Anish Kumar Vishwakarma et al / Indian Journal of Computer Science and Engineering (IJCSE)

Color Image Enhancement Techniques:

A Critical Review
Anish Kumar Vishwakarma
Department of Electronics and Communication Engineering
Jabalpur Engineering College, Jabalpur (M.P.), India
E-mail: [email protected]
Agya Mishra
Department of Electronics and Communication Engineering
Jabalpur Engineering College, Jabalpur (M.P.), India
E-mail: [email protected]
Abstract
Image enhancement is one of the key issues in high quality pictures such as digital camera and HDTV. Since
Image clarity is very easily affected by lighting, weather, or equipment that has been used to capture the image.
These conditions lead to image may suffer from loss of information. As a result, many techniques have
developed known as image Enhancement techniques to recover the information in an image. This paper presents
a literature review on some of the image Enhancement techniques for color image enhancement like, Contrast
Stretching, Histogram Equalization and its improvement versions, Homomorphic Filtering, Retinex, and
Wavelet Multiscale Transform. Comparison of all the techniques concludes the better approach for its future
research.
Keywords: Image enhancement; Homomorphic Filtering; Retinex

1. Introduction
Color images provide more and richer information for visual perception than that of the gray images. Color
image enhancement plays an important role in Digital Image Processing [1]. The purpose of image enhancement
is to get finer details of an image and highlight the useful information. During poor illumination conditions, the
images appear darker or with low contrast. Such low contrast images needs to be enhanced. In the literature
many image enhancement techniques such as gamma correction, contrast stretching, histogram equalization, and
Contrast-limited adaptive histogram equalization (CLAHE) have been discussed. These are all old techniques
which will not provide exact enhanced images and gives poor performance in terms of Root Mean Square Error
(RMSE), Peak Signal to Noise Ratio (PSNR) and Mean Absolute Error (MAE) [5]. Use of the old enhancement
technique will not recover exact true color of the images. Recently, Retinex, Homomorphic and Wavelet Multi-
Scale techniques have been popular for enhancing images. These methods are shown to perform much better
than those listed earlier [7]. This report is organized as follows: Section II presents brief description of image
Enhancement techniques, Section III gives comparison between various enhancement techniques in tabular
form, Section IV presents performance comparison of those techniques and finally, conclusion is presented in
Section V.

2. Image Enhancement Techniques

Image enhancement is basically improving the Interpretability or perception of information in images for human
viewers and providing better input for other automated image processing techniques [1]. There exist many
techniques that can enhance a digital image without spoiling it. The enhancement methods can broadly be
divided in to the following two categories:
1. Spatial Domain Methods 2. Frequency Domain Methods

In spatial domain techniques, we directly deal with the image pixels. The pixel values are manipulated to
achieve desired enhancement. In frequency domain methods, the image is first transferred in to frequency
domain. It means that, the Fourier Transform of the image is computed first. All the enhancement operations are
performed on the Fourier transform of the image and then the Inverse Fourier transform is performed to get the
resultant image. Image enhancement is applied in every field where images are ought to be understood and
analyzed. For example, medical image analysis, analysis of images from satellites etc. In this section we briefly
describe the various image enhancement techniques.

ISSN : 0976-5166 Vol. 3 No. 1 Feb -Mar 2012 39

 Anish Kumar Vishwakarma et al / Indian Journal of Computer Science and Engineering (IJCSE)

2.1. Contrast Stretching

Contrast stretching technique is used to stretch the dynamic range of an image. Dynamic range is the range
between the minimum intensity value and the maximum intensity value of an image. Mathematically, Contrast
Stretching is given by [3],

′
, , (1)

Where, ′ , is the new dynamic range image, d is the new dynamic range value, , is the input image,
is the minimum intensity value of the input image, is the maximum intensity value of the input image,
and is the offset point of the new dynamic range for ′ , . This transformation will provide good visual
representation of the original scene but some of the detail maybe loss due to saturation and clipping as well as
due to poor visibility in under-exposure regions of the image.

2.2. Histogram Equalization

Histogram Equalization (HE) [1], is a technique that made contrast adjustment using image’s histogram. This
technique is based on the idea of remapping the histogram of the scene to a histogram that has a near-uniform
probability density function. Histogram Equalization redistributes intensity distribution. If the histogram of any
image has many peaks and valleys, it will have peaks and valleys after equalization but the peaks and valleys
will be shifted. This technique improves contrast and the goal of Histogram Equalization is to obtain a uniform
histogram. In general, Histogram Equalization can be divided into three types, Global Histogram Equalization
(GHE), Adaptive Histogram Equalization (AHE), and Block-based Histogram Equalization (BHE) [4].
In Global Histogram Equalization (GHE), each pixel is assigned a new intensity value based on
previous cumulative distribution function. To perform Global Histogram Equalization (GHE), the original
histogram of the grayscale image needs to be equalized. The cumulative histogram from the input image needs
to be equalized to 255 by creating the new intensity value by applying [4];

′
(2)

Where, ′ is the new intensity level, d is the new dynamic range value, is the offset point of new dynamic
range for ′ , C(x) is the normalized cumulative value, is the maximum value in normalized cumulative
value, and is the minimum value in normalized cumulative value. Lastly, the normalized cumulative
histogram is used as the mapping functions of the original image. This technique increased the contrast of the
image but lighting condition under uneven illumination may sometimes turn to be more uneven.

2.3. Homomorphic Filter

Homomorphic Filtering [3], is sometimes used for image enhancement. It simultaneously normalized the
brightness across an image and increases the contrast. Here, Homomorphic Filtering is used to remove
multiplicative noise. Illumination and reflectance are not separable, but their approximate locations in the
frequency domain may be located. Since illumination and reflectance are combined multiplicatively, the
component are made additive by taking the logarithm of the image intensity, so that these linearly in the
frequency domain. Illumination variations can be thought of as a multiplicative noise, and can be reduced by
filtering in the log domain.

Input image  Forward Linear Inverse Output image

Log Fourier Processing Fourier EXP
Transform Transform

Fig. 1. Block Diagram of Homomorphic Filtering[3]

2.4. Retinex

Retinex theory was first proposed by Edwin Land in 1964 [6]. There are many algorithms based on Retinex
theory such as Multi Scale Retinex (MSR) [8], Multi Scale Retinex with modified color restoration (MSRCR)

ISSN : 0976-5166 Vol. 3 No. 1 Feb -Mar 2012 40

 Anish Kumar Vishwakarma et al / Indian Journal of Computer Science and Engineering (IJCSE)

[7], Fast Multi Scale Retinex (FMSR) [10] etc. According to Land [6], image is composed by two parts namely
the incident light and the reflectance of the object. This can be represented by,

⁄ (3)

Where, L represents value of incident light, R represents the value of object’s reflection, and E represents the
value of reflected light.
Single Scale Retinex, is the most basic technique for Retinex based algorithm. In Single Scale Retinex
method proposed by Jobson et al. [9], the illumination , is estimated by applying a linear Low Pass Filter
(LPF) for an input color image, , . The output color image , is obtained by subtracting the log
signal of the estimated illumination, which is 2D convolution of Gaussian surround function and original image
of component as per Equation (4):

, log , log , , (4)

Where , , , , is the retinex output for channel ‘i’ , is the image value for channel,
denotes convolution and F(x, y) is a Gaussian surround function. The convolution , , represents
illumination estimation and is a convolution of Gaussian functions with original image.
The Gaussian Surround Function F(x, y) is given by Equation (5).

, , (5)

,
K is selected such that;
, 1
Where x & y are spatial coordinates, c is the standard deviation of Gaussian distribution that determines the
scale.

   Fig. 2. Image processing by single-scale Retinex [6]

 
The Multiscale retinex is given by Equation (6):

, ∑ , (6)

is the weighting factor. The Multiscale retinex with color restoration is given by Equation (7):

, , , , , (7)
, log , log ,

=gain constant, =strength of non linearity.

The comparison of the enhanced color images obtained from different techniques is of done by taking
parameters Mean Square Error (MSE), and Peak Signal to Noise Ratio (PSNR). These parameters for color
images are given by equation (8) and (9).

1 ′ ′ ′
, , , , , , 8

ISSN : 0976-5166 Vol. 3 No. 1 Feb -Mar 2012 41

 Anish Kumar Vishwakarma et al / Indian Journal of Computer Science and Engineering (IJCSE)

10log ⁄
(9)

Where , , , , , , ′ , , ′ , , ′ , are image pixels of original and enhanced image of image

size of and , , are the MSE between the R component, G component and B component
of the original and the enhanced image.

2.5. Wavelet Multi-scale Transform

Wavelet analysis [12] has proven to be a powerful image processing tools in recent years. When images are to
be viewed or processed at multiple resolutions, the wavelet transform (WT) is the mathematical tool of choice.
In addition to being an efficient, highly intuitive framework for the representation and storage of multi-
resolution images, the WT provides powerful insight into an images spatial and frequency characteristics. The
image detail parts are stored in the high-frequency parts of image transformed by wavelet and the imagery
constant part is stored in low-frequency part. Because the imagery constant part determines the dynamic range
of image, the low-frequency part determines the dynamic range of image. We attenuate the low-frequency part
in order to compress the dynamic range.
But the details must be loss when the low-frequency part is attenuated [11]. As some details are stored
in the high-frequency parts very well, the image reconstructed by inverse wavelet transform (IWT) has more
detail.

3. Performance Comparison
 In [4] simulation results of different histogram equalization techniques were compared, for comparison
three quality measurements were used: flatness ( ), Contrast-per-pixel (C), and average absolute mean
brightness error (AAMBE).Observed data from [4] is compared as given in Table-1.

Table 1. Comparison of different histogram equalization techniques

Method Flatness Value Contrast per pixel AAMBE Average Execution
Time (ms)
GHE 2079.2001 4.6450 35.9358 30.93
BBHE 1226.1185 4.2699 22.1924 32.19
DSIHE 1217.9784 4.5453 16.7719 36.77
MMBEBHE 1217.7566 4.2225 18.5597 65.55
BHENM 115.9456 5.1296 16.6760 90.67

This paper concludes that histogram equalization extension BHENM technique shows the improved image
contrast, flatness, and simultaneously preserved the image brightness as compared to previous version of
Histogram equalization techniques, but it requires 90.67 ms (Table 1) average execution time which is three
times higher than the GHE.
 In [5] different color image enhancement techniques were compared by taking parameters Root Mean
Square Error (RMSE), Peak Signal to Noise Ratio (PSNR), and Mean Absolute Error (MAE), which is
presented in Table 2. In this paper Retinex model with different surround function such as Gaussian, Laplacian,
and Gamma distribution as the illumination estimation function were used and there results were also presented
in Table 2. This paper conclude that Use of the Laplacian distributed surround function gives some better results
compared to Gaussian distribution function. And also the use of the Gamma distributed surround function gives
still better results compared to both Gaussian and Laplacian distribution function. Compared to Histogram
Equalization, Gamma Correction and Homomorphic filtering techniques, the Retinex method of image
enhancement is giving us more contrast enhanced results without destroying any parts of the image that does not
require enhancement.

Table 2. [5]
Errors Measurment Homom- Homomorphic Gaussian Laplacian Gamma Gamma Histogram CLAHE
Between orphic Method Surround Distribution Distribution Correction Equalization
Original Method with
& Noisy Thresholding

RMSE 128.93 125.44 105.38 75.16 75.47 71.99 124.54 53.17 102.05
PSNR 13.56 14.10 17.59 24.35 24.27 25.21 14.25 31.27 18.23
MAE 113.89 111.39 91.02 65.04 65.15 62.77 110.61 39.55 85.31

ISSN : 0976-5166 Vol. 3 No. 1 Feb -Mar 2012 42

 Anish Kumar Vishwakarma et al / Indian Journal of Computer Science and Engineering (IJCSE)

 The papers [6],[7],[8] describe performance of Retinex model with different color correction and color
restoration techniques. An efficient color correction technique with single scale Retinex (SSR) is shown in
figure 3 which was used in [6].

Fig. 3. Luminance signal processing [6]

This method shows good contrast improvement compare to Gamma correction, Histogram equalization, and its
improved methods denoted by BBHE, DSIHE, RMSHE and RSIHE, but this method has still large computation
time because of large scale Gaussian filter.
In [8] Adaptive Multi Scale Retinex with Gaussian filter of various sizes were used. A visual contrast
measure (VCM) and halo artifact measure were used to evaluate the image produced using various Gaussian
filters and weights. Figure 4 shows VCM results corresponding to different test images and figure 5 shows the
averaged maximum color differences using an SSR (single Scale Retinex) model. In the case of a Gaussian filter
less than 80, the color difference was greater than three, resulting in halo artifacts. Therefore, when considering
the VCM and halo artifacts, the size of the large Gaussian filter must be greater than 200.

Fig. 4. VCM values for test images varying scale of Fig. 5. Halo artifacts for varying the weight of large
Gaussian filter [8] Gaussian filter [8]

This paper conclude that proposed color correction method based on MSR (Multi Scale Retinex) takes into
account the dominant chromaticity in an image to improve local contrast and color rendition.
In [7] a new technique for the enhancement of color image using Multi scale Retinex with modified
color restoration was proposed. The operational sequence of the proposed algorithm for Multi Scale Retinex
method based on modified color restoration technique is shown in figure 6. Table 3 shows that proposed method
in this paper has nearly the same contrast performance and PSNR values compared to NASA’s Multi Scale
Retinex method of Jobson’s et al.

Table 3. Multi scale Retinex with color restoration [7]

Parameter NASA Proposed
Method Method in[7]
Contrast 0.7717 0.7362
Luminance 1.3281 1.8091
PSNR (dB) 37.6 37.8

ISSN : 0976-5166 Vol. 3 No. 1 Feb -Mar 2012 43

 Anish Kumar Vishwakarma et al / Indian Journal of Computer Science and Engineering (IJCSE)

This method provides balance between contrast and luminance. This method offers better pixel distribution
compared to other methods, but performs in unnatural color rendition.

Original Image (RGB)

Separate into R, G, B Channels

Convolve with Gaussian Surround

Single Scale Retinex

Multi Scale Retinex

Color Restoration Technique

Combine R, G, B Channels

Enhanced Color Image

Fig. 6.  Flow Sequence of Color Restoration Method [7]

4. Comparison of Different Techniques

Comparison of observations given in all references is discussed here.
Table 4. Comparison between different color image enhancement techniques
S No. Enhancement Domain Measuring Advantages Disadvantages
Technique/Algorithm Parameters
1. Contrast Stretching Spatial Good visual representation some of the detail maybe loss due to
[3] - of the original scene. saturation and clipping.

2. Histogram Spatial RMSE- 53.17 Image has uniform Cannot adapt the local information of
Equalization PSNR (dB)- histogram the image and preserve the brightness
[2] [4] 31.27 Produce optimal of the original image.
contrast
Fast
3. Contrast Limited Spatial RMSE -102.05 Enhanced local contrast Noise amplification in flat region and
Adaptive Histogram PSNR (dB)- ring artifacts at strong edges.
Equalization 18.23
[2] [4]
4. Homoporphic Frequency RMSE -125.44 Remove multiplicative Problem of bleaching of the image
Filtering PSNR (dB)- Noise
[3] [5] 14.10
Simultaneous gray-level
range compression and
contrast enhancement
5. Single Scale Retinex Frequency Exceptional promise for Gray level violation problem
[6] [8] - dynamic range ‘Washed out’appearance
compression
6. Multi Scale Retinex Frequency PSNR (dB)- Provides dynamic Unnatural color rendition
[7] [8] 37.8 range compression
Preserve most of the detail ‘Washed out’appearance but less than
SSR

Above table concludes that classical method Contrast stretching losses some of the detail information of the
images during enhancement, Histogram Equalization gives better results but it can not preserve the brightness of
the original image, so Adaptive histogram Equalization, homomorphic filtering overcome this problem but not
for the multiplicative noise. Retinex is the advance technique for the color image enhancement because it is
based on color constancy theory and provide dynamic range compression and preserve most of the detail of an

ISSN : 0976-5166 Vol. 3 No. 1 Feb -Mar 2012 44

 Anish Kumar Vishwakarma et al / Indian Journal of Computer Science and Engineering (IJCSE)

image but it still has the problem observed in above table, so it is necessary to use some transform for the
remedy of this problem.

5. Conclusion

Critical review concludes that Histogram equalization cannot preserve the brightness and color of the original
image and Homomorphic filtering technique has the problem of bleaching of the image. Modern techniques
Retinex (SSR & MSR) performs much better than those listed above because it is based on the color constancy
theory, but it still suffers from color violation and unnatural color rendition problem, as The Wavelet transforms
is the very good technique for the image denoising and input images always faces the noise during image
processing so in future work there is scope of applying Wavelet transform with Retinex to improve the image
enhancement results such as color rendition problem, PSNR, minimum MSE .

References
[1] Rafael, C. Gonzalez and Richard E. Woods. “Digital Image Processing,” 2nd edition, Prentice Hall, 2002.
[2] Gabriel Thomas, Daniel Flores-Tapia, and Stephen Pistorius, “Histogram Specification: A Fast and Flexible Method to Process Digital
Images,” IEEE Transactions on Instrumentation and Measurement, Vol. 60, no. 5, May 2011, pp.1565-1578
[3] Mohd Firdaus Zakaria, Haidi Ibrahim, and Shahrel Azmin Suandi, “A Review: Image Compensation Techniques,”2nd International
Conference on Computer Engineering and Technology 2010, vol. 7, pp. 404-408.
[4] Nyamlkhagva Sengee, Altansukh Sengee, and Heung-Kook Choi, “Image Contrast Enhancement using Bi-Histogram Equalization
with Neighborhood Metrics,” IEEE Trans.Consumer Electronics, vol.56, no. 4, pp. 2727-2734, Nov 2010.
[5] Dileep MD, and A. Sreenivasa Murthy, “A Comparison Between Different Color Image Contrast Enhancement Algorithms,”
Proceedings of IEEE ICETECT 2011, pp. 708-712.
[6] Youhei Terai, Tomio Goto, Satoshi Hirano, and Masaru Sakuari, “Color Image Contrast Enhanncement by Retinex Model”, 13th IEEE
International Symposium on Consumer Electronics 2009, pp. 392-393.
[7] Hanumantharaju M.C., Ravishankar M., Rameshbabu D.R., and Ramchandran S., “Color Image Enhancement using Multiscale
Retinex with Modified Color Restoration Technique,” Second International Conference on Emerging Applications of Information
Tecnology 2011, pp. 93-97.
[8] In-su Jang, Tae-Hyoung Lee, Ho-Gun Ha, and Yeong-Ho Ha, “Adaptive Color Enhancement Based on Multi-Scaled Retinex using
Local Contrast of the Input Image” 2011.
[9] Li He, Ling Luo, and Jin Shang, “An Image Enhancement Algorithm Based Retinex Theory,” International Workshop on Education
Technology and Computer Science 2009, pp. 350-352.
[10] Xiong Jie, Han Li-na, Geng Guo-hua, and Zhou Ming-quan, “Real Color Image Enhancement by Illumination-Reflectance Model and
Wavelet Transform,” International Forum on Information Technology and Aplications 2009, pp. 691-695.
[11] Feng Xiao, Mingquan Zhou, and Guohua Geng, “Detail Enhancement and Noise Reduction with Color Image Detection Based on
Wavelet Multi-scale,” 2011, pp. 1061-1064.
[12] Sangjin Kim, Wonseok Kang, Eunsung Lee, and Joonki Paik, “Wavelet-Domain Color Image Enhancement using Filtered Directional
Bases and Frequency- Adaptive Shrinkage,” IEEE Trans. on consumer Electronics, vol. 56, no. 2, May 2010.

ISSN : 0976-5166 Vol. 3 No. 1 Feb -Mar 2012 45