Local Contrast Enhancement

Transcription

1 Local Contrast Enhancement Marco Bressan, Christopher R. Dance, Hervé Poirier and Damián Arregui Xerox Research Centre Europe, 6 chemin de Maupertuis, Meylan, France ABSTRACT We introduce a novel algorithm for local contrast enhancement. The algorithm exploits a background image which is estimated with an edge-preserving filter. The background image controls a gain which enhances important details hidden in underexposed regions of the input image. Our designs for the gain, edge-preserving filter and chrominance recovery avoid artifacts and ensure the superior image quality of our results, as extensively validated by user evaluations. Unlike previous local contrast methods, ours is fully automatic in the sense that it can be directly applied to any input image with no parameter adjustment. This is because we exploit a trainable decision mechanism which classifies images as benefiting from enhancement or otherwise. Finally, a novel windowed TRC mechanism based on monotonic regression ensures that the algorithm takes only 0.3 s to process a 10 MPix image on a 3 GHz Pentium. Keywords: image enhancement, contrast adjustment, tone reproduction operator, local contrast enhancement, exposure correction, TRC, user evaluation, decision mechanism. 1. INTRODUCTION Almost half a trillion digital camera and camera phone images will be captured in On the other hand, it is already easy for users to integrate their own content into value-added workflows such as online photofinishing and content-sharing communities. Regardless of the final medium where the images will be managed, shared and visualized the quality expectations of consumers are likely to grow steadily. Acquisition conditions, user expertise, compression algorithms and sensor quality, can seriously degrade image quality. Image enhancement compensates for these degradations, improving subsequent analysis, distribution and display. While perceptual measures for the quality of contrast, sharpness, noise, exposure, colour can be defined, the final judgement remains subjective. Image processing software often requires considerable expertise to be employed efficiently. Consumers are often not willing to make the effort to manually enhance their pictures, though they would appreciate the results. The result is a demand for fully automated image enhancement tools that require no user interaction. In this paper we focus on automatically correcting one of the most common degradations, namely the joint existence of over- and underexposed regions. This problem arises in settings with back lighting, indoor lighting coupled with bright windows and flash. Our local contrast enhancement (LCE) algorithm enhances such images by locally correcting exposure, preserving contrast. With respect to existing work in this domain, LCE has three clear strengths: it is fully automatic, it is thoroughly validated through pairwise preference tests and it is fast. The optimized version of the algorithm processes a 10 MPix image in less than 0.3 seconds on a 3.2Ghz Pentium machine. Developing an algorithm free of user interaction is not only about automatically setting the parameters to work across a wide range of images but also preventing the application of the enhancement in cases where the image may not benefit. Since we prefer inaction to degradation a decision mechanism becomes necessary as a preprocessing stage of the algorithm. On an image set representative of consumer photography, the default settings of our system result in 8% degraded images and 46% significantly improved images. Our decision mechanism enables approximately 40% of the images to be processed with LCE with less than 1% degraded. Section 2 presents previous work on contrast correction. Section 3 describes the inner workings of LCE, while section 4 describes a decision mechanism which we have incorporated after a first series of user tests. Section 5 reports on the results of the evaluation of the system, and section 6 concludes. Authors addresses: Firstname.Lastname@xrce.xerox.com

2 2. PREVIOUS WORK Contrast enhancement is typically achieved via tone mapping techniques applied to luminance values. In general, tone mapping approaches are classified in two categories depending on their global (spatially uniform) or local (spatially varying) nature. The former are generally referred to as Tone Reproduction Curves (TRCs) and the latter Tone Reproduction Operators (TROs): see Devlin, 1 DiCarlo and Wandell 2 and Fattal et al. 3 A complete evaluation of TRO performance focused on high dynamic range display appears in Ledda et al. 4 TRCs are simple and computationally efficient but limited by their global nature. 5, 6 They generally result in loss of local contrast for some image regions giving a washed out appearance. Global histogram adjustment by Ward et al. 7 is an exception which preserves local contrast while still being based on a global approach. Limitations due to the global nature of this technique are observed in those images where luminance is uniformly distributed over the whole range. The simplest TROs, such as CLAHE, 8 apply TRCs locally. CLAHE is not focused on subjective image quality but rather on bringing out all detail present in an image, even at the cost of enhancing noise. For this reason it is more frequently used in medical imaging or remote sensing than in photography. Generative models that address the problem of recovering reflectance from an image are also a very common approach. In this model, an image is considered as the product I = R.L where R is the reflectance and L the illuminance at any given pixel. These approaches are also called background-image techniques. Since this problem is ill-posed additional assumptions are introduced on reflectance and illuminance to bound the solution. The earliest approach along this line is homomorphic filtering which assumes strong changes R and smooth changes in L. 9 Under these assumptions, both factors can be separated in log-space via band-pass filters. These assumptions do not hold in the presence of strong edges, which leads to halo and banding artifacts. More complex TROs, such as multiscale versions of the retinex theory of color vision 10 have been proposed 11, 12 for recovering reflectance but in the gradient domain. When compared with homomorphic approaches, the halos produced by these techniques are reduced but not completely eliminated. Multiscale techniques can also be applied in the context of partial differential equations. Tumblin and Turk 13 perform a hierarchical decomposition of an image into levels with different levels of detail using anisotropic diffusion. Anisotropic diffusion creates artificial flat regions. This technique is computationally intensive and, even though it avoids halo artifacts, it can produce an artificial sensation since it overemphasizes detail. Artefacts introduced by anisotropic diffusion can be avoided if the bilateral filter 14 is used within the same hierarchical scheme. 15 The notion here is that if the gain map has sharp edges in the same locations that the image, halos are prevented. Nevertheless, this approach is too slow for modern digital photographic production printing speeds of megapixel images per minute and is has nothing intrinsic to prevent it from being applied to images which have already been enhanced in this way. 3. LCE ALGORITHM Figure 1 illustrates the main steps of our algorithm. First we downsample the luminance in order to estimate a background image via an edge-preserving filter. This background image is used to estimate a local gain. The gain is used to estimate the local TRCs in full resolution space using monotonic regression. The output luminance is computed by bilinear interpolation of these TRCs. Finally, chrominance is recovered to obtain the enhanced output image. These steps seek to satisfy two of the main objectives of the algorithm: quality and speed. The objective of safety which states that inaction is preferred to potential degradation is addressed by the decision mechanism detailed in the following section. We now motivate the choices in the algorithm in more detail. The first step transforms the input image I in into a predefined luminance-chrominance space such as YIQ, CIELab, or HSV. The luminance image Y in is downsampled by an integer factor to a resolution close to some target without prior smoothing yielding image Ŷ in. The target downsampling rate R = integer(max(w, H)/512) has proved effective throughout the experiments. The main purpose of working at a lower resolution is to accelerate the costly edge-preserving filter and gain computations.

3 Figure 1. LCE Algorithm. The background layer is estimated by applying a bilateral filter 14 to Ŷ in obtaining Ŷ blf. An effective, fast implementation of the bilateral filter has been suggested by Durand. 15 The background layer Ŷ blf is then used to estimate the gain Ĝ ( Ĝ = min M G, max ( L ) ) m G,ρ+ (1) max(3, Ŷ blf ) where M G and m G are the maximum and minimum tolerated gain, L represents a limit on low-exposure luminance levels, and ρ is an offset value. Offset value ρ is defined as ρ =1 L/C where C is a cut-off value which prevents from applying gain to bright regions. Figure 2 shows our choice of gain. The proposed gain was defined to deal with problems in over-exposed and under-exposed regions. Sensor noise is generally found on dark image regions and, in this situation, applying an excessive gain enhances noise. We avoid this situation by limiting the maximum gain value and introducing L. In bright image regions a common approach is to reduce luminance differences as can be seen in the blue curve in figure 2. This might cause the flattening of shiny patches on specular surfaces such as metal or parts of faces given an unrealistic appearance. In the presence of flat, bright background the result is a grayed out effect. If bright regions are, instead, stretched as would occur if a hyperbolic tangent curve-type approach is chosen for the gain there is a risk of introducing posterization artifacts in flat regions. Since, in many situations, bright regions correspond to sky this risk is high. These undesirable results are avoided in our scheme by introducing a cut-off value above which the luminance levels are left untouched. Gain parameters L = 40, C = 160, M G =3andm G = 1 were chosen. These same values have shown to be robust throughout a large variety of images and have been consistently used throughout the experiments. LCE avoids contrast enhancement in bright regions. This is related to our objective of safety. It was observed throughout user tests that people appreciate the enhancement of dark regions more than processing bright regions. Additionally, due to the mentioned problems, the risk of degrading an image by enhancing a bright region is higher and more difficult to avoid than with a dark region. Having defined the gain Ĝ, a first approach could be to upsample it and apply the result directly to Y in. Results would have been similar to using a low-pass filtered version of the image to estimate the gain, rendering useless the effort of using an edge-preserving filter. The typical problems with low-pass filters such as haloes and banding effects would arise in this case. Instead, LCE uses a fast implementation simultaneously avoiding unwanted artifacts. We first compute the enhanced low-resolution image using the computed (low-resolution) gain Ŷ out = Ŷ in.ĝ. We wish to apply this gain to the full resolution image so we first divide this image into windows. In our experiments the number of subwindows was fixed to 64 (8 8 grid) regardless of the image size. We then regress the TRC on each subwindow. In particular, we fit the minimum squared error monotonic curve to the samples (Ŷ in, Ŷ in ). This

4 Figure 2. In red, choice of gain for LCE algorithm together with role of parameters. Parameters were fixed after for all images based on user preference. is easy to achieve by dynamic programming. Figure 3 illustrates this estimation for a sample set of input and output luminance values. Since we only constrain monotonicity and not strict monotonicity, there is a risk of TRCs which have constant segments, for instance that for input luminance in the figure. Such segments are detected and merged with neighbouring regions into strictly increasing portions. Having estimated the TRCs for each window of the image, we apply them to the full-resolution image Y in to finally obtain Y out. Blocking artifacts are avoided by obtaining the value for each pixel from the result of interpolating the TRCs for the neighboring windows to the pixel. One possible approach is illustrated in Figure 4. Taking this approach, the output luminance is given by, Y out (x, y) = 9 n=1 TRC n( Yin (x, y) ) exp ( dn/σ) 9 n=1 exp( dn/σ) (2) The final stage is to perform chrominance restoration to recover the enhanced color image I out. A common requirement on this stage is to use the modified luminance while preserving hue and saturation. 16 Different approaches arise from different definitions for these color dimensions. If HSV color space is used, hue and saturation are preserved under a multiplicative model: linearly scaling the RGB coordinates. If, instead, YIQ or any other linear transform of RGB is used, hue and saturation are preserved under an additive model: linearly shifting the RGB coordinates. An increase in luminance can be seen, depending on the model, as a positive shift or a scaling by a value larger than one on RGB space. It is interesting to notice that each of this two transformations preserve the hue but alters the saturation in the other space. An increase of luminance modelled by scaling the RGB coordinates will result in a saturation increase in YIQ space. On the other hand, a positive shift in RGB space will result in a saturation decrease in HSV space. An unexpensive approach for overcoming

5 Figure 3. On the top row, the result of applying a traditional gain approach to the luminance levels. On the bottom row, the same images after our proposed gain has been applied. Figure 4. The Tone Reproduction Curve applied to any given image pixel is the result of linearly combining the TRCs of neighboring windows, using distance as weight.

6 this situation is to simply use a combination of the two models. If I c in,c=1, 2, 3 are the three RGB coordinates for the input image, R = Y out /Y in the luminance ratio, and D = Y out Y in the luminance difference, then the corresponding output coordinates are obtained by, I c out =(α)(ri c in)+(1 α)(d + I c in) (3) where α is a saturation parameter that if α = 0 YIQ saturation is preserved and the image usually looks less saturated, and if α = 1 HSV saturation is preserved and the image might appear overly saturated. Setting α to a fixed value gave similar results to more complex schemes based on gain and chrominance and nonlinear color spaces such as CIE L*a*b* and CIE L*u*v*, or gamma-based corrections. Throughout the experiments the value α =0.6 has been used with good performance across a wide range of images. 4. DECISION MECHANISM The LCE decision mechanism (DM) efficiently determines whether the enhancement will be applied to an image, ensuring safety of the approach. In our case image improvement or degradation is only related to user preference. Given image I, we can estimate p I, the probability of a subject preferring the enhanced version to the original image. The optimal DM would be one which is able to accurately threshold this probability. The details of this estimation can be found in Section 5 and we for this section we assume we have a sufficiently large image training set labelled with probabilities of user preference. Observation of labelled data confirmed some very reasonable features which could influence user preference. Firstly, people do not want noise to be enhanced. The most common sources of noise are sensor noise and compression artifacts. Notice that our algorithm will only make noise explicit if it is present in the darker image regions. Sensor noise usually concentrates on these regions. It is also well known that stretching luminance values in flat image regions can result in posterization artifacts and can have a graying effect on almost black regions. Test participants noticed these effects and disliked the results. Instead, those images where users valued most the algorithm were those where the underexposed image regions showed a high level of interesting detail. Since LCE practically does not alter medium and high exposed image regions, it has no perceivable effect on those images with normal exposure levels. These images came out in the test as samples where the user does not care whether the algorithm is applied or not. Though LCE does not degrade these images, we also included a feature to measure brightness (perceived luminance) change in the DM so that we do not apply LCE if we can anticipate it will have no perceivable result. Except for the brightness feature, all other features are local in nature. We made a first set of candidate features and refined this set by analyzing their predictive qualities using correlation measures on normalized feature values. The five features finally used by the decision mechanism are: Brightness change measures the variation in brightness from the low-resolution input image to an estimated output image. Computed on luminance histograms. Brightness change is defined by B = L L (1/3)( H in (L) H out (L) ). Detail gain measures the spread of the difference between the low-resolution input image and its low-pass filtered version. The detail gain/loss is the difference between this measure on the enhanced image and on the input image. Detail gain is defined by D = std(y out Ŷout) std(y in Ŷin). Noise (sensor) is the expected value after applying a Laplacian filter (LAP ) to a chrominance channel (Q from YIQ). High values on the luminance gradient image are not considered since they could yield a high response on the Laplacian filter, confusing edges with noise. Denoting the gradient of the luminance by G and the empirical expectation by E treating pixels as independent samples, this is given by N = E[ Q LAP G(x, y) <k].

7 Compression Artifacts are measured by the presence of JPEG blocking artifacts and can be computed from the distribution of chrominance along a regular grid of the image. For implementation details see 17 Flatness is measured as the negative entropy of the input image and is defined as, F = L H in (L) log(h in (L)). Except for the global measure on compression artifacts, all these features can be computed in an image ROI. In our implementation, we divide the image in an equal 8 8 grid, such that 64 subwindow feature values are obtained per image. For each feature we count the number of times the features is above (below) a certain threshold and classify these counts. This ensures intuitive results from the decision mechanism, e.g. don t enhance if there are more noisy windows. The optimal threshold values were obtained by assuming independence and cross-validating on a measure of class separability, e.g. divergence. For the dimensionality reduction we also evaluated PCA trained on a 5000 image non-labelled data-set. Class separability results were comparable to our threshold approach, but the straightforward interpretation is lost. Figure 5 shows the different features used by the decision mechanism on a sample image. Figure 5. Features used for the Decision Mechanism on a sample image. Last cell shows actual feature values in block counts and DM decision for this particular image.

8 We can pose the design of the decision mechanism as a regression problem or as a classification problem. In our experience, regression has proved a more complex problem to tackle than classification. Some reasons that can account for this difficulty are insufficient training data, noise generally present in subjective image evaluation, lack of agreed upon rating standards, inadequate choice of features, etc. Based on these results we decided to consider the DM as a binary image classifier which indicates if LCE should be applied to image I (DM(I) =1) or not (DM(I) = 0). The choice of a classifier has also been suggested in other works dealing with (subjective) measures of image quality. 18 A training set for this classifier was generated from the user evaluated images and the label 1 (enhance) was assigned if the probability of preference was clear p I > 0.5+δ and 0 if p I < 0.5 δ. A low value for delta would allow noisy data (strong user disagreement) in our training set, while a high value can excessively reduce the amount of training data. We have observed values between 0.15 <δ<0.35 yield very similar results and fixed δ = 0.2 throughout the experiments and in our current implementation. A total of 128 images satisfied this condition, 73 negative samples (DM should reject) and 55 positive samples. Best results were obtained using the naive Bayes classifier with the smoothed feature histograms as nonparametric probability estimates. Notice that these we had to estimate the probability of 4 discrete random variables taking values between 0 and 64 (number of blocks). The JPEG value was discretized in a uniform distribution using JPEG values obtained from a 5000 image dataset independent of our training data, 10 bins were used. Our application required that we provide three different configuration settings for the decision mechanism: risky, default and conservative. The risky option would allow a large number of images to pass and would be activated when there is a certainty on the image quality. The conservative option would reject most of the images. A false positive (FP) occurs when the decision mechanism decides to accept an image that the enhancer degrades. A false negative (FN) occurs when a rejected image would have improved with the enhancement. Finally, we have a global measure of rejection (GR) which indicates how many images the DM will reject, regardless the effect LCE can have on them. Notice that different weights on FP and FN, as well as different criteria for GR can result in completely different policies for setting the thresholds and priors on the naive Bayes classifier. Typically, the cost of a false positive will be much higher than the cost of a false negative. Our approach takes this assymetry into account and its performance is summarized in Table 1. Table 1. DM performance for the three settings required by our implementation. FP FN GR conservative 2% 69% 76% default 8% 54% 59% risky 25% 36% 35% Notice that in Table 1 FP is not a global measure of image degradation but the percentage of negatively labelled images which the DM incorrectly accepts. The global probability of degrading an arbitrary image is much lower. In our large scale general preference test (test 3 in Table 2) the percentage of degraded images observed by participants was approximately 7%. If the images in this test were representative of an arbitrary dataset the global degradation rate of our method would be ( ) < 1%. Another estimation of the global degradation rate can be obtained from the process in designing tests 5 and 6 in the same table. The images for these tests were obtained from a 3500 image dataset including images from a large variety of sources. An initial screening of the 3500 was made and roughly 200 images where there was a possible negative difference were selected. Approximately 40% of these images were actually considered as degraded by the users. Even if our screening missed half of the relevant images, and assuming the percentage holds, these numbers indicate that the probability of global degradation is ( /3500) 0.08 = 0.4%. Actual degradation will strongly depend on image set characteristics such as sensor quality and compression rate, so, if a global degradation number has to be estimated, this should be done from user preference evaluations on specific scenarios.

9 5. EVALUATION We designed LCE to be automatic, fast and user validated. We have seen that our implementation, both the method and decision mechanism, is fully automatic since it relies on a set of parameters which ensure robust performance across a wide range of images. Some of these parameters could be fine-tuned for specific applications but the values introduced in this article have been extensively validated. Regarding speed, it takes 0.3sec for our system to enhance a 10 MPix image. Approximately 12% of this time is employed by the decision mechanism. These speeds allow for efficient automated implementations in image-rich workflows. For instance, a high speed printer is typically able to print 100 color pages per minute. If the default settings are used, we can see from Table 1 that the enhancement is only applied to approximately 40% of the images. So the final average is 0.14 seconds per image (more than MPix images per minute). User validation (algorithm quality) was tested through user preference evaluations. All tests were performed under uncontrolled conditions to keep the range of feasible applications for our algorithm as wide as possible. For this reason, the tests placed no restrictions on lighting conditions, print quality, image source and user expertise. Another feature common to most of the tests is the inclusion of the don t care option. This option allows for faster processing, robust results and answers correlated with a general rather than detailed perception. In a pairwise evaluation for a given image, if p i is the probability of preferring the processed image and q i the probability of a participant not caring, we define the user preference on this image as p i +0.5q i. The assumption being that, if required to, half of the people who do not care would choose the processed image while half would prefer the original. Another important topic in evaluations is how we pose preference to the users, a generally purpose-based question. Since the final objective of our method is to produce pleasant images, and that the most likely application of this kind of technology is within photofinishing environments, our question was of commercial nature. Each participant was proposed to assume the image were his and it had been printed at different print shops. He was then asked if he had to pay for only one of the printouts, which one would that be. User preference evaluations were done at three different stages of LCE design. We identify these phases as the algorithm design stage (AD), algorithm performance stage (AP) and decision mechanism design (DMD) stage. Experiments in the AD stage are made to determine the best approaches and an optimal set of parameters. These are the most complex tests usually involving complex ranking procedures when several images are analyzed simultaneously. AP tests were simple pairwise comparison tests contrasting LCE with another technique or, directly, with the original image. Besides a global measure of preference, when these tests are performed on a representative image sample we can use the result to predict the proportion of images LCE will improve/degrade. Finally, DMD are those tests whose sole purpose is to label data for training the decision mechanism. Table 2 details some of the main tests performed, the stage concerned, the resources involved and the main conclusions drawn from the test. In this table, the tests are in chronological order and one can understand the structure of the underlying research. For instance, when doing the general preference test (test 3) on what we thought was the final version of the algorithm, we observed users disliked some features related with our chrominance restoration scheme. We had the option of preventing this situation by predicting color changes through the decision mechanism or directly altering the algorithm. Since simply changing the chrominance model solved this problem at no increase in rejected images, we made the second choice. But the new scheme had to be once again evaluated (test 4) and the images from previous tests were no longer valid as labelled training samples for the decision mechanism, so new data had to be generated (tests 5 and 6). 6. CONCLUSIONS Enhancing contrast is a basic component of any suite of image enhancement algorithms. Contrast is closely linked with exposure in the sense that both over and under-exposed regions have the effect of diminishing the perceived contrast in those regions. LCE, the algorithm presented in this article, focuses on locally enhancing contrast in underexposed image regions using a background-based approach. Algorithms falling in this category can be very slow and introduce artifacts. We have shown our approach is both fast and does not introduce any artifacts. Furthermore LCE is fully automatic in the sense that images which do not benefit from enhancement with the default parameters are excluded from enhancement by a decision mechanism.

10 Figure 6. Before and after Local Contrast Enhancement for two sample images.

11 test stage users images method highlights screening comparison AD way ranking Ranking done on original im- of LCE age, CLAHE, 8 global con- with alternatives trast 6 and LCE with 3 diff. parameters. Plackett-Luce statistic indicate probability of preferring our method to original image is highest and above 0.7 original vs. local AD pairwise com- select appropriate gain for experiment parison LCE general preferencparison AP pairwise com- large-scale test to identify key algorithm issues prior to de- sign of decision mechanism chrominance approach AD three-way com- original vs. additive, as addiparisotive/multiplicative approach multiplicative for chrominance restoration. or mixed image preference DMD method image labelling for decision labelling mechanism. extended image DMD method image labelling for decision preference mechanism, using prototype labelling version of DM for selecting informative images Table 2. Some of the main experiments performed for evaluating user preference. ACKNOWLEDGMENTS We would like to acknowledge the valuable insights and observations contributed by Reiner Eschbach. REFERENCES 1. K. Devlin, A. Chalmers, A. Wilkie, and W. Purgathofer, Star: Tone reproduction and physically based spectral rendering, in State of the Art Reports, Eurographics, Sep J. DiCarlo and B. Wandell, Rendering high dynamic range images, in Proc. of the SPIE: Image Sensors, 3965, pp , R. Fattal, D. Lischinski, and M. Werman, Gradient domain high dynamic range compression, ACM Transactions on Graphics 21 3, pp , P. Ledda, A. Chalmers, T. Troscianko, and H. Seetzen, Evaluation of tone mapping operators using a high dynamic range display, in Proc. ACM SIGGRAPH 05, pp , J. Tumblin and H. Rushmeier, Tone reproduction for realistic images, IEEE Computer Graphics and Applications 13, pp , Nov R. Eschbach, B. Waldron, and W. Fuss, US Patent : Image-dependent luminance enhancement, Xerox Corporation. 7. G. W. Larson, H. Rushmeier, and C. Piatko, A visibility matching tone reproduction operator for high dynamic range scenes, IEEE Transactions on Visualization and Computer Graphics 3, pp , K. Zuiderveld, Contrast limited adaptive histogram equalization, in Graphic Gems IV, A. Press, ed., pp , J. Stockham, Image processing in the context of a visual model, in Proc. of the IEEE, 60, pp , 1972.

12 10. E. Land and J. McCann, Lightness and retinex theory, Journal of the Optical Society of America 1, pp. 1 11, Jan K. Chiu, K. Herf, M. Shirley, P. Swamy, S. Wang, and K. Zimmerman, Spatially nonuniform scaling functions for high contrast images, in Proc. Graphics Interface 93, M. Kaufmann, ed., pp , D. Jobson, Z. Rahman, and G. Woodell, A multiscale Retinex for bridging the gap between color images and the human observation of scenes, IEEE Transactions on Image Processing 10, pp , July J. Tumblin and G. Turk, LCIS: A boundary hierarchy for detail-preserving contrast reduction, in Proc. ACM SIGGRAPH 99, A. Rockwood, ed., pp , C. Tomasi and R. Manduchi, Bilateral filtering for gray and color images, in ICCV 98: Proc. of the 6th Intl. Conference on Computer Vision, I. C. Society, ed., pp , F. Durand and J. Dorsey, Fast bilateral filtering for the display of high dynamic range images, ACM Transactions on Graphics 21 3, pp , C. Yang and J. Rodriguez, Efficient luminance and saturation processing techniques for color images, Journal of Visual Communication and Image Representation 8, pp , Sep Z. Fan and R. de Queiroz, Identification of bitmap compression history: JPEG detection and quantizer estimation, IEEE T. on Image Processing 12, Feb R. Datta, D. Joshi, J. Li, and J. Wang, Studying aesthetics in photographic images using a computational approach, in Proc. ECCV 2006, A. Leonardis, H. Bischof, and A. Pinz, eds., 3953, pp , 2006.