixels have a common intensity level but they were allocated                                                place the pixels continuously across the given intensity range,
by equalization to discrete intensity levels. Here, we propose to                                           we propose to allocate these pixels to more than one intensity
remove the gaps between intensities by matching the original                                                value. In order to preserve the image integrity, these values
histogram to a smooth distribution instead of a uniform                                                     cannot be separated from each other by a large intensity
distribution. The target histogram is designated as a running                                               distance. To this end, we place pixels to the neighborhood
average of the original histogram.                                                                          intensities according to the histogram matching principle.
   Let the original histogram be given by                                                                      Let there be m pixels of the same original intensity i and
                       H = {hi }, i = 1, 2, · · · , L,                                                (7)   they are to be allocated to intensity j. That is, we allocate
                                                                                                                                        i     j
where there are possibilities that some hi = 0 giving a                                                                                Vuv → Vuv .                          (10)
discontinuous histogram. A smoothed histogram is obtained                                                   From the smoothed histogram Hs , a cumulative histogram C s
from passing a running window through the bins of the                                                       is constructed according to equation (5). Giving
histogram. The smoothed histogram is given as
                    Hs = {hs }, i = 1, 2, · · · , L,
                           i                                                                          (8)                       C s = {cs }, cs =
                                                                                                                                        j     j           hs .
                                                                                                                                                           k                (11)
where each element of the histogram is                                                                                                              k=1

                    m                                                                                       The two, original and smoothed, cumulative histograms are
                m   j=1 hj , for 1      ≤m≤ L  2
    hs     =          L                                                                               (9)   aligned with pixels indexed by their entries in the cumulative
     i           1
                L−m   j= L +1 hj ,      for L + 1 ≤ m < L
                                            2                                                               histogram. We have

   This smoothed histogram then replaces the role of the                                                          i         j
                                                                                                               Vuv,m ↔ Vuv,n , i, j = 1, 2, · · · , L, m, n = 1, 2, · · · , N,
desired histogram Hd in equation (4) as the final target in the                                                                                                               (12)
histogram equalization process. The contrast-enhanced result                                                where pixels are indexed by subscripts m, n indicating their
is illustrated in Fig. 3.                                                                                   positions in different cumulative histograms. Because of the
   An inspection of Fig. 3(a) and a comparison with Fig. 2(a)                                               nature of the cumulative distribution, pixels with higher indices
could reveal that the artefact has been satisfactorily removed.                                             m, n have higher intensities.
It is further noticed that due to a reduced extension of the                                                   It is noted that the pixel counts, in most cases, do not result
intensity range, details in the texture are preserved. In Fig.                                              in multiples and the resultant histogram cannot be continuous.
3(b), the resultant histogram is depicted. It is seen that the                                              For example, if i < j, then a portion of pixels to be assigned to
discontinuity in the high intensity regions are mitigated. Pixels                                           intensity j needs to be derived from original pixels of intensity
originated from high intensities are assigned with values that                                              i+1. On the other hand, if i > j, then part of pixels of assigned
are different but close to their originals. In this context, we see                                         to intensity j will be obtained from pixels of original intensity
that the smoothed histogram approach is effective in producing                                              i − 1.
an image with satisfactory objective viewing comfort.                                                          In order to relax this constraint, we propose to distribute
   However, it is also noted that there are still gaps between
                                                                                                            the assignments across the neighbouring intensities. For a
histogram elements. On the other hand, the complete intensity                                                                                                          i
                                                                                                            particular intensity i, the corresponding pixels Vuv,m for
range has not yet been fully utilized to convey the maximum
                                                                                                            m = mmin · · · mmax are shuffled by drawing a sample index
amount of information to the viewer. A further refinement in
                                                                                                            from a uniform probability distribution. That is,
contrast enhancement is therefore proposed in the next section
                                                                                                                                     i       j
with the aim to increase the information content.                                                                                   Vuv,m → Vuv,m′ ,                        (13)

                                                                                                            where m′ ∼ U( · ), U( · ) ∈ [mmin , mmax ] is an uniform
                                       14000                                                                probability distribution. The result is that pixels of the same
                                       12000                                                                original intensity from a spatial diversity are assigned with
                                                                                                            new intensities. A contrast-enhanced image using this method
                                                                                                            is shown in Fig. 4.

                                                                                                               It is observed from Fig. 4(a) that salt-and-pepper artefacts
                                                                                                            do not appear in the result. The image also appears satisfactory
                                            0   0.1   0.2   0.3   0.4   0.5   0.6   0.7   0.8   0.9    1
                                                                                                            for objective viewing. An inspection of Fig. 4(b) illustrates that
                 (a)                                              (b)
                                                                                                            the histogram is continuous and covers the complete intensity
                                                                                                            spectrum. The performance of the method is further verified
 Fig. 3.   Result from smoothed histogram equalization (a), histogram (b).                                  in experiments described below.


                                            9000                                                               histogram is developed such that the information contained

                                                                                                               in the image can be better visualized by making use of the
                                            6000                                                               complete intensity range. Satisfactory results from experiments
                                                                                                               conducted on a set of colour images are obtained and verified

                                                                                                               the performance of the proposed approach.

                                            1000                                                                                            R EFERENCES
                                                0   0.1   0.2    0.3   0.4   0.5   0.6   0.7   0.8   0.9   1
                                                                                                                [1] S. C. Pei, Y. C. Zeng, and C. H. Chang, “Virtual restoration of ancient
                  (a)                                                  (b)                                          chinese paintings using color contrast enhancement and lacuna texture
                                                                                                                    synthesis,” IEEE Trans. on Image Processing, vol. 13(3), pp. 416–429,
Fig. 4.   Result from continuous histogram equalization (a), histogram (b).                                         2004.
                                                                                                                [2] D. Zheng, Y. Zhao, and J. Wang, “An efficient method of license plate
                                                                                                                    location,” Pattern Recognition Letters, vol. 26, pp. 2431–2438, 2005.
                                                                                                                [3] S. Lee, L. M. Chang, and M. Skibniewski, “Automated recognition of
                            V. E XPERIMENTS                                                                         surface defects using digital color image processing,” Automation in
   A collection of 40 test images were taken under a variety                                                        Construction, vol. 15, pp. 540–549, 2006.
                                                                                                                [4] J. Frosio, G. Ferrigno, and N. A. Borghese, “Enhancing digital cephalic
of environment conditions in outdoor scenes. In addition to                                                         radiography with mixture models and local gamma correction,” IEEE
objective viewing, two further assessment metrics are also                                                          Trans. on Medical Imaging, vol. 25(1), pp. 113–121, 2006.
examined. These are the entropy and the contrast metric, which                                                  [5] N. M. Kwok and A. B. Rad, “A modified particle filter for simultaneous
                                                                                                                    localization and mapping,” Journal of Intelligent and Robotic Systems:
is given by                                                                                                         Theory and Applications, vol. 46(4), pp. 365–482, 2006.
                                       1            1                                                           [6] F. Li, X. Jia, and D. Fraser, “Superresolution reconstruction of multi-
                                                                                                                    spectral data for improved classification,” IEEE Geoscience and Remote
            τuv = |9 × Vuv −                                    Vu+m,v+n |.                           (14)          Sensing Letters, vol. 6(4), pp. 689–693, 2009.
                                    m=−1 n=−1                                                                   [7] S. Y. Chen, Y. F. Li, and J. W. Zhang, “Vision processing for realtime
                                                                                                                    3d data acquisition based on coded structured light,” IEEE Trans. on
The entropy adopted is the Shannon’s entropy calculated on                                                          Image Processing, vol. 17(2), pp. 167–176, 2008.
all test images in the intensity domain. The contrast metric                                                    [8] J. G. M. Schavemaker, M. J. T. Reinders, J. J. Gerbrands, and E. Backer,
is also calculated in the colour space. The two assessment                                                          “Image sharpening by morphological filtering,” Pattern Recognition,
                                                                                                                    vol. 33, pp. 997–1012, 2000.
metrics, quantifying the information content, are illustrated in                                                                                          e
                                                                                                                [9] J. L. Starck, F. Murtagh, E. J. Cand` s, and D. L. Donoho, “Gray and
box plots in Fig. 5.                                                                                                color image contrast enhancement by the curvelet transform,” IEEE
   In the plot of entropies, Fig. 5(a), original image entropies                                                    Trans. on Image Processing, vol. 12(6), pp. 706–717, 2003.
                                                                                                               [10] M. J. Seow and V. K. Asari, “Ratio rule and homomorphic filter for
(left column) contain a wide range of values. In particular,                                                        enhancement of digital colour image,” Neurocomputing, vol. 69, pp.
there is an extended lower quartile. The entropies from the                                                         954–958, 2006.
proposed approaches are shown in the next three columns.                                                       [11] M. Hanmandlu, D. Jha, and R. Sharma, “Color image enhancement by
                                                                                                                    fuzzy intensification,” Pattern Recognition Letters, vol. 24, pp. 81–87,
There is a noticeable increase in the entropies. Similarly,                                                         2003.
the plot of contrasts in Fig. 5(b) also illustrates increases of                                               [12] M. S. Shyu and J. J. Leou, “A genetic algorithm approach to color image
contrasts in the resultant images. A sample of resultant images                                                     enhancement,” Pattern Recognition, vol. 31(7), pp. 871–880, 1998.
                                                                                                               [13] Q. Chen, X. Xu, Q. Sun, and D. Xia, “A solution to the deficiences of
obtained from the proposed methods, together with the original                                                      image enhancement,” Signal Processing, vol. 90, pp. 44–56, 2010.
images, is shown in figures 6 to 9 where satisfactory contrast                                                  [14] W. C. Kao and Y. J. Chen, “Multistage bilateral noise filtering and
enhancements are observed.                                                                                          edge detection for color image enhancement,” IEEE Trans. on Consumer
                                                                                                                    Electronics, vol. 51(4), pp. 1346–1351, 2005.
                           VI. C ONCLUSION                                                                     [15] H. D. Cheng and X. J. Shi, “A simple and effective histogram equal-
                                                                                                                    ization approach to image enhancement,” Digital Signal Processing,
   This paper has presented methods for image contrast en-                                                          vol. 14, pp. 158–170, 2004.
hancement. It is illustrated that artefacts frequently appears                                                 [16] T. K. Kim, J. K. Park, and B. S. Kang, “Contrast enhancement system
                                                                                                                    using spatially adaptive histogram equalization with temporal filtering,”
from the linear histogram equalization process and causes                                                           IEEE Trans. on Consumer Electronics, vol. 44(1), pp. 82–87, 1998.
uncomfortable viewing. In order to resolve this problem, pixel                                                 [17] H. Zhu, H. Y. Chan, and F. K. Lam, “Image contrast enhancement by
intensities in the original image histogram is matched to a                                                         constrained local histogram equalization,” Computer Vision and Image
                                                                                                                    Understanding, vol. 73(2), pp. 281–290, 1999.
smoothed version of itself. A further step giving a continuous                                                 [18] Y. Wang, Q. Chen, and B. Zhang, “Image enhancement based on equal
                                                                                                                    area dualistic sub-image histogram equalization method,” IEEE Trans.
                                                                                                                    on Consumer Electronics, vol. 45(1), pp. 68–75, 1999.
                                                                                                               [19] J. Y. Kim, L. S. Kim, and S. H. Hwang, “An advanced contrast en-
                                                                                                                    hancement using partially overlapped sub-block histogram equalization,”
                                                                                                                    IEEE Trans. on Circuits and Systems for Video Technology, vol. 11(4),
                                                                                                                    pp. 475–484, 2001.
                                                                                                               [20] J. A. Stark, “Adaptive image contrast enhancement using generalizations
                                                                                                                    of histogram equalization,” IEEE Trans. on Image Processing, vol. 9(5),
                                                                                                                    pp. 889–896, 2000.
                                                                                                               [21] S. D. Chen and A. R. Ramli, “Minimum mean brightness error bi-
                  (a)                                                   (b)                                         histogram equalization in contrast enhancement,” IEEE Trans. on Con-
                                                                                                                    sumer Electronics, vol. 49(4), pp. 1310–1319, 2003.
Fig. 5. Box plots of (a) entropies and contrast metrics (b): (i) original image,                               [22] C. Wang and Z. Ye, “Brightness preserving histogram equalization with
(ii) linear histogram equalization, (iii) smoothed histogram equalization and                                       maximum entropy: a variational perspective,” IEEE Trans. on Consumer
(iv) continuous histogram equalization.                                                                             Electronics, vol. 51(4), pp. 1326–1334, 2005.

              Fig. 6.   Original images.

   Fig. 7.     Liner histogram equalized images.

Fig. 8.      Smoothed histogram equalized images.

Fig. 9.   Continuous histogram equalized images.


To top