ieeegeoscience and remote sensing letters 1 contrast

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 1

Contrast Enhancement Using Dominant BrightnessLevel Analysis and Adaptive Intensity

Transformation for Remote Sensing ImagesEunsung Lee, Sangjin Kim, Wonseok Kang, Doochun Seo, and Joonki Paik, Senior Member, IEEE

Abstract—This letter presents a novel contrast enhancement ap-proach based on dominant brightness level analysis and adaptiveintensity transformation for remote sensing images. The proposedalgorithm computes brightness-adaptive intensity transfer func-tions using the low-frequency luminance component in the waveletdomain and transforms intensity values according to the trans-fer function. More specifically, we first perform discrete wavelettransform (DWT) on the input images and then decompose theLL subband into low-, middle-, and high-intensity layers using thelog-average luminance. Intensity transfer functions are adaptivelyestimated by using the knee transfer function and the gammaadjustment function based on the dominant brightness level ofeach layer. After the intensity transformation, the resulting en-hanced image is obtained by using the inverse DWT. Althoughvarious histogram equalization approaches have been proposedin the literature, they tend to degrade the overall image qualityby exhibiting saturation artifacts in both low- and high-intensityregions. The proposed algorithm overcomes this problem usingthe adaptive intensity transfer function. The experimental resultsshow that the proposed algorithm enhances the overall contrastand visibility of local details better than existing techniques. Theproposed method can effectively enhance any low-contrast imagesacquired by a satellite camera and is also suitable for other variousimaging devices such as consumer digital cameras, photorealistic3-D reconstruction systems, and computational cameras.

Index Terms—Adaptive intensity transfer function, contrast en-hancement, discrete wavelet transform (DWT), dominant bright-ness level analysis, remote sensing images.

I. INTRODUCTION

FOR several decades, remote sensing images have playedan important role in many fields such as meteorology,

agriculture, geology, education, etc. As the rising demandfor high-quality remote sensing images, contrast enhancementtechniques are required for better visual perception and colorreproduction.

Manuscript received November 1, 2011; revised February 25, 2012; acceptedMarch 16, 2012. This work was supported by the Chung-Ang Universityexcellent freshman scholarship grants, by the Basic Science Research Pro-gram through the National Research Foundation of Korea funded by theMinistry of Education, Science, and Technology (2009-0081059), and by theKorea Aerospace Research Institute in the Korea Multi-Purpose Satellite-3A(KOMPSAT-3A) project.

E. Lee, S. Kim, W. Kang, and J. Paik are with the Department of Image,Chung-Ang University, Seoul 156-756, Korea (e-mail: [email protected]).

D. Seo is with the Calibration/Validation Team, Satellite Data InformationDepartment, Korea Aerospace Research Institute, Daejeon 305-333, Korea.

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/LGRS.2012.2192412

Histogram equalization (HE) [1] has been the most popularapproach to enhancing the contrast in various application areassuch as medical image processing, object tracking, speechrecognition, etc. HE-based methods cannot, however, maintainaverage brightness level, which may result in either under- oroversaturation in the processed image. For overcoming theseproblems, bi-histogram equalization (BHE) [2] and dualisticsubimage HE [3] methods have been proposed by using decom-position of two subhistograms. For further improvement, therecursive mean-separate HE (RMSHE) [4] method iterativelyperforms the BHE and produces separately equalized subhis-tograms. However, the optimal contrast enhancement cannot beachieved since iterations converge to null processing.

Recently, the gain-controllable clipped HE (GC-CHE) hasbeen proposed by Kim and Paik [5]. The GC-CHE methodcontrols the gain and performs clipped HE for preserving thebrightness. Demirel et al. have also proposed a modified HEmethod which is based on the singular-value decompositionof the LL subband of the discrete wavelet transform (DWT)[6], [7]. In spite of the improved contrast of the image, thismethod tends to distort image details in low- and high-intensityregions.

In remote sensing images, the common artifacts causedby existing contrast enhancement methods, such as driftingbrightness, saturation, and distorted details, need to be mini-mized because pieces of important information are widespreadthroughout the image in the sense of both spatial locations andintensity levels. For this reason, enhancement algorithms forsatellite images not only improve the contrast but also minimizepixel distortion in the low- and high-intensity regions.

To achieve this goal, we present a novel contrast enhance-ment method for remote sensing images using dominant bright-ness level analysis and adaptive intensity transformation. Morespecifically, the proposed contrast enhancement algorithm firstperforms the DWT to decompose the input image into a setof band-limited components, called HH, HL, LH, and LLsubbands. Because the LL subband has the illumination infor-mation [6], the log-average luminance is computed in the LLsubband for computing the dominant brightness level of theinput image [8]–[10]. The LL subband is decomposed into low-,middle-, and high-intensity layers according to the dominantbrightness level.

The adaptive intensity transfer function is computed in threedecomposed layers using the dominant brightness level, theknee transfer function [11], and the gamma adjustment function

1545-598X/$31.00 © 2012 IEEE


2 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS

Fig. 1. Block diagram of the proposed contrast enhancement algorithm.

[12], [13]. Then, the adaptive transfer function is applied forcolor-preserving high-quality contrast enhancement. The re-sulting enhanced image is obtained by the inverse DWT(IDWT).

II. ANALYSIS OF DOMINANT BRIGHTNESS LEVELS

In spite of increasing demand for enhancing remote sensingimages, existing histogram-based contrast enhancement meth-ods cannot preserve edge details and exhibit saturation artifactsin low- and high-intensity regions. In this section, we present anovel contrast enhancement algorithm for remote sensing im-ages using dominant brightness level-based adaptive intensitytransformation as shown in Fig. 1.

If we do not consider spatially varying intensity distribu-tions, the correspondingly contrast-enhanced images may haveintensity distortion and lose image details in some regions. Forovercoming these problems, we decompose the input imageinto multiple layers of single dominant brightness levels.

To use the low-frequency luminance components, we per-form the DWT on the input remote sensing image and thenestimate the dominant brightness level using the log-averageluminance in the LL subband [8]. Since high-intensity valuesare dominant in the bright region, and vice versa, the dominantbrightness at the position (x, y) is computed as

D(x, y) = exp

1

NL

∑(x,y)∈S

{logL(x, y) + ε}

(1)

where S represents a rectangular region encompassing (x, y),L(x, y) represents the pixel intensity at (x, y), NL representsthe total number of pixels in S, and ε represents a sufficientlysmall constant that prevents the log function from diverging tonegative infinity.

The decomposed low-, middle-, and high-intensity layersare shown in Fig. 2. The low-intensity layer has the dominantbrightness lower than the prespecified low bound. The high-intensity layer is determined in the similar manner with theprespecified high bound, and the middle-intensity layer has

Fig. 2. Image decomposition based on the dominant brightness levels andcontrast enhancement results. (a) Original image. (b) Dominant intensityanalysis. (c) Enhanced result image. (d–f) Low-, middle-, and high-intensitylayers. (g–i) Enhanced low-, middle-, and high-intensity layers.

the dominant brightness in between low and high bounds. Thenormalized dominant brightness varies from zero to one, and itis practically in the range between 0.5 and 0.6 in most images.For safely including the practical range of dominant brightness,we used 0.4 and 0.7 for the low and high bounds, respectively.

III. EDGE-PRESERVING CONTRAST ENHANCEMENT USING

ADAPTIVE INTENSITY TRANSFORMATION

Based on the dominant brightness in each decomposedlayer, the adaptive intensity transfer function is generated.


LEE et al.: CONTRAST ENHANCEMENT 3

Fig. 3. (a) Knee transfer functions for three layers using the correspondingknee points and spline interpolation. bl and bh represent low and high bounds,respectively, of intensity, and ©,×, and � represent low-, middle-, and high-intensity layers, respectively. (b) Adaptive intensity transfer functions for threelayers.

Since remote sensing images have spatially varying intensitydistributions, we estimate the optimal transfer function in eachbrightness range for adaptive contrast enhancement. The adap-tive transfer function is estimated by using the knee trans-fer [11] and the gamma adjustment functions [12], [13]. Forthe global contrast enhancement, the knee transfer functionstretches the low-intensity range by determining knee pointsaccording to the dominant brightness of each layer as shown inFig. 3(a). More specifically, in the low-intensity layer, a singleknee point is computed as

Pl = bl + wl(bl −ml) (2)

where bl represents the low bound, wl represents the tuningparameter, and ml represents the mean of brightness in the low-intensity layer.

For the high-intensity layer, the corresponding knee point iscomputed as

Ph = bh − wh(bh −mh) (3)

where bh represents the high bound, wh represents the tuningparameter, and mh represents the mean brightness in the high-intensity layer.

In the middle-intensity layer, two knee points are computedas

Pml = bl − wm(bml −mm) + (Pl − Ph) (4)

Pmh = bh + wm(bmh −mm) + (Pl − Ph) (5)

where wm represents the tuning parameter and mm representsthe mean brightness in the middle-intensity layer.

The global image contrast is determined by tuning parameterwi for i ∈ {l,m, h}. Although the contrast is more enhancedas the wi increases, the resulting image is saturated and con-tains intensity discontinuity. In this letter, we can adjust onlythe middle-intensity tuning parameter wm for reducing suchartifacts.

Fig. 3(a) shows knee transfer functions for three layers usingthe corresponding knee points and spline interpolation.

Fig. 4. (a) Original image from the Computer Vision Group-University ofGranada image database [16]; contrast-enhanced images by using (b) thestandard HE, (c) RMSHE, (d) GC-CHE, (e) Demirel’s, and (f) the proposedmethods.

Since the knee transfer function tends to distort image detailsin the low- and high-intensity layers, additional compensa-tion is performed using the gamma adjustment function. Thegamma adjustment function is modified from the original ver-sion by scaling and translation to incorporate the knee transferfunction as

Gk(L) =

{(L

Mk

) 1γ

−(1− L

Mk

) 1γ

+ 1

},

for k ∈ {l,m, h} (6)

where M represents the size of each section intensity range,such as Ml = bl, Mm = bh − bl, and Mh = 1− bh, L rep-resents the intensity value, and γ represents the prespecifiedconstant.

The prespecified constant γ can be used to adjust the localimage contrast. As γ increases, the resulting image is saturatedaround bl/2, bh − bl/2, and 1− bh/2. Therefore, the γ valueis selected by computing maximum values of adaptive transferfunction in ranges {0 ≤ L < (bl/2)}, {bl ≤ L < (bh − bl/2)},and {bh ≤ L < (1− bh/2)}, which are smaller than bl/2, bh −bl/2, and 1− bh/2, respectively.

The proposed adaptive transfer function is obtained by com-bining the knee transfer function and the modified gammaadjustment function as shown in Fig. 3(b). Three intensity-transformed layers by using the adaptive intensity transfer func-tion are fused to make the resulting contrast-enhanced imagein the wavelet domain. We extract most significant two bitsfrom the low-, middle-, and high-intensity layers for generating


4 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS

Fig. 5. (a) Original satellite image from Korea Aerospace Research Institute(KARI); contrast-enhanced images by using (b) the standard HE, (c) RMSHE,(d) GC-CHE, (e) Demirel’s, and (f) the proposed methods.

the weighting map, and we compute the sum of the two-bit values in each layer. We select two weighting maps thathave two largest sums. For removing the unnatural bordersof fusion, weighting maps are employed with the Gaussianboundary smoothing filter. As a result, the fused image F isestimated as

F = W1×cl+(1−W1)×{W2×cm+(1−W2)×ch} (7)

where W1 represents the largest weighting map, W2 representsthe second largest weighting map, cl represents the contrast-enhanced brightness in the low-intensity layer, cm representsthe contrast-enhanced brightness in the middle-intensity layer,and ch represents the contrast-enhanced brightness in thehigh-intensity layer. Since (7) represents the point operation,the pixel coordinate (x, y) is omitted. The fused LL sub-band undergoes the IDWT together with the unprocessed HL,LH, and HH subbands to reconstruct the finally enhancedimage.

IV. EXPERIMENTAL RESULTS

For evaluating the performance of the proposed algorithm,we tested three low-contrast remote sensing images as shownin Figs. 4–6(a). The performance of the proposed algorithmis compared with existing well-known algorithms includingstandard HE, RMSHE, GC-CHE, and Demirel’s methods.

For the experiment, we used γ = 1.4, bl = 0.4, and bh =0.7. For three different intensity layers, wl = 1, wm = 3, andwh = 1 were used.

Fig. 6. (a) Original satellite image from KARI; contrast-enhanced images byusing (b) the standard HE, (c) RMSHE, (d) GC-CHE, (e) Demirel’s, and (f) theproposed methods.

Figs. 4–6 show the results of contrast enhancement using thestandard HE, RMSHE, GC-CHE, Demirel’s, and the proposedmethods. As shown in Figs. 4–6(b), the results of the standardHE method show under- or oversaturation artifacts because itcannot maintain the average brightness level.

Although RMSHE and GC-CHE methods can preserve theaverage brightness level, and better enhance overall imagequality, they lost edge details in low- and high-intensity rangesas shown in Figs. 4–6(c) and (d). On the other hand, Demirel’smethod could not sufficiently enhance the low-intensity rangeas shown in Figs. 4–6(e) because of the singular-value con-straint of the target image. Figs. 4–6(f) show the results ofthe proposed contrast enhancement method. The overall imagequality is significantly enhanced with preserving the averagebrightness level and edge details in all intensity ranges.

For performance evaluation, we used the measure of en-hancement (EME) [15], which is computed

EME =1

k1k2

k2∑l=1

k1∑k=1

Imax(k, l)

Imin(k, l) + cln

Imax(k, l)

Imin(k, l) + c(8)

where k1k2 represents the total number of blocks in an im-age, Imax(k, l) represents the maximum value of the block,Imin(k, l) represents the minimum value of the block, and crepresents a small constant to avoid dividing by zero. In thisletter, we used 8 × 8 blocks and c = 0.0001.

EME values for different enhancement methods are listed inTable I. Comparison of EME values show that the proposedmethod outperforms existing enhancement methods.


LEE et al.: CONTRAST ENHANCEMENT 5

TABLE IEME VALUES OF FIVE DIFFERENT ENHANCEMENT METHODS

V. CONCLUSION

In this letter, we have presented a novel contrast enhancementmethod for remote sensing images using dominant brightnessanalysis and adaptive intensity transformation. The proposedalgorithm decomposes the input image into four wavelet sub-bands and decomposes the LL subband into low-, middle-, andhigh-intensity layers by analyzing the log-average luminanceof the corresponding layer. The adaptive intensity transferfunctions are computed by combining the knee transfer functionand the gamma adjustment function. All the contrast-enhancedlayers are fused with an appropriate smoothing, and the pro-cessed LL band undergoes the IDWT together with unprocessedLH, HL, and HH subbands. The proposed algorithm can effec-tively enhance the overall quality and visibility of local detailsbetter than existing state-of-the-art methods including RMSHE,GC-CHE, and Demirel’s methods. Experimental results demon-strate that the proposed algorithm can enhance the low-contrastsatellite images and is suitable for various imaging devices suchas consumer camcorders, real-time 3-D reconstruction systems,and computational cameras.

REFERENCES

[1] R. Gonzalez and R. Woods, Digital Image Processing, 3rd ed.Englewood Cliffs, NJ: Prentice-Hall, 2007.

[2] Y. Kim, “Contrast enhancement using brightness preserving bi-histogramequalization,” IEEE Trans. Consum. Electron., vol. 43, no. 1, pp. 1–8,Feb. 1997.

[3] Y. Wan, Q. Chen, and B. M. Zhang, “Image enhancement based on equalarea dualistic sub-image histogram equalization method,” IEEE Trans.Consum. Electron., vol. 45, no. 1, pp. 68–75, Feb. 1999.

[4] S. Chen and A. Ramli, “Contrast enhancement using recursive mean-separate histogram equalization for scalable brightness preservation,”IEEE Trans. Consum. Electron., vol. 49, no. 4, pp. 1301–1309,Nov. 2003.

[5] T. Kim and J. Paik, “Adaptive contrast enhancement using gain-controllable clipped histogram equalization,” IEEE Trans. Consum.Electron., vol. 54, no. 4, pp. 1803–1810, Nov. 2008.

[6] H. Demirel, C. Ozcinar, and G. Anbarjafari, “Satellite image contrastenhancement using discrete wavelet transform and singular value decom-position,” IEEE Geosci. Reomte Sens. Lett., vol. 7, no. 2, pp. 3333–337,Apr. 2010.

[7] H. Demirel, G. Anbarjafari, and M. Jahromi, “Image equalization basedon singular value decomposition,” in Proc. 23rd IEEE Int. Symp. Comput.Inf. Sci., Istanbul, Turkey, Oct. 2008, pp. 1–5.

[8] E. Reinhard, M. Stark, P. Shirley, and J. Ferwerda, “Photographic tone re-production for digital images,” in Proc. SIGGRAPH Annu. Conf. Comput.Graph., Jul. 2002, pp. 249–256.

[9] L. Meylan and S. Susstrunk, “High dynamic range image rendering with aretinex-based adaptive filter,” IEEE Trans. Image Process., vol. 15, no. 9,pp. 2820–2830, Sep. 2006.

[10] S. Chen and A. Beghdadi, “Nature rendering of color image basedon retinex,” in Proc. IEEE Int. Conf. Image Process., Nov. 2009,pp. 1813–1816.

[11] Y. Monobe, H. Yamashita, T. Kurosawa, and H. Kotera, “Dynamic rangecompression preserving local image contrast for digital video camera,”IEEE Trans. Consum. Electron., vol. 51, no. 1, pp. 1–10, Feb. 2005.

[12] S. Lee, “An efficient contrast-based image enhancement in the compresseddomain using retinex theory,” IEEE Trans. Circuit Syst. Video Technol.,vol. 17, no. 2, pp. 199–213, Feb. 2007.

[13] W. Ke, C. Chen, and C. Chiu, “BiTA/SWCE: Image enhancement withbilateral tone adjustment and saliency weighted contrast enhancement,”IEEE Trans. Circuit Syst. Video Technol., vol. 21, no. 3, pp. 360–364,Mar. 2010.

[14] S. Kim, W. Kang, E. Lee, and J. Paik, “Wavelet-domain color im-age enhancement using filtered directional bases and frequency-adaptiveshrinkage,” IEEE Trans. Consum. Electron., vol. 56, no. 2, pp. 1063–1070, May 2010.

[15] S. S. Agaian, B. Silver, and K. A. Panetta, “Transform coefficienthistogram-based image enhancement algorithms using contrast entropy,”IEEE Trans. Image Process., vol. IP-16, no. 3, pp. 741–758, Mar. 2007.

[16] Computer Vision Group-University of Granada (CVG-UGR) ImageDatabase. [Online]. Available: http://decsai.ugr.es/cvg/dbimagenes

ieeegeoscience and remote sensing letters 1 contrast

Documents