paper special section on circuit, system, and...

IEICE TRANS. FUNDAMENTALS, VOL.E96–A, NO.6 JUNE 20131185

PAPER Special Section on Circuit, System, and Computer Technologies

LDR Image to HDR Image Mapping with OverexposurePreprocessing

Yongqing HUO† ,††a), Fan YANG††, Vincent BROST††, Nonmembers, and Bo GU†††, Member

SUMMARY Due to the growing popularity of High Dynamic Range(HDR) images and HDR displays, a large amount of existing Low Dy-namic Range (LDR) images are required to be converted to HDR format tobenefit HDR advantages, which give rise to some LDR to HDR algorithms.Most of these algorithms especially tackle overexposed areas during ex-panding, which is the potential to make the image quality worse than thatbefore processing and introduces artifacts. To dispel these problems, wepresent a new LDR to HDR approach, unlike the existing techniques, it fo-cuses on avoiding sophisticated treatment to overexposed areas in dynamicrange expansion step. Based on a separating principle, firstly, accordingto the familiar types of overexposure, the overexposed areas are classifiedinto two categories which are removed and corrected respectively by twokinds of techniques. Secondly, for maintaining color consistency, color re-covery is carried out to the preprocessed images. Finally, the LDR image isexpanded to HDR. Experiments show that the proposed approach performswell and produced images become more favorable and suitable for appli-cations. The image quality metric also illustrates that we can reveal moredetails without causing artifacts introduced by other algorithms.key words: HDR, overexposed area, highlight removal, principal compo-nent analysis, image quality metric

1. Introduction

In recent years, HDR images and monitors have gained sig-nificant interest in industry. Because of their superexcellentperformance, they will be the development trend in the fu-ture. Lately, researchers have focused on HDR image pro-cessing, including capturing, coding, compression, display-ing and so on [1], [2]. HDR monitors are capable of dis-playing simultaneously bright highlights and dark shadows,image or video displayed on them looks more close to whathuman eyes see in real world. Furthermore, HDR displayshave greatly extended the limited dynamic range of conven-tional cathode ray tube (CRT), liquid crystal display (LCD),and projector-based displays [3]. The developments in HDRhardware and display technique indicate that HDR displaytechnology is now getting ready for the consumer market inmost fields, from entertainment to scientific visualization.The availability of HDR displays and a large base of LDRimages necessitate the generation of HDR images from LDR

Manuscript received September 3, 2012.Manuscript revised January 19, 2013.†The author is with the School of Communication and Infor-

mation Engineering, University of Electronic Science and Tech-nology of China, Chengdu City, Sichuan Province, 611731 China.††The authors are with LE2I-CNRS 6306 Laboratory, Univer-

sity of Burgundy, 21078 Dijon, France.†††The author is with Global Information and Telecommunica-

tion Institute, Waseda University, Tokyo, 169-0051 Japan.a) E-mail: [email protected]

DOI: 10.1587/transfun.E96.A.1185

images, which not only makes it possible to re-use the ex-isting legacy contents on HDR displays, but also solves theproblem of capturing difficulty for HDR images [4]. Thisneed to expand the range of LDR image to create HDR de-piction which matches real-world luminance values as faith-fully as possible.

In general, professionals (photographers, moviemak-ers) always try to avoid excessive overexposure at imageacquiring moments. Because overexposure distracts andobscures surface detail, makes image segmentation, objectrecognition and stereo matching difficult. However, over-exposed area presence in image is hackneyed and unavoid-able for some applications. Thus, it is necessary to developa method to cope with overexposure during dynamic rangeexpansion. Most of existing LDR to HDR algorithms per-form sophisticated process to overexposed areas using theway different from that to the non-overexposed areas dur-ing dynamic range expanding. Meanwhile, most of themmake a general assumption that highly saturated pixels needto be expanded much more than the rest. As a result, brightimage areas are largely boosted. However, the sophisticatedtreatment to overexposed areas makes the dynamic range ex-pansion complicated and is the potential to make the imagequality worse than that before processing, through the intro-duction of objectionable artifacts; the large boosting to thebright image areas results in contouring artifacts for brightobject sometimes.

To dispel these problems, we propose a new LDR toHDR approach. Unlike the existing techniques, our ap-proach classifies and preprocesses the overexposure beforeexpansion instead of treating them especially in dynamicrange expanding step. It mainly consists in two novelties:the first one is the separation of overexposed area process-ing and dynamic range expansion steps; the second one isthe classification of the overexposure and the utilization ofdifferent methods for different overexposure types. The firstnovelty escapes special treatment to overexposed areas dur-ing expanding and facilitates the dynamic range expansionstep. This avoids the formation of the artifacts introducedby the dynamic range expansion step. The second noveltyensures the effect of overexposure processing. The prepro-cessing to overexposure is the foundation and key factor forthe quality of HDR image. Both novelties make contribu-tions to the high quality HDR images. In our approach,according to the types of overexposure, highlight removaltechnique or Principal Component Analysis (PCA) is usedto remove or correct the overexposure. Moreover, in order

Copyright c© 2013 The Institute of Electronics, Information and Communication Engineers

1186IEICE TRANS. FUNDAMENTALS, VOL.E96–A, NO.6 JUNE 2013

to maintain color consistency, we carry out color recoveryto the preprocessed image.

The remainder of the paper is organized as follows: wesummarize the related works in Sect. 2, followed by the pro-posed approach description in Sect. 3. Section 4 discussesthe implementation details and experiment results. Conclu-sions and perspectives are given in the last section.

2. Related Works

In the last few years, several of LDR to HDR algorithmshave been proposed (reader can see review in Banterle etal. [5]). Daly and Feng addressed the problem of dynamicrange expansion by bit-depth extension technique and de-contouring method [6], [7]. They were among the first re-searchers to concern this issue, although the specifics ofHDR displays were not taken into account. Using a globalmethod, Akyuz and colleagues performed two experiments[8] which revealed that in many circumstances a linear con-trast scaling works surprisingly well for mapping LDR con-tent onto HDR screens, based on the hypothesis that the in-put LDR image has high quality without compression ar-tifacts. Masia et al. presented a simple expansion methodbased on γ transformation [9]. It first computes γ accord-ing to the key value of image, then boosts image using γ asexponential.

The global algorithms expand all pixels using the samefunction. Considering the different particularity of pixelsin overexposed and non-overexposed areas, Banterle et al.proposed a general framework to expand LDR content withoverexposure for HDR monitors [10], [11]. They first usethe iTMO (inverse Tone Mapping Operator) for initial ex-pansion, and then compute an expand-map to reconstructlost luminance profiles in overexposed areas and attenuatequantization or compression artifacts that can be enhancedduring expansion. In almost concurrent work, Meylan et al.transformed the problem to a simple highlight detection byluminance thresholding, applying a piece-wise linear map-ping function that allocates more range to those highlights[12], [13]. Masia et al. proposed a classification expan-sion method which divides image into interesting objectsand background, not into overexposed component and non-overexposed component [14]. User chooses linear functionswith different slopes to expand the objects and background.EI-Mahdy et al. presented a LDR to HDR algorithm basedon camera sensor response curve [15]. It first uses the in-verse of response curve to perform inverse tone mapping,then performs the dodge or burn operation to selectively in-crease or decrease the luminance values of pixels.

Rempel et al. gave a video expansion technique [16],in which, the image frame is first filtered to remove arti-facts due to the compression algorithms of the media; thenthe intensity is linearized and a binary mask is created bythresholding the saturated pixels; Finally, they compute abrightness enhancement map in real time as a blurred ver-sion of the binary mask, combined with an edge stoppingfunction to retain contrast of prominent edges. The contrast

of the LDR image is scaled according to the enhancementmap. Kovaleski and Oliveira presented a similar method[17] to Rempel et al.’s, which substitutes a bilateral filter forthe combination of a Gaussian blur and an edge stoppingfunction used by Rempel. Banterle and colleagues extendedtheir original framework to process video by designing atemporally-coherent version of the expand-map [18]. Didykand colleagues advanced another expansion technique forvideo [19]. They classified a scene into three components:diffuse, reflections, and light sources, enhanced only reflec-tions and light sources. This method is suitable for highquality video enhancement thanks to the temporal coherenceof the segmentation and the expansion function. It is also in-teractive, which rely on user assistance to guide the process.

Most of the LDR to HDR algorithms described abovedeal with the overexposed areas during expanding by intri-cate methods. In contrast with these algorithms, based onthe fact that human vision is more sensitive to image contentdistortion than luminance distortion, we facilitate the dy-namic range expansion as far as possible to averting imagecontent distortion. We devote to excluding mazy measureduring dynamic range expanding by overexposed area pre-processing; eliminating the influence of overexposure be-fore expanding. This separated approach makes the expand-ing forthright and improves the image to be more pleasantand maneuverable for consumer in amusement and securitysurveillance applications.

3. LDR to HDR Approach Description

In this section, we propose a new LDR to HDR approach.Figure 1 shows our framework, which mainly contains threestages: overexposed areas classification, overexposed areaspreprocessing and dynamic range expansion.

Fig. 1 Structure of the proposed approach for LDR to HDR.

HUO et al.: LDR IMAGE TO HDR IMAGE MAPPING WITH OVEREXPOSURE PREPROCESSING1187

3.1 Overexposed Areas Classification

In practice, there are usually two types of overexposure inimage: one is induced by Specular Reflection of object sur-face which we called SR overexposure; another one is in-duced by Light Source in the image such as sun or candlewhich we called LS overexposure. These two types of over-exposure will be preprocessed respectively by two differentadequate techniques. So, we first classify the image accord-ing to the type of overexposed areas. The classification pro-cedure is performed by a method similar to Didyk et al.’smethod [19]. A Support Vector Machine (SVM) with kernelas Eq. (1) makes an initial classification.

k(z, z′) = exp(−γ||z − z′||2) (1)

where z are overexposed area features including 1st- and2nd-order image statistics (mean, median), geometric fea-tures (size, shape) and neighborhood characteristics (con-trast, smoothness of neighboring pixels). These features aredesigned based on the statistics of 1600 manually classifiedoverexposed areas. In order to correct potential errors of theclassification procedure, after the initial classification step,any misclassified images will be corrected by user. So, theclassification is semi-automatic.

The proposed approach chooses overexposed area pre-processing method in function of the classification results.The SR overexposure is performed by the highlight removaltechnique. For the LS overexposure, our experiments showthat the highlight removal technique cannot obtain satisfiedresults in certain cases. So, for this kind of overexposure,we will use PCA technique to correct it. In the followingsubsection, we will introduce these two kinds of techniques.

3.2 Overexposed Areas Preprocessing

3.2.1 SR Overexposed Area Processing

The SR overexposure is called highlight, and the highlightremoval technique [20], [21] is an essential subject in thefield of computer vision. Most of these techniques [22],[23] depend on the Dichromatic Reflection Model (DRM)introduced by Shafer [24]. They are capable of removingthe specular reflection of object surface and recovering thesurface information.

According to the DRM model, the radiance spectrumat each pixel position is the mixed spectrum as a linear com-bination of the diffuse component and the specular compo-nent:

I(x) = ID(x) + IS (x) (2)

where I(x) is the reflected light color at position x capturedby a RGB camera, ID(x) is the diffuse component and IS (x)is the specular component. The diffuse component ID(x) canbe expressed by the difference between original image I andspecular component IS(x) [22]:

Fig. 2 From left to right: original image, SR overexposed area image,and image after overexposure removal.

ID(Λmax) = I − maxi∈{r,g,b} Ii − Λmax∑

i∈{r,g,b} Ii

1 − 3Λmax(3)

where Λmax = max(Λr,Λg,Λb) is maximum diffuse chro-maticity. The diffuse chromaticity is defined as:

Λi =IDi∑

i∈{r,g,b} IDi

(4)

Since object surface materials may vary from point topoint, Λi changes from pixel to pixel. Estimating Λi forevery pixel from a single image is a non-trivial problem. So,a bilateral filter is used to approximate Λmax.

In this method, we firstly define the chromaticity oforiginal image as:

σi =Ii∑

i∈{r,g,b} Ii(5)

Then, the maximal chromaticity σmax = max(σr, σg, σb) oforiginal image I is bilateral filtered. The bigger one betweenthe original σmax and its filtered version σF

max is used as anapproximation of Λmax to compute the diffuse image. Fig-ure 2 shows the SR overexposed areas removal results.

3.2.2 LS Overexposed Area Processing

The proposed approach correct the LS overexposed area byPCA technique, which includes two stages: overexposedarea detection and overexposed area processing.

A. Overexposed area detection

In this stage, the overexposed areas in image are detectedbased on the difference between the original image and theModified Specular Free (MSF) image [25], [26]. The MSFimage is obtained by adding the mean of the minimum ofRGB color value of original image to the Specular Free (SF)image as shown in Eq. (6). The SF image is calculated bysubtracting the minimum of RGB color value in each pixellevel (see Eq. (7)).

MSFi(x, y) = SFi(x, y) + Imin (6)

SFi(x, y) = Ii(x, y) −min(Ir(x, y), Ig(x, y), Ib(x, y)) (7)

where: i ∈ {r, g, b}, Ii(x, y) is the value of ith color channelat (x, y) pixel position of original image I, Imin is the meanof Imin(x, y) for all the pixels in the image.

Then, the difference between the original image and theMSF image is used to detect overexposed areas. If the dif-ferences in three color channels of a pixel are all bigger than


Fig. 3 Original images (top) and detected LS overexposed areas(bottom).

Fig. 4 LS overexposed area processing flowchart.

the threshold, the pixel is overexposed; otherwise, the pixelis non-overexposed. The threshold value is set to Imin. Fig-ure 3 shows some LS overexposed areas detection results.

B. Overexposed area processing

After detecting the LS overexposed areas, we correct themby PCA technique. Figure 4 shows the LS overexposed areaprocessing flowchart.

Since there are three color channels, RGB image canbe represented by three principal components weighted bythree eigenvectors:

I = V1P1 + V2P2 + V3P3 (8)

where I is RGB value of image, Vi and Pi are principal com-ponent vectors and corresponding principal components re-

spectively. It has been proved by experiments that in mostcases, the second principal component corresponding to thesecond largest eigenvalue contains the part or whole of theoverexposure component [25]. Sometimes it is not true, soa threshold value should be set to detect whether the secondlargest principal component contains overexposure compo-nent. Before this, the fidelity ratio fi need to be defined as:

fi =σi∑3

j=1 σ j

× 100 (9)

where σi is the ith eigenvalue. If f2 < threshold value, thesecond principal component was detected as overexposurecomponent, it will not be used in image reconstruction. Thethreshold value is set to 2 [25].

In experiments, it was found that the first principalcomponent still contains part of overexposure. To eliminat-ing this overexposure, histogram equalization was appliedto the first principal component. The reconstructed imagecan be expressed as:

IRec = V1Ph + V2P2 + V3P3 (10)

where Ph is histogram equalized principal component andIRec is the reconstructed image. The second item of the rightside is included or removed in reconstruction according tofidelity ratio of second largest eigenvector.

Through plenty of examinations, we found that thecolor of the reconstructed image IRec is shifted. The orig-inal color of the image without overexposure is obtained bysecond order polynomial transformation. The main problemis to find the weight function for polynomial transformation.This weight function is calculated using transformation be-tween detected non-overexposed part of original image andcorresponding part of reconstructed image as:

Id = WMd (11)

The weight function W can be obtained by using leastsquare pseudo inverse matrix calculation as:

W = Id

[MT

d Md

]−1MT

d (12)

where Id is non-overexposed part in original image, Md

is the second order polynomial extension of correspondingnon-overexposed part of reconstructed image. [ ]−1 and [ ]T

are the inverse and the transpose of matrix respectively. Weapply the weight function to the second order polynomialextension M of reconstructed image IRec to achieve the over-exposure corrected image IDiff of the original image I:

IDiff = WM (13)

Although the overexposed areas have been processedtill now, we noted that in some cases the color in non-overexposed areas of image IDiff is not exactly the same asthe original image I, and sometimes the difference is visible.Furthermore, there are noise and artifacts in image IDiff . So,we combine the luminance of the original image and IDiff toobtain the final overexposure corrected image ID:


ID(x, y)=

{I(x, y) × L

LDiff(x, y) � over exposed area

IDiff (x, y) (x, y) ∈ over exposed area(14)

where, L and LDiff are the luminance information of originalimage and image IDiff respectively.

Figure 5 shows the original image I, images IDiff andID. From the figure, we can notice that the color in non-overexposed areas of image ID is the same as that of originalimage I; but the color of image IDiff has apparent differenceto original image I. From the images in right column, wealso can find that image IDiff without combination has noiseand artifacts.

3.3 Dynamic Range Expansion

After overexposed area preprocessing, the image is moresuitable for expanding and more close to the normal ex-posure image. Akyuz et al. [8] got a result by two exper-

Fig. 5 From top to bottom: original image I, final overexposurecorrected image without and with combination step.

iments: LDR image does not necessarily require sophisti-cated treatment to produce a compelling HDR experience;simply boosting the range of an LDR image linearly to fitthe HDR monitor can equal or even surpass the appearanceof a true HDR image. However, this result is obtained basedon the assumption that LDR image has high quality. If theimage contains overexposed area, the result is not alwaystrue.

Since overexposed areas in images are removed andcorrected using our approach, according to the results ob-tained by Akyuz et al., we use simple linear expansionmethod to expand the image. This will bring forth two ad-vantages: the first one is that the expanding method is globaland linear, which makes the LDR to HDR algorithm allevi-ated; the second one is that the straightforwardness of theexpanding method avoids causing artifacts in incorrectly ex-posed areas. The linear expansion calculation is given inEq. (15):

Lh = Ih × Ll − Lmin

Lmax − Lmin(15)

where Ll is the luminance of the pixel being scaled, Lmin andLmax are the minimum and maximum luminance of the inputimage. Ih is the maximum input intensity of the HDR mon-itor, Lh is the luminance of the scaled pixels. This operationwas applied to all pixels individually.

4. Experiments and Results

We implemented the proposed algorithm with Matlab2011bon Intel Core i5-2520M CPU @ 2.5 GHz, 4.00 GB RAM,win32 system. The maximum illumination of HDR moni-tor is set to 3000 cd/m2, according to Brightside’s 37” HDRmonitor. We used a set of images to test our approach;they represent a wide range of lighting conditions from verydark to very bright, including overexposed and underex-posed scenes. Figure 6 shows a subset of the tested images.

In order to validate the proposed scheme, we operatedon the test images and compared to other three algorithms:Banterle et al.’s operator [10], Meylan et al.’s function [12]and LDR2HDR [16] (all the three algorithms perform so-phisticated treatment to overexposed areas during expand-ing). The comparison is performed mainly by tone mapped

Fig. 6 Some original test images, from left top to right down: Pear, Sculpture, Ball, Roberto, Tahoe,Fish, Person, Church, Rosette and Sea.


Fig. 7 Experiment result illustration; for each image group, from the top left in clockwise order: atone-mapped version of Banterle et al.’s HDR output [10]; a tone-mapped version of Meylan et al.’s HDRoutput [12]; a tone-mapped version of Rempel et al.’s HDR output [16]; and a tone-mapped version ofour HDR output (all tone-mapped version produced using Reinhard et al.’s TMO [27]).


Fig. 8 From left to right in each row: comparison between the original LDR image (reference image)and HDR image generated by Banterle et al.’s operator [10], Meylan et al.’s function [12], LDR2HDR[16], and our proposed approach.

version of the generated HDR images and the image qualitymetric results.

4.1 Tone Mapped Results

Because of limitations of the print medium, we cannotpresent the generated HDR images here directly, which aresupplied as additional materials. We mapped them to LDRwith Reinhard et al.’s photographic tone mapping operator[27] and showed a part of these LDR images in Fig. 7. Fromthe tone mapped versions, we observe that the proposedapproach works well without inducing distortion or losingfidelity. Moreover, our method can reveal details whichare invisible in the original LDR image, as (a) (c) (d) (e)in Fig. 7. Furthermore, the proposed algorithm can totallyeliminate the SR overexposure, as (a) (b) (c) (e) in Fig. 7; itcan make the glaring LS overexposure more soft and com-fortable, especially in Figs. 7(f) and 7(g), the tone mappedversion of our HDR image is the most pleasing one of thefour images. In brief, the results show that the proposed ap-proach can make the images containing overexposed areasmore popular and proper for entertainment and surveillanceapplications.

4.2 Image Quality Metric Results

In order to further validate the performance of the proposedmethod, we choose a novel image quality metric introducedby Aydin et al. [28] to assess the quality of generated HDRimages (the original LDR images as reference images). Thismetric identifies distortions between two images, indepen-dently of their respective dynamic ranges. The metric usesa model of the human visual system, and classifies visiblechanges between a reference and a compared image. Theresult image generated by the metric is a summary imagewith red, green and blue pixels. The authors identify that

red pixels mean reversal of visible contrast (when contrastpolarity is reversed in the compared image with respect tothe reference image); green pixels imply loss of visible con-trast (when visible contrast in the reference image becomesinvisible in the compared image); blue pixels denote ampli-fication of invisible contrast (when invisible contrast in thereference image becomes visible in the compared image).In our test, the original LDR images are reference images,the generated HDR images are compared images. The met-ric parameters are set to the default values: Typical LCD,viewing distance of 0.5 m, pixel per visual degree 30, peakcontrast 0.0025.

We compared the original LDR images (reference im-age) with HDR images generated by four algorithms consid-ered in the comparison. A part of result images are showedin Fig. 8. These images show that the metric result of our ap-proach has more blue pixels and little red and green pixels.According to the authors of the metric, the larger blue valuemeans that the LDR to HDR method can reveal more imagedetails; the larger red and green value imply that the methodintroduced more contrast reversal and contrast loss. Thus,we can declare that our approach reveals more details andinduces little contrast loss and reversal. This will be provedby the numerical results in Table 1. In addition, from the lastimage of each row, we can note that between the HDR im-age generated by our method and the original LDR image,there is no loss, amplification or reversal of contrast in theoverexposed areas.

Table 1 gives the red, green and blue pixel percentages(the ratio of the red/green/blue pixel number to the total im-age pixel number) of result images obtained by Aydin et al.’sHDR metric. The data in Table 1 also shows that our methodobtains more blue pixels and little red, green pixels. Thesedata similarly imply that the proposed approach can arousemore image visibility and induce little contrast loss and re-versal in average.


Table 1 Red, green, blue pixel percentage of HDR metric result images: Larger blue percentage indi-cates more detail reveal; larger red percentage means more contrast reversal; and larger green percentagedenotes more contrast loss.

Fig. 9 Comparison between the original LDR image and (a) LS overex-posure processed image, and (b) SR overexposure processed image.

We also tested and validated our overexposed area pro-cessing methods with Aydin et al.’s image quality metric.We compared the original LDR image (reference image)with LDR image after preprocessed. Referring to the orig-inal images in Fig. 6, the results in Fig. 9 indicate that theoverexposed area preprocessing techniques indeed and justprocess the overexposed areas. Figure 9(b) infers that thehighlight removal technique can eliminate the specular re-flection and recover the surface information (the informationof overexposed area is visible after processing, it is coloredblue). Both techniques induce slight detail loss and con-trast reversal; however, these are negligible (2.17% greenpercentage and 3.62% red percentage per pixel for Church;0.39% green percentage and 0.43% red percentage per pixelfor Pear).

These good performances owe to the separationmethod and the appropriate techniques for different overex-posure. The separation method eliminates the special treat-ment to the overexposed areas, which facilitates the dynamicrange expanding and avoids image distortion. The propertechniques for overexposures guarantee the effect of the im-age preprocessing, which favors the high quality of the gen-erated HDR images.

5. Conclusions and Perspectives

In this paper, we have presented a new LDR to HDR ap-

proach for dynamic range expansion of legacy, low dynamicrange images. Unlike some existing algorithms, we sepa-rate the overexposed area treatment from the dynamic rangeexpanding step. This separated scheme dispels the poten-tial to make the image appear worse than before processingand prevents the objectionable artifacts in significant areaof image. Moreover, our approach process the overexposedarea by different techniques according to their types, whichmakes it sophisticated and proper for the most familiar kindsof overexposure in practice.

The experiments demonstrate that the proposed ap-proach can make result images more favorable and natural,facilitating the applications such as image segmentation, ob-ject recognition, film, television, video surveillance etc. Theimage quality metric indicates that our approach works wellin comparison to the algorithms using mazy treatment tooverexposed areas in expansion. Furthermore, it should benoted that despite we describe this study focusing on imagewith overexposure, our approach also gets pleasing resultsfor underexposed areas in images (see Figs. 7(a), 7(d) and7(e)).

The results obtained by the metric also conclude thatthe overexposed area preprocessing methods perform well;it allows initiating more details in the most significant areasof images and leading little negligible detail loss and con-trast reversal.

In the actual framework, the overexposed area classi-fication is not automatic; we would like to study some au-tomatic classification method. Moreover, we will exploreother high efficiency overexposure removal techniques toimprove performance of the approach. In the future, wealso want to optimize our algorithms and implement themon hardware embedded system.

References

[1] J.H. Wang, D. Xu, and B. Li, “Exposure fusion based on steerablepyramid for displaying high dynamic range scenes,” Optical Engi-neering, vol.48, no.11, 117003, Nov. 2009.

[2] T. Jinno, H. Kaida, X. Xue, A. Nicola, and M. Okuda, “μ-Law basedHDR coding and its error analysis,” IEICE Trans. Fundamentals,vol.E94-A, no.3, pp.972–978, March 2011.

[3] D. Xu, C. Doutre, and P. Nasiopoulos, “Correction of clipped pixelsin color images,” IEEE Trans. Vis. Comput. Graphics, vol.17, no.3,


pp.333–344, March 2011.[4] S. Battiato, A. Castorina, and M. Mancuso, “High dynamic range

imaging for digital still camera: an overview,” J. Electronic Imaging,vol.12, no.3, pp.459–469, July 2003.

[5] F. Banterle, K. Debattista, A. Artusi, S. Pattanaik, K. Myszkowski,P. Ledda, M. Bloj, and A. Chalmers, “High dynamic range imag-ing and low dynamic range expansion for generating HDR content,”Computer Graphics Forum, vol.28, no.8, pp.2343–2367, 2009.

[6] S. Daly and X. Feng, “Bit-depth extension using spatiotemporal mi-cro dither based on models of the equivalent input noise of the visualsystem,” Proc. Color Imaging VIII: Processing, Hardcopy, and Ap-plications, vol.5008, 455, SPIE, June 2003.

[7] S. Daly and X. Feng, “Decontouring: Prevention and removal offalse contour artifacts,” Proc. Human Vision and Electronic ImagingIX, vol.5008, 455, SPIE, June 2004.

[8] A.O. Akyuz, R. Fleming, B.E. Riecke, E. Reinhard, and H.H.Bulthoff, “Do HDR displays support LDR content?: A psychophys-ical evaluation,” ACM Trans. Graphics, vol.26, no.3, pp.1–7, 2007.

[9] B. Masia, S. Agustin, R. Fleming, O. Sorkine, and D. Gutierrez,“Evaluation of reverse tone mapping through varying exposure con-ditions,” ACM Trans. Graphics, vol.28, no.5, pp.1–8, 2009.

[10] F. Banterle, P. Ledda, K. Debattista, and A. Chalmers, “Inverse tonemapping,” Proc. 4th International Conf. on Computer Graphics andInteractive Techniques in Australasia and Southeast Asia, pp.349–356, New York, 2006.

[11] F. Banterle, P. Ledda, K. Debattista, A. Chalmers, and M. Bloj, “Aframework for inverse tone mapping,” Visual Comput., vol.23, no.7,pp.467–478, 2007.

[12] L. Meylan, S. Daly, and S. Susstrunk, “The reproduction of specu-lar highlights on high dynamic range displays,” Proc. IST/SID 14thColor Imaging Conference, Springfield: Society for Imaging Sci-ence and Technology (IS&T), pp.333–338, 2006.

[13] L. Meylan, S. Daly, and S. Susstrunk, “Tone mapping for high dy-namic range displays,” Proc. Human Vision and Electronic ImagingXII, Bellingham, SPIE, pp.649210-1–649210-12, 2007.

[14] B. Masia, R. Fleming, O. Sorkine, and D. Gutierrez, “Selective re-verse tone mapping,” Proc. CEIG, 2010.

[15] A. El-Mahdy and H. El-Shishiny, “High-quality HDR renderingtechnologies for emerging applications,” IBM J. Research and De-velopment, vol.54, no.6, pp.8-1–8-15, 2010.

[16] A.G. Rempel, M. Trentacoste, H. Seetzen, H.D. Young, W. Heidrich,L. Whitehead, and G. Ward, “Ldr2Hdr: On-the-fly reverse tonemapping of legacy video and photographs,” ACM Trans. Graphics,vol.26, no.3, Article 39, 2007.

[17] R.P. Kovaleski and M.M. Oliveira, “High quality brightness en-hancement functions for real-time reverse tone mapping,” VisualComput., vol.25, pp.539–547, April 2009.

[18] F. Banterle, P. Ledda, K. Debattista, and A. Chalmers, “Expand-ing low dynamic range videos for high dynamic range applications,”SCCG’08: Proc. 4th Spring Conf. on Computer Graphics, pp.349–356, New York, ACM, 2008.

[19] P. Didyk, R. Mantiuk, M. Hein, and H.P. Seidel, “Enhancement ofbright video features for hdr displays,” Eurographics Symposium onRendering, vol.27, no.4, pp.1265–1274, 2008.

[20] D. Xu, C. Doutre, and P. Nasiopoulos, “Saturated-pixel enhance-ment for color images,” Proc. 2010 IEEE International Symposiumon Circuits and Systems (ISCAS), pp.3377–3380, Aug. 2010.

[21] E. Elboher and M. Werman, “Recovering color and details of clippedimage regions,” Int. J. Computer Information Systems and IndustrialManagement Applications, vol.3, pp.812–819, 2011.

[22] Q. Yang, S. Wang, and N. Ahuja, “Real-time specular highlightremoval using bilateral filtering,” Proc. ECCV (2010), pp.87–100,2010.

[23] R. Tan and K. Ikeuchi, “Separating reflection components of tex-tured surfaces using a single image,” IEEE Trans. Pattern Anal.Mach. Intell., vol.27, no.2, pp.178–193, 2005.

[24] S.A. Shafer, “Using color to separate reflection components,” Color

Res. App., vol.10, no.4, pp.210–218, 1985.[25] P. Koirala, M. Hauta-Kasari, and J. Parkkinen, “Highlight removal

from single image,” Proc. Advanced Concepts for Intelligent VisionSystems (Springer, 2009), vol.5807, pp.176–187, 2009.

[26] H.L. Shen, H.G. Zhang, S.J. Shao, and J.H. Xin, “Chromaticity-based separation of reflection components in a single image,” PatternRecognit., vol.41, pp.2461–2469, 2008.

[27] E. Reinhard, M. Stark, P. Shirley, and J. Ferwerda, “Photographictone reproduction for digital images,” ACM Trans. Graphics, vol.21,no.3, pp.267–276, 2002.

[28] T.O. Aydin, R. Mantiuk, K. Myszkowski, and H.P. Seidel, “Dy-namic range independent image quality assessment,” Proc. ACMSIGGRAPH Papers, pp.1–10, Los Angeles, CA, 2008.

Yongqing Huo received her B.S. degree incommunication enginnering in 2002, M.S. de-gree and Ph.D. degree in signal and informationprocessing in 2005 and 2007 from the Univer-sity of Electronic Science and Technology ofChina. She is currently an assistant professorwith the college of communication and infor-mation engineering at University of ElectronicScience Technology of China. And now she isproceeding her postdoctoral research in Univer-sity of Burgundy. Her current research interests

include image analysis and HDR image processing.

Fan Yang received the B.S. degree inelectrical engineering from the University ofLanzhou, China, in 1982 and the M.S. (D.E.A.)(computer science) and Ph.D. degrees (imageprocessing) from the University of Burgundy,France, in 1994 and 1998, respectively. She iscurrently a full professor and member of LE2ICNRS-UMR, Laboratory of Electronic, Com-puting, and Imaging Sciences at the Universityof Burgundy, France. Her research interests arein the areas of patterns recognition, neural net-

work, motion estimation based on spatio-temporal Gabor filters, paral-lelism and real-time implementation, and, more specifically, automatic faceimage processing algorithms and architectures. Pr. Yang is member of thefrench research group ISIS (Information, Signal, Images and Vision), shelivens up she livens up the theme C: Algorithm Architecture Adequation.

Vincent Brost received the Master Degreeand Ph.D. degrees from University of Burgundy.He is currently a associate professor at the Uni-versity of Burgundy and member of LE2I (Lab-oratory of Electronic, Computing, and Imag-ing), France. His research interests are in theareas of rapid prototyping system, VLIW archi-tecture, and Electronic Devices.


Bo Gu received his B.E. degree in com-puter science from Tianjin University, Tianjin,China, in 2004, and M.E. degree from PekingUniversity, Beijing, China, in 2007, respec-tively. From 2007 to 2011, he was a researchengineer at Sony Digital Network Applications,Inc., Japan. He is currently a research asso-ciate at Global Information and Telecommuni-cation Studies, Waseda University. He receivedthe IEICE Young Researcher’s Award in 2011,and the IEICE Communication Quality Confer-

ence Premium Award in 2012.

paper special section on circuit, system, and...

Documents