ASINGLE pixel on color LCD is generally composed of

Transcription

1 3818 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 10, OCTOBER 2013 Luma-Chroma Space Filter Design for Subpixel-Based Monochrome Image Downsampling Lu Fang, Oscar C. Au, Ngai-Man Cheung, Aggelos. K. Katsaggelos, Houqiang Li, and Feng Zou Abstract In general, subpixel-based downsampling can achieve higher apparent resolution of the down-sampled images on LCD or OLED displays than pixel-based downsampling. With the frequency domain analysis of subpixel-based downsampling, we discover special characteristics of the luma-chroma color transform choice for monochrome images. With these, we model the anti-aliasing filter design for subpixel-based monochrome image downsampling as a human visual system-based optimization problem with a two-term cost function and obtain a closedform solution. One cost term measures the luminance distortion and the other term measures the chrominance aliasing in our chosen luma-chroma space. Simulation results suggest that the proposed method can achieve sharper down-sampled gray/font images compared with conventional pixel and subpixel-based methods, without noticeable color fringing artifacts. Index Terms Subpixel rendering, down-sampling, human visual system, improve resolution. I. INTRODUCTION ASINGLE pixel on color LCD is generally composed of three individual color components, typically three color elements ordered (on various displays) either as blue, green, and red (BGR), or as red, green, and blue (RGB). These three color components, known as subpixels, are fused together to appear as a single color to human due to the blurring effect by the optics and spatial integration by nerve cells in the human eyes [1], [2]. Subpixel-based techniques were originally proposed for the rendering of monochrome fonts on LCDs. Before that, pixel-based font rendering was used and the finest detail that can be displayed on an LCD was a single pixel. Then, researchers found that, by adjusting subpixel values of neighboring pixels, the number of units that may be independently controlled to reconstruct the image increases, Manuscript received May 28, 2012; revised December 20, 2012 and April 22, 2013; accepted April Date of publication May 13, 2013; date of current version August 28, This work was supported in part by the Research Grants Council of the Hong Kong Special Administrative Region GRF under Project , and the Natural Science Foundation of China under Contract The associate editor coordinating the review of this manuscript and approving it for publication was Prof. A. N. Rajagopalan. L. Fang and H. Li are with the Department of Electronic Engineering and Information Science, the University of Science and Technology of China, Hefei , China ( fanglu@ustc.edu.cn; lihq@ustc.edu.cn). O. C. Au and F. Zou are with Hong Kong University of Science and Technology, Kowloon , Hong Kong ( eeau@ust.hk; fengzou@ust.hk). N.-M. Cheung is with the Singapore University of Technology and Design, Singapore ( ngaiman_cheung@sutd.edu.sg). A. K. Katsaggelos is with Northwestern University, IL USA ( aggk@eecs.northwestern.edu). Color versions of one or more of the figures in this paper are available online at Digital Object Identifier /TIP IEEE Fig. 1. (a) Letter m in italic, (b) pixel rendered m with jagged edges, (c) subpixel rendered m with smooth edges. and it becomes possible to shift the apparent position or orient the edge of the font by one or two subpixels to achieve better edge reconstruction. Methods that take into account the interaction between displaying technology and human visual system are called as subpixel rendering algorithms [1]. In 1998, Microsoft announced a subpixel-based font display technology, ClearType, which is able to improve the readability of small text on regular LCD with three vertical stripe subpixels. In Fig. 1, we illustrate the difference between pixelbased rendering and subpixel rendering (ClearType). We can see that pixel-based rendering causes sawtooth artifacts in the sloping edges. With ClearType, we can borrow sub-pixels from neighboring pixels, and features of text as small as a fraction of a pixel can be displayed to better reconstruct the shape information and reduce the staircase artifacts [3], [4]. Since the number of individual and controllable reconstruction points in LCD can be increased by three times by considering subpixels, application of subpixel rendering in down-sampling scheme may lead to improvement in apparent resolution. Daly et. al. [5] have proposed a simple subpixelbased down-sampling pattern which we call Direct Subpixelbased Down-sampling (DSD), and it decimates the red, green, and blue components (i.e., the three subpixels) of the pixel alternately in horizontal direction from three different pixels in original large image as shown in Fig. 2(c). Let (r, g, b) and (R, G, B) be the RGB components of one pixelinlowandhighresolutionimages respectively. From the signal point of view, the RGB signal of one pixel in an image (either high resolution or low resolution image) occupies the same pixel location. Direct Pixel-based Down-sampling (DPD) sampling pattern decimates in a way that the RGB three signal of one pixel in DPD image is copied from one pixel in high resolution image, i.e., the RGB components of (i, j) th pixel in DPD image are copied from the (3i 2, 3 j 1) th pixel of high resolution image, such that r i, j = R 3i 2,3 j 1, g i, j = G 3i 2,3 j 1, b i, j = B 3i 2,3 j 1 as shown in Fig. 2(a).

2 FANG et al.: LUMA-CHROMA SPACE FILTER DESIGN 3819 Fig. 2. (a) Direct Pixel-based Downsampling (DPD) (b) magnified result of DPD, where grass is broken due to aliasing artifacts (c) Direct Subpixel-based Downsampling (DSD) (d) magnified result of DSD, where grass is smooth but has color fringing artifacts. In other words, DPD has regular 3:1 signal sampling spacing from signal point of view. While for DSD, thanks to the fact that a pixel in an LCD can be decomposed to red, green and blue three controllable subpixels, we can control each color component of input high resolution image separately to obtain DSD image, i.e., r i, j = R 3i 2,3 j 2, g i, j = G 3i 2,3 j 1, b i, j = B 3i 2,3 j as shown in Fig. 2(c). Intuitively, DSD achieves higher sampling frequency than the whole-pixelbased DPD sampling from signal point of view. In particular, for a monochrome image, one color component is enough to represent the whole pixel information. Consequently, DSD has effectively three times higher horizontal sampling frequency (1:1) than DPD (3:1), which contributes a lot to the advantage of DSD over DPD. From the displaying point of view, the RGB signal of one pixel will be projected to different locations of LCD according to the exact physical pattern of LCD. Such displaying effect works for all color images (either high resolution image or low resolution DPD and DSD images). For example, the Red, Green and Blue displaying unit in conventional RGB stripe LCD is arranged in a horizontal way, such that the red and blue signal will be projected to the left and right positions of the green signal, as shown in Fig. 2(a) and 2(c). Although all images may inherit a spatial offset when rendering, the red and blue signal in subpixel-based down-sampling is already a shifted version of green signal, which fortunately conceals out with the offset in rendering stage. In other words, the spatial arrangement of subpixel-based image actually fits well with the spacing of LCD display matrix. Fig. 3. Example to show how color fringing artifacts occur in subpixel-based down-sampling. Fig. 2(b) and 2(d) are the corresponding results of applying DPD and DSD on a gray image respectively. It is clear that the aliasing artifacts in pixel-based down-sampled image result in broken grass (Fig. 2(b)). And subpixel-based down-sampling may introduce annoying perceptual color artifacts around edges (Fig. 2(d)). Fig. 3 shows how color fringing artifacts occur in subpixel-based down-sampling. Suppose there is a vertical edge in the original high resolution image, with a left pixel belonging to object 1 of white color, and the right two pixels belonging to object 2 of black color. After DSD takes the red component from the left pixel and green/blue components from the right two pixels to form one pixel, the pixel in the low resolution image is red, not black and not white. It is this artificial red color that causes the color fringing artifact. While examining Fig. 2(d), it is interesting to see that the grass in subpixel-based down-sampled image appears continuous with the help of colors, indicating that the color

3 3820 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 10, OCTOBER 2013 Fig. 4. An illustration of the differences between DPD and DSD. fringing artifacts caused in subpixel-based down-sampling mix together with the extra apparent luminance details, and DSD effectively preserves more original information than DPD does for non-strictly-vertical sloping edge. We use a very simple example in Fig. 4 to describe how does DPD or DSD work on the thin sloping edge, where the square area represents a pixel, the rectangular area represents a subpixel, and the black region indicates that the RGB components are (0, 0, 0). It can be shown that for a non-strictly-vertical sloping edge, DSD helps to preserve part of the whole-pixel information (R, G or B subpixels alternately) even in vertical direction, achieving better continuity of the sloping edge than DPD does. Although subpixel rendering can increase apparent luminance resolution, color fringing artifacts may appear in a given orientation on the color subpixel arrangement as verified in Fig. 2(d). Therefore, it is necessary to design a proper filter to suppress color fringing artifacts without significantly damaging the improved apparent luminance resolution. Based on psychophysical experiments, Platt defined an error metric in frequency domain, and derived the anti-aliasing filter coefficients by minimizing the error metric [3]. In [4], Betrisey et. al. applied the results of Platt in [3] to monochrome (gray-scale) font rendering which is the basis of the Microsoft s ClearType system. In [1], Gibson used a five-tap low-pass filter to smooth the results of the subpixel-based rendered image. However, all these low-pass filters relieve color fringing artifacts at the expense of blurring artifacts. Later, Klompenhouwer and Haan in [6] proposed to deal with the consequences of the subpixel structure based on frequency domain analysis. Under the assumption of bandlimited flat baseband signal, they further show the horizontal frequency spectrum of pixel and subpixel samplings. It is meaningful that [6] covers both Vertical stripe and Delta- Nabla subpixel arrangements. Nevertheless, the paper did not clearly discuss how to balance the trade-off between luminance sharpness and chrominance distortion when designing the subpixel polyphase filter. Hirakawa and Gu in [7] introduced a rendering error cancellation embedding technique for color flat-panels, based on their previous work of frequencydomain display analysis for color filter array design [8]. They considered the modeling of image formation in low-level human vision with the spatial and temporal contrast sensitivity function (i.e., a low-pass effect on the granularity of individual elements). Then, the different HVS responses to full color signal and single color signal were compared and optimized in Fourier domain. The proposed rendering error cancellation embedding successfully reduced the undesirable effects caused by rendering error in RGBW Quad panel. Unfortunately, this technique cannot be applied to subpixel rendering with traditional RGB vertical or diagonal panel. Frequency-domain analysis approach used to be proposed for demosaicking problem as well. For example, Alleysson et al. defined a model that characterizes a one-color per spatial position image (CFA image) as a coding into luminance and chrominance of the corresponding three colors per spatial position image (full color image), based on which, the demosaicking algorithm was specifically designed [9], [10]. Inspired by [9], [10], we can think of subpixels as just having a rectangular pixel of CFA image, then we can borrow the idea of frequency-domain analysis for subpixel problem, where the equivalent vertical Fourier may go from π to +π, and the horizontal Fourier axis may go from 3π to +3π. Although the frequency-domain analysis content of our work is inspired by [9], [10], how to achieve frequency characteristics and design corresponding sampling schemes for subpixelbased subsampling are not straightforward. In [8], Hirakawa examined the problem of color filter array design by explicitly considering the spectral wavelength representation induced by the choice of array pattern. They showed that the spectral replicas of the chrominance signals induced by CFA patterns are centered around frequencies are overlapping with the luminance signal spectrum. [8] talked about the cut off frequencies of luma and chroma based on human vision. However, [8] did not tell how to render subpixels, since the analysis was proposed for designing new CFAs. In this paper, we try to find out how much extra apparent resolution of subpixel-based down-sampling is actually gained by involving Human Visual System effect. Our proposed scheme is different from aforementioned methods, in particular: We specifically investigate unique luma-chroma space definition for subpixel-based subsampling, with which, the frequency spectrum of subsampled image resembles the original image the most. We propose to deal with a penalty function by jointly considering the frequency characteristics of subpixelbased subsampling in multi-channels (both luma and chroma), with the purpose of jointly minimizing the apparent luminance loss and chrominance distortion. We further introduce a cross-channel balance weight (between apparent luminance loss and chrominance distortion) by taking into account of different sensitivities of our HVS on luminance and chrominance. The corresponding closed-form solution is very simple for real implementation, and its inherent luma-chroma weight is simply controllable to intuitively reflect the different preferences (sensitivities) of different human eyes on luminance and chrominance. The rest of this paper is organized as follows. Section II introduces the detail of the proposed luma-chroma space antialiasing filter design for DSD (LC-DSD). In Section III, we evaluate LC-DSD using a number of gray-scale images

4 FANG et al.: LUMA-CHROMA SPACE FILTER DESIGN 3821 and font images. Finally, we draw the conclusions in Section IV. II. ANTI-ALIASING FILTER DESIGN IN LUMA-CHROMA SPACE FOR DSD (LC-DSD) A. Frequency Analysis in Luma-Chroma Space Suppose an input large image of size dm dn is to be down-sampled by a factor of d to a small image of size M N, where d is an integer. Let C k be the three color components (k = 1, 2, 3 representing R, G and B respectively) of the input large image, each of size dm dn. In the downsampling processing, one pixel out of a d d block in C k will be sampled and the sampling location can be different for different k. Let(m k, n k ) be the sampling location inside the d d block for C k such that m k and n k are values between 1tod. We define a pseudo image C,d k of size dm dn with the corresponding sampled values of C k at the sampling locations and zero elsewhere such that C,d k (i, j) = C k (i, j), for i = di + m k, j = dj + n k, 0 i M 1, 0 j N 1 0, otherwise M 1N 1 = C k (i, j) δ ( i (di + m k ), i (dj + n k ) ). (1) i =0 j =0 Considering one pixel in the stripe RGB LCD contains 3 subpixels and there is no natural downsampling pattern such as DSD for d:1(d = 3) downsampling, we thus investigate the case of d = 3 for DPD and DSD sampling patterns as shown in Fig. 2. For other downsampling ratio d = 3, the analytical model can be extended accordingly. Given d = 3, we have C,3 k (i 0, j 0 ) C k (i 0, j 0 ), for i 0 = 3i + m k, j 0 = 3 j + n k, = 0 i M 1, 0 j N 1 0, otherwise M 1 N 1 = C k (i 0, j 0 ) δ ( i 0 (3i +m k ), i 0 (3 j +n k ) ). (2) i =0 j =0 Taking the Fourier transform of (2), we have (3), shown at the bottom of the page, where represents the corresponding Fourier transform, represents convolution and stands for the complex conjugation. (ξ) T = [ δ(ξ ), δ(ξ), δ(ξ 1 3 )], (η) T = [ δ(η ), δ(η), δ(η 1 3 )], p k = e j 2π 3 m k and q k = e j 2π 3 n k. Considering DSD-based pseudo image (denoted as C,DSD k ), where (m 1, n 1 ) = (2, 1), (m 2, n 2 ) = (2, 2), (m 3, n 3 ) = (2, 3), wehavep 1 = p 2 = p 3 = e j 2π 3, q 1 = e j 2π 3, q 2 = e j 2π 3, q 3 = 1. Substituting p k and q k into (3), we obtain the frequency representation of C k,dsd, i.e., Ĉ,DSD 1 (ξ, η) = R DSD (ξ, η) 1 e j 2π 3 e j 2π 3 1 = 9 (ξ)t e j 2π 3 1 e j 2π 3 (η) R(ξ, η). (4) e j 2π 3 e j 2π 3 1 Similarly, we have Ĝ DSD (ξ, η) and B DSD (ξ, η). Examining (4), there are nine replicated spectra in Ĉ,DSD k (ξ, η), each with different scaling factor of Ĉk (ξ, η). Inspired by the fact that the human visual system (HVS) perceives a color stimulus in terms of luminance and chrominance attributes rather than in terms of RGB values [11], we consider to analyze the frequency characteristics of DSD image in the luma-chroma space than directly in the RGB space. According to [12], [13], there are various definitions of the luma-chroma space based on the Smith and Pokorny cone responses. We will choose a definition with some nice properties. Consider a generalized luma-chroma space, Y = ay r R + ag Y G + ab Y B U = a U r R + ag U G + ab U B (5) V = av r R + ag V G + ab V B, where Y is the generalized luminance component, U and V are the generalized chrominance components. a k X are the coefficients for the linear color space transform where X can be Y, U or V and k can be r, g, orb, such that ay r +ag Y +ab Y = 1, au r + ag U + ab U = 0andar V + ag V + ab V = 0. The corresponding Fourier representation of Y, U or V component is X(ξ, η) = a r X R(ξ, η) + a g X Ĝ(ξ, η) + ab X B(ξ, η). (6) For the DSD image, it becomes X DSD (ξ, η) = a r X R DSD (ξ, η)+a g X ĜDSD(ξ, η)+a b X B DSD (ξ, η). (7) Substituting R DSD, Ĝ DSD and B DSD from (4) into (7) and denote k 1 = e j 2π 3, k 2 = e j 2π 3,wehave X DSD (ξ, η) = 1 k 1 X l (ξ +,η + ) k 1 X(ξ +,η) k 1 X r (ξ +,η ) 9 1T 3 X l (ξ, η + ) X(ξ, η) X r (ξ, η ) 1 3, (8) k 2 X l (ξ,η + ) k 2 X(ξ,η) k 2 X r (ξ,η ) where 1 T 3 = [1, 1, 1], ξ ± = ξ ± 1 3, η± = η ± 1 3,and X l (ξ, η) = k 2 a r X R(ξ, η) + k 1 a g X Ĝ(ξ, η) + ab X B(ξ, η) X r (ξ, η) = k 1 a r X R(ξ, η) + k 2 a g X Ĝ(ξ, η) + ab X B(ξ, η). (9) p k q k δ(ξ + 1 Ĉ,3 k (ξ, η) = 1 3,η+ 1 3 ) + p kδ(ξ,η ) + p kqk δ(ξ 1 3,η+ 1 3 ) 9 +q k δ(ξ + 1 3,η)+ δ(ξ,η) + q k δ(ξ 1 3,η) Ĉ k (ξ, η) +pk q kδ(ξ + 1 3,η 1 3 ) + p k δ(ξ,η 1 3 ) + p k q k δ(ξ 1 3,η 1 3 ) 1 = 9 (ξ)t p kq k p k p k qk q k 1 qk (η) pk q k pk Ĉk (ξ, η), p k q k (3)

5 3822 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 10, OCTOBER 2013 For DPD, the X DPD (ξ, η) can be expressed in a similar way where X l (ξ, η) = X r (ξ, η) = X(ξ, η) due to (m k, n k ) = (2, 2) for all RGB color components. Signal processing theory governs that downsampling in spatial domain would lead to some form of spectra replication in frequency domain. If the downsampling factor is 3:1 (l = 3), there would be 9 replicated spectra situated at (ξ, η), whereξ { { 1 3, 0, 1 3} and η 1 3, 0, 1 } 3. For DPD, it is X(ξ, η) at all the 9 locations. To prevent aliasing (the replicated spectra overlapping with each other), an ideal low-pass filter with cut-off frequency of 1/6 should be applied before the 3:1 DPD [14]. For DSD, it is X(ξ, η) located at ( 1 3, 0), (0, 0) and ( 1 3, 0) with scaling factors k 1 9, 1 9 and k 2 9 respectively, and X l (ξ, η) or X r (ξ, η) at the remaining 6 locations with corresponding scaling factors according to (8). Consider now the special case of a monochrome image. The three color components are identical such that R = G = B and R(ξ, η) = Ĝ(ξ, η) = B(ξ, η). This implies Ŷ (ξ, η) = (a r Y + ag Y + ab Y ) R(ξ, η) = R(ξ, η) Û(ξ, η) = (a r U + ag U + ab U ) R(ξ, η) = 0 V (ξ, η) = (a r V + ag V + ab V ) R(ξ, η) = 0. (10) When DSD is applied to the monochrome image, there will still be 9 replicated spectra for X DSD,whereX is Y, U or V. The corresponding X l and X r become scaled versions of R(ξ, η) also, X l (ξ, η) = (k 2 a r X + k 1a g X + ab X ) R(ξ, η) X r (ξ, η) = (k 1 a r X + k 2a g X + ab X ) R(ξ, η). (11) The Ŷ DSD is effectively R at all the 9 locations with different scaling factors. And the Û DSD and V DSD contain only 6 nonzero spectra because the 3 center ones at ( 1 3, 0), (0, 0), and ( 1 3, 0) are zero. The 6 non-zero spectra are also R with different scaling factors, given by Û l (or V l ) and Û r (or V r ) from (11). For the X DSD, the 3 center spectra at ( 1 3, 0), (0, 0), and ( 1 3, 0) are always X with scaling factors k 1 9, 1 9 and k 2 9 which do not change with color space definition. The 3 spectra on the left and the 3 on the right are scaled versions of X l (ξ, η) and X r (ξ, η), which in turn are scaled versions of R(ξ, η) according to (11). But the scaling factors of R(ξ, η) would change with the color space definition. This provides the possibility for us to choose a color space such that X DSD resembles X the most. We can multiply X DSD by 9 so that the center spectrum at (0, 0) is exactly X. While the two spectra at ( 1 3, 0) and ( 1 3, 0) would not change with the color space definition, we will seek to find a color space such that the X l (ξ, η) and X r (ξ, η) are minimal, which is true when the scaling factors in (11), i.e. k 2 a r X + k 1a g X + ab X and k 1 a r X +k 2a g X +ab X, are minimized. As k 2a r X +k 1a g X +ab X and k 1 a r X + k 2a g X + ab X are conjugated, we can minimize any one of these two scaling factors. For the Ŷ l (or Ŷ r ), min a r Y,ag Y,ab Y k 1 a r Y + k 2a g Y + ab Y s.t. a r Y + ag Y + ab Y = 1 k 1 = e j 2π 3 k 2 = e j 2π 3 ay r, ag Y, ab Y (0, 1). (12) It can be shown that (12) is convex and the optimal solution is ay r = a g Y = ay b = 1 3. The corresponding minimal value of the magnitude of Ŷ l (or Ŷ r ) is k 1 ay r + k 2a g Y + ab Y =0. For the Û l (or Û r ), min a r U,ag U,ab U k 1 a r U + k 2a g U + ab U s.t. a r U + ag U + ab U = 0 V a r U, ag U < 0, ab U = 1 2 k 1 = e j 2π 3 k 2 = e j 2π 3. (13) Note that we add a common constraint au b = 1 2 for the purpose of U [ 128, 128] [11]. The optimal solution of (13) is au r = ag U = 1 4 and ab U = 1 2. The corresponding minimal value of the magnitude of Û l (Û r ) is k 1 au r +k 2a g U +ab U = 3 4. Similarly, we do the same for V l ( V r ). The corresponding optimal solution is a r = ab V = 1 4. = 1 2 and ag V Recall that the monochrome image implies Ŷ = R = Ĝ = B, Û = 0and V = 0. The choice of ay r = ag Y = ab Y = 1 3 implies Ŷ l = Ŷ r = 0. The choice of au r = ag U = 1 4 and au b = 1 2 implies Û l = Û r = 3 4 Ŷ. The choice of ar V = 1 2 and a g V = ab V = 1 4 implies V l = 3 4 k 2Ŷ, V r = 3 4 k 1Ŷ. In our chosen color space, (8) can be simplified to Ŷ DSD = 1 0 k 1 Ŷ (ξ +,η) 0 9 1T 3 0 Ŷ (ξ, η) 0 1 3, (14) 0 k 2 Ŷ (ξ,η) 0 Û DSD = 1 k 1 Ŷ (ξ +,η + ) 0 k 1 Ŷ (ξ +,η ) 12 1T 3 Ŷ (ξ, η + ) 0 Ŷ (ξ, η ) 1 3, (15) k 2 Ŷ (ξ,η + ) 0 k 2 Ŷ (ξ,η ) and V DSD = T 3 Ŷ (ξ +,η + ) 0 k 2 Ŷ (ξ +,η ) k 2 Ŷ (ξ, η + ) 0 k 1 Ŷ (ξ, η ) k 1 Ŷ (ξ,η + ) 0 Ŷ (ξ,η ) 1 3. (16) In Fig. 5, we show the typical magnitude spectra of such Ŷ DSD and Û DSD. Those of DPD are also shown for comparison. The magnitude spectrum of V DSD is similar to that of Û DSD. B. Cut-Off Frequency Selection in Luma-Chroma Space Let A = 1 3 which is the horizontal and vertical shifts of the replicated spectra. From signal processing theory, the optimal anti-aliasing filter for DPD is an ideal low-pass filter with cut-off frequency of A/2. Let fc H and fc V be the horizontal and vertical cut-off frequencies of an ideal low-pass filter respectively. Then fc,dpd H = f c,dpd V = A 2 = 1 6 for DPD. As the 9 replicated spectra of X DSD and X DPD are different, the fc,dsd H and f c,dsd V may be different from those of DPD. The vertical cut-off frequency of DSD would be simply fc,dsd V = A 2 = 1 6, due to the almost identical replicated spectra in the vertical direction. Consider the horizontal direction.

6 FANG et al.: LUMA-CHROMA SPACE FILTER DESIGN 3823 Fig. 5. Magnitude spectra of Y and U of Direct Pixel-based Downsampling (DPD) and Direct Subpixel-based Downsampling (DSD), where (0, 0) represents the frequency, fc H and fs H are the Nyquist cut-off frequency and sampling frequency in Horizontal direction respectively. For the luminance component Ŷ DSD, the replicated spectra at η = 1 3 and η = 1 3 are zero. Thus f H c,dsd should be as large as possible to retain the most of the center spectrum at (0, 0). For the chrominance components, the replicated spectra at η = 1 3 and η = 1 3 are scaled and phase shifted versions of Ŷ for both U and V components, according to (15) and (16). Such replicated spectra, if not removed, may cause aliasing to the chrominance components. Therefore fc,dsd H should be as small as possible to remove the most of such replicated spectra. Consider both the luminance and chrominance situation, we formulate the optimal fc,dsd H selection as a joint optimization problem with one cost term being luminance distortion in Ŷ DSD and the second term being chrominance aliasing in Û DSD and V DSD. As depicted in Fig. 2(d) (where the grass in DSD image is continuous with the help of color information), subpixel rendering provides extra apparent luminance resolution with the expense of color fringing artifacts. How to find a good tradeoff between apparent luminance sharpness and chrominance distortion is a tough problem. Fortunately, due to the fact that our human eyes are relatively more sensitive in luma channel than in chroma channel, the color fringing artifact tends to be less noticeable compared to the improvement of apparent luminance resolution. Based on aforementioned observation and prior knowledge, we introduce K (denoted as lumachroma weight ) in the penalty function to be the weighting factor that captures the different sensitivities of the HVS to luminance and chrominance errors, such that min f H c,dsd s.t. KSloss Y ( f c,dsd H ) SU alias ( f c,dsd H ) SV alias ( f c,dsd H ) [ fc,dsd H 0, 1 ], (17) 2 where Sloss Y ( f c,dsd H ) is the distortion due to the loss of the high frequency detail of the center spectrum in Ŷ DSD, Salias U ( f c,dsd H ) and Salias V ( f c,dsd H ) are distortions due to the aliasing artifacts in Û DSD and V DSD. As studied in [15], [16], the majority of the spectral energy of a typical signal is highly concentrated in the lower frequency, and the Laplacian probability density function tends to be a good model to approximate the magnitude of the frequency spectrum. We thus approximate the normalized Fig. 6. An illustration of S Y loss ( f H c,dsd ) and SU alias ( f H c,dsd ). spectrum of Ŷ using a zero-mean circularly symmetric ( Laplacian model with variance 2λ 2 0, i.e., Ŷ E 0 2λ 1 0 exp f λ 0 ), where E 0 is the total energy of Ŷ. In Fig. 6, the blue curve represents the center spectrum Ŷ at (0, 0) in Ŷ DSD,andthe purple curve represents Û r at (0, 1 3 ) in Û DSD. The area of the red and green regions are the luminance distortion (Sloss Y ) and chrominance aliasing (Salias U ) respectively for any f H c,dsd. With the Laplacian approximation, (17) is re-written as min KS Y fc,dsd H loss ( f c,dsd H ) SU alias ( f c,dsd H ) SV alias ( f c,dsd H ) s.t.s Y loss ( f H c,dsd ) = f H c,dsd f H Salias U ( f c,dsd H ) = c,dsd f H Salias V ( f c,dsd H ) = c,dsd Û r ( f ) = 3 4 Ŷ ( f ) V r ( f ) = 3 4 k 1Ŷ ( f ) Ŷ ( f ) df Û r ( f A) df V r ( f A) df

7 3824 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 10, OCTOBER 2013 Ŷ ( f ) = E ( 0 exp f ) 2λ 0 λ [ 0 fc,dsd H 0, 1 ]. (18) 2 By taking derivative of {KSloss Y ( f c,dsd H )+ 1 2 SU alias ( f c,dsd H )+ 1 2 SV alias ( f c,dsd H )} with respect to f H c,dsd and setting the result to be zero, we have K Ŷ ( fc,dsd H ) =1 2 Û r ( fc,dsd H A) +1 2 V r ( fc,dsd H A), (19) in which K helps to balance the different sensitivities of HVS to the luminance and chrominance errors. Obviously, Ŷ ( fc,dsd H ) = 2 1 Û r ( fc,dsd H A) V r ( fc,dsd H A) for K = 1, indicating that the instantaneous decrease of high frequency detail loss of Ŷ DSD is equal to the instantaneous increase of chrominance aliasing distortion in Û DSD and V DSD. Substituting Û r ( f A) = 3 4 Ŷ ( f A), V r ( f A) = 3 4 k 1Ŷ ( f A) and Ŷ ( f ) = E 0 2λ 0 exp( f λ 0 ) into (19), we have f H c,dsd = A 2 + λ 0 2 ln ( 4 3 K ). (20) It can be shown that the objective function tends to favor the minimization of luminance distortion as the increase of K, resulting in a higher cut-off frequency fc,dsd H and preserving more high frequency luminance information. While the bad effect of increasing K is the increase of color fringing artifact. Since different human eyes may have different sensitivities in luma or chroma channel, the value of K could be adaptively controllable, i.e., for someone who are less sensitive to chroma, a large K could be chosen, vice versa. For example, the optimal cut-off frequency is the Nyquist cut-off frequency fc,dsd H = A/2 whenk = 3 4 ;wehave f H c,dsd > A/2 for K > 3 4 ;and f H c,dsd < A/2 fork < 3 4. But typically, HVS is more sensitive to luminance than chrominance [17], [18] and thus we expect K > 1. C. Real Implementation To investigate how strongly K may affect the LC-DSD result, we simulate LC-DSD under various K values for a large number of gray and font images respectively. To objectively measure how does luminance sharpness vary with different K values, we simply apply high-pass filter [1, 1] on each of LC-DSD images and compute l 1 -norm high frequency energy (HFE) value for each of LC-DSD images [19]. Note that we use the HFE value of PDAF as normalization reference, hence the HFE value of PDAF is 1 and image with higher HFE value tends to be sharper. The average HFE values versus K for gray and font images are plotted as follows. Based on (17), the penalty function tends to favor the minimization of luminance distortion as the increase of K, resulting in a higher cut-off frequency fc,dsd H, as shown in (20). It is well known that downsampling with a higher fc,dsd H tends to preserve more high frequency information. In other words, a higher K implies that the down-sampled image tends to have higher HFE value, as observed in Fig. 7. Another observation is that the HFE value tends to be approximately Fig. 7. High Frequency Energy (HFE) values of gray and font images downsampled via LC-DSD under different K values. constant when K 4 for gray images and K 12 for font images. The possible explanation for the constant HFE value may be the compact and band-limited frequency spectra, so that the gain is negligible although K ( fc,dsd H )islarge enough. In general, the HFE values of font images are slightly higher than those of gray images, due to extremely high contrast between black color and white color in font images. As suggested by our experiments, the luma-chroma weight K is set to be 4 for gray images and 12 for font images. Consider the spatial implementation of LC-DSD, we need to truncate the ideal low-pass filter to be k k tap filter, where k = 9 is a good choice under our investigation of how anti-aliasing filter affects image contrast in [20]. Due to the symmetric property, a total of 15 multiplications are needed for each subsampled pixel (there are M N subsampled pixels). Therefore, the computational complexity of LC-DSD is basically 15MN multiplications, which is comparable to those of other concerned algorithms, i.e., O(MN) for PDAF (Pixel-based Downsampling with Anti-aliasing Filter) [14], Betrisey [4] and Gibson [1] (which are subpixel-based methods). III. EXPERIMENTAL RESULTS We evaluate the proposed LC-DSD by comparing it with PDAF (Pixel-based Downsampling with Anti-aliasing Filter) [14], Betrisey [4] and Gibson [1] (which are subpixel-based methods). As PDAF usually introduces blurring artifacts, a simple way is applying post-processing sharpening filter to improve the sharpness of PDAF image. In our experiment, the unsharp mask filter is chosen since it fits the behavior of the algorithm to some extent, and we denote PDAF followed by unsharp filter as PDAF-UF. Fig. 8(a) and 8(b) are various gray and font testing images, which would be downsampled by a factor of 3 both horizontally and vertically using concerned methods: LC-DSD, PDAF, PDAF-UF, Betrisey and Gibson. Both the objective and subjective performance measurements will be considered. The objective measurements will evaluate the luminance sharpness and chrominance distortion of concerned methods respectively. To measure the apparent luminance sharpness, we apply a simple high-pass filter in Luma channel (Y component) of the concerned image,

8 FANG et al.: LUMA-CHROMA SPACE FILTER DESIGN 3825 Fig. 8. (a) Gray-scale testing images 1 15, (b) font testing images then compute the corresponding l 1 -norm high frequency energy (HFE) value. A larger HFE value indicates more high frequency energy (higher apparent luminance sharpness). In terms of chrominance distortion (color fringing artifacts) measurement, we propose to compute the PSNR values of concerned subpixel-based methods in Chroma channels (U and V components). Note that the reference U and V values to compute MSE are simply 128 due to the monochrome property, PSNR U = 10 log 10 (U x 128) 2, (21) where U x is the U component of the concerned image x. Similarly, we have PSNR V. It can be shown in Table I that the HFE values of Gibson and Betrisey are comparable to those of PDAF, indicating that the down-sampled images using Gibson and Betrisey may appear similar sharpness level as those using PDAF. While the HFE values of PDAF-UF and LC-DSD are significantly larger than PDAF, Gibson and Betrisey. Such higher values suggest that PDAF-UF and LC-DSD effectively retain more high frequency detail than PDAF, Gibson or Betrisey, leading to sharper downsampled images. For l:1(l = 3) downsampling ratio, we use a common pixel-based decimation method (with bi-cubic filter) to resize L to be 3M 3N and then apply the proposed LC-DSD. We simulate this method for downsampling ratios of 4:1 and 5:1. Due to limited space, we only show HFE values of down-sampled images for the case 4:1 in Table I, from which, similar conclusions can be obtained as 3:1 downsampling. The performance of the chrominance distortion is shown in Table II. As we expected, all the subpixel-based methods introduce color fringing artifacts, and the proposed LC-DSD suffers relatively more color fringing artifacts compared to Betrisey and Gibson. Fortunately, the color fringing artifacts are not as noticeable as the extra apparent luminance sharpness achieved in subpixel-based methods, as verified in Fig. 9 and 10. Since the special color fringing artifacts and blurring artifacts are best seen directly via LCD display, TABLE I LUMINANCE SHARPNESS MEASURE FOR IMAGES :1 Downsampling Image PDAF PDAF-UF Gibson Betrisey LC-DSD Ave :1 Downsampling Image PDAF PDAF-UF Gibson Betrisey LC-DSD Ave we only show some typical down-sampled images, and we recommend the readers to view our results on our homepage directly: fanglu/gray.htm.we can see that

9 3826 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 10, OCTOBER 2013 Fig. 9. Comparison of typical down-sampled images using PDAF and LC-DSD (with left part being PDAF and right part being LC-DSD). TABLE II CHROMINANCE DISTORTION MEASURE FOR IMAGES 1 15 PSNR U PSNR V Image Gibson Betrisey LC-DSD Gibson Betrisey LC-DSD Ave all the PDAF image, Betrisey image and Gibson image suffer burring artifacts to a certain extent as shown in Fig. 10. On a sharpness level, the down-sampled gray images using PDAF- UF and LC-DSD appear much sharper than those of PDAF, Betrisey and Gibson as depicted in Fig. 9 and 10. For better comparison, we particularly combine down-sampled images using PDAF and LC-DSD in Fig. 9. It is clear that LC-DSD retains more high frequency detail, leading to much better visual quality than PDAF. Consider the relatively low PSNR values in Chroma channel of LC-DSD, we combine a post-processing chroma-channel TABLE III LUMINANCE AND CHROMINANCE MEASURE FOR IMAGES Luminance Measure Chrominance Measure Image PDAF-UF PLC-DSD LC-DSD LC-DSD Ave low-pass filter with the proposed LC-DSD, which is denoted as Post-processed LC-DSD ( PLC-DSD ). Suppose a multichannel filter is applied in luma and chroma channels, i.e., the filter in each of Y, U and V channels could be particularly designed separately, denoted as H Y, H U and H V respectively, such that yuv = H Y H U 0 YUV = H YUV YUV, (22) 0 0 H V where yuv = [ y, u, v] T and YUV = [ Y, U, V ] T represent the YUV components of down-sampled low resolution image and high resolution input image that reshaped in row-wise to be column vectors respectively, is the normal matrix multiplication.

10 FANG et al.: LUMA-CHROMA SPACE FILTER DESIGN 3827 Fig. 10. Comparison of typical down-sampled font images using PDAF, Gibson, PDAF-UF, PLC-DSD and LC-DSD. The remaining problem is to determine the post-processing filter H U and H V in chroma channel. Intuitively, from Fig. 5(d), the optimal choice of chrominance filters would be HU = 0andH V = 0 due to the monochrome property. We compare the performance of PLC-DSD with other concerned methods on typical testing images both objectively and subjectively. In particular, we take PDAF-UF and LC-DSD for comparison. The luminance sharpness and chrominance distortion measurements for font source images are shown in in Table III. As we expected, the chrominance distortion in LC-DSD has been completely removed due to HU = 0 and HV = 0 (the PSNR value of U and V components is infinite). The subjective results of concerned methods for font images are depicted in Fig. 10. It seems that Gibson (subpixel-based) produces similar down-sampled image as PDAF (pixel-based) at first glance. Nevertheless, as verified in Fig. 10(b) and 10(g), the strong edges (such as sloping edges) in Gibson image appear smoother than those in PDAF with less jaggies due to subpixel-based processing. Similar situation exists for PDAF- UF and LC-DSD. As can be seen from Fig. 10(c) and 10(h), PDAF-UF produces sharp image with the sacrifice of staircase artifacts, resulting in jaggy sloping edges especially for italic fonts. On the contrary, LC-DSD effectively achieves high apparent resolution with better reconstruction of edge information, as verified in Figs. 10(e) and 10(j). In terms of the chrominance distortion, we may be happy with the result of PLC-DSD at the first glance, since all the color fringing artifacts have been removed due to the particular post-processing in chroma channel. Nevertheless, if we check with the details of sloping lines, it is surprising that subpixel effect disappears as well, resulting in jaggy sloping lines

11 3828 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 10, OCTOBER 2013 especially for italic fonts, as shown in Fig. 10(d) and 10(i). Therefore, simply removal of all the color information is not adoptable for subpixel-based situation. Because the color fringing artifacts in subpixel-based image actually inherit part of high frequency details, and subpixel rendering works by taking account of our human eyes are more sensitive to extra apparent luminance resolution than color fringing artifacts. This is also why the ClearType (installed in Microsoft windows system) always allows little color effect to help with the rendering of small fonts. IV. CONCLUSION Based on the special frequency characteristics of monochrome images, we choose a color space in which the spectrum of subpixel-based down-sampled image resembles the original image the most. With the chosen color space, a joint optimization scheme is proposed to balance the luma distortion and chroma aliasing. Simulation results show that the proposed method is effective in preserving apparent luminance sharpness without noticeable chrominance artifacts. In particular, it works well for font images. V. ACKNOWLEDGMENT The authors would like to thank Dr. Y. Chen for his insightful discussion and the Associate Editor Prof. A. N. Rajagopalan and three anonymous reviewers for their valuable comments. [15] L. Fang, K. Tang, O. C. Au, and A. K. Katsaggelos, Anti-aliasing filter design for subpixel downsampling based on frequency analysis, IEEE Trans. Image Process., vol. 21, no. 3, pp , Mar [16] E. Y. Lam and J. W. Goodman, A mathematical analysis of the DCT coefficient distributions for images, IEEE Trans. Image Process., vol. 9, no. 10, pp , Oct [17] R. Balasubramanian, C. A. Bouman, and J. P. Allebach, Sequential scalar quantization of color images, J. Electron. Imag., vol. 3, no. 1, pp , [18] C. A. Parraga, G. Brelstaff, T. Troscianko, and I. Moorhead, Color and luminance information in natural scenes, J. Opt. Soc. Amer. A, vol. 15, no. 3, pp , [19] L. Fang, O. C. Au, K. Tang, and X. Wen, Novel 2-D MMSE subpixelbased image downsampling, IEEE Trans. Circuits Syst. Video Technol., vol. 22, no. 5, pp , May [20] K. Tang, O. C. Au, L. Fang, Z. Yu, and Y. Guo, How anti-aliasing filter affects image contrast: An analysis from majorization theory perspective, in Proc. IEEE Int. Conf. Multimedia Expo, Jul. 2011, pp Lu Fang received the B.S. degree from the University of Science and Technology of China (USTC), Hefei, China, in 2007, and the Ph.D. degree from the Hong Kong University of Science and Technology (HKUST), Hong Kong, in She used to visit Northwestern University, Evanston, IL, USA, in 2010, under the support of Prof. A. K. Katsaggelos. From 2011 to 2012, she was the Post- Doctoral Research Fellow at HKUST and the Singapore University of Technology and Design (SUTD), Singapore. She is currently an Associate Professor at USTC. His current research interests include multimedia processing, image and video coding, and machine learning. REFERENCES [1] S. Gibson. Sub-Pixel Font Rendering Technology [Online]. Available: [2] L. Fang and O. C. Au, Subpixel-based image downsampling with minmax directional error for stripe display, IEEE J. Sel. Topics Signal Process., vol. 5, no. 2, pp , Apr [3] J. C. Platt, Optimal filtering for patterned displays, IEEE Signal Process. Lett., vol. 7, no. 7, pp , Jul [4] C. Betrisey, J. F. Blinn, B. Dresevic, B. Hill, G. Hitchcock, B. Keely, D. P. Mitchell, J. C. Platt, and T. Whitted, Displaced filtering for patterned displays, in SID Int. Symp. Dig. Tech. Papers, vol , pp [5] S. Daly, Methods and systems for improving display resolution in images using sub-pixel sampling and visual error filtering, U.S. Patent , Aug. 19, [6] M. A. Klompenhouwer and G. de Haan, Subpixel image scaling for color matrix displays, in SID Symp. Dig. Tech. Papers, 2002, pp [7] K. Hirakawa and J. Gu, High resolution subpixel and subframe rendering for color flatpanel and projector displays, in Proc. Int. Conf. Imag. Process., 2011, pp [8] K. Hirakawa and P. J. Wolfe, Fourier domain display color filter array design, in Proc. Int. Conf. Imag. Processing, 2007, pp. III-429 III-432. [9] D. Alleysson, S. Susstrunk, and J. Herault, Color demosaicing by estimating luminance and opponent chromatic signals in the Fourier domain, in Proc. Color Imag. Conf., 2002, pp [10] D. Alleysson, S. Susstrunk, and J. Herault, Linear demosaicing inspired by the human visual system, IEEE Trans. Image Process., vol. 14, no. 4, pp , Apr [11] E. Hamilton. (1992, Sep.). JPEG File Interchange Format Version 1.02 [Online]. Available: [12] C. Smith and J. Pokorny, Spectral sensitivity of the foveal cone photopigments between 400 and 500 nm, Vis. Res., vol. 15, no. 2, pp , [13] D. Travis, Effective Color Displays, vol. 4. San Francisco, CA, USA: Academic, 1991, pp [14] R. C. Gonzalez and R. E. Woods, Digital Image Processing. Beijing, China: Publishing House of Electronics Industry, Oscar C. Au received the B.A.Sc. degree from the University of Toronto, Toronto, ON, Canada, in 1986, the M.A. and Ph.D. degrees from Princeton University, Princeton, NJ, USA, in 1988 and 1991, respectively. After being a Postdoctoral Researcher with Princeton University for one year, he joined the Hong Kong University of Science and Technology (HKUST), Hong Kong, as an Assistant Professor, in He has been a Professor with the Department of Electronic and Computer Engineering, Director of Multimedia Technology Research Center, and the Director of the computer engineering program at HKUST. His current research interests include video and image coding and processing, watermarking and light weight encryption, and speech and audio processing. He has published over 50 technical journal papers, 320 conference papers, and 70 contributions to international standards. He was a Chair of Screen Content Coding AdHoc Group in the JCTVC for the ITU-T H.265 HEVC video coding standard. He has 18 granted U.S. patents and is applying for over 80 more on his signal processing techniques. He is a Board of Governor Member of the Asia Pacific Signal and Information Processing Association (APSIPA). He is/was an Associate Editors of IEEE TCSVT, IEEE TIP, and IEEE TCAS1. He is on the Editorial Boards of Journal of Visual Communication and Image Representation, Journal of Signal Processing Systems, APSIPA TSIP, JMM, and Journal of Franklin Institute. He has been Chair of IEEE CAS Technical Committee on Multimedia Systems and Applications (MSATC), Chair of SP TC on MMSP, and Chair of APSIPA TC on Image, Video and Multimedia (IVM). He is a member of CAS TC on Video Signal Processing and Communications (VSPC), CAS TC on Digital Signal Processing, SP TC on IVMSP, SP TC on Information Forensics and Security, and ComSoc TC on Multimedia Communications. He served on the Steering Committee of IEEE TMM, and IEEE ICME. He served on the organizing committee of ISCAS in 1997, IEEE ICASSP in 2003, the ISO/IEC MPEG 71st Meeting in 2005, IEEE ICIP in 2010, and other conferences. He was a General Chair of PCM in 2007, IEEE ICME in 2010, and the International Packet Video Workshop in He received the Best Paper Awards in SiPS 2007, PCM 2007, and MMSP He is an IEEE Distinguished Lecturer in 2009 and 2010, and has been keynote speaker for multiple times.

12 FANG et al.: LUMA-CHROMA SPACE FILTER DESIGN 3829 Ngai-Man Cheung is an Assistant Professor with the Singapore University of Technology and Design, Singapore. He received the Ph.D. degree in electrical engineering from the University of Southern California, Los Angeles, CA, USA, in From 2009 to 2011, he was a Post-Doctoral Researcher with the Image, Video and Multimedia Systems Group, Stanford University, Stanford, CA, USA. He has held research positions with Texas Instruments Research Center Japan, Nokia Research Center, IBM T. J. Watson Research Center, Yorktown Heights, NY, USA, HP Labs Japan, Hong Kong University of Science and Technology, Hong Kong, and Mitsubishi Electric Research Labs. His current research interests include image and video processing, signal processing, and multimedia communication. Aggelos K. Katsaggelos (F 98) received the Diploma degree in electrical and mechanical engineering from the Aristotelian University of Thessaloniki, Thessaloniki, Greece, in 1979, and the M.S. and Ph.D. degrees in electrical engineering from the Georgia Institute of Technology, Atlanta, GA, USA, in 1981 and 1985, respectively. In 1985, he joined the Department of Electrical Engineering and Computer Science, Northwestern University, Evanston, IL, USA, where he is currently a Professor of the AT&T chair. He was previously the holder of the Ameritech Chair of Information Technology from 1997 to He is the Director of the Motorola Center for Seamless Communications, a Academic Staff Member, NorthShore University Health System, an Affiliated Faculty with the Department of Linguistics and he has an appointment with the Argonne National Laboratory. He has published extensively in the areas of multimedia processing and communications (180 journal papers, over 400 conference papers, and 40 book chapters) and he holds 19 international patents. He is the co-author of Rate-Distortion Based Video Compression (Kluwer, 1997), Super-Resolution for Images and Video (Claypool, 2007), and Joint Source-Channel Video Transmission (Claypool, 2007). He was an Editor-in-Chief of the IEEE SIGNAL PROCESSING MAGAZINE from 1997 to 2002, a BOG Member of the IEEE Signal Processing Society from 1999 to 2001, and a member of the Publication Board of the IEEE Proceedings from 2003 to He is a fellow of the SPIE in 2009 and the recipient of the IEEE Third Millennium Medal in 2000, the IEEE Signal Processing Society Meritorious Service Award in 2001, the IEEE Signal Processing Society Technical Achievement Award in 2010, an IEEE Signal Processing Society Best Paper Award in 2001, an IEEE ICME Paper Award in 2006, an IEEE ICIP Paper Award in 2007, and an ISPA Paper Award in He was a Distinguished Lecturer of the IEEE Signal Processing Society from 2007 to Houqiang Li (M 10) received the B.S., M.Eng., and Ph.D. degrees from the University of Science and Technology of China (USTC), Hefei, China, in 1992, 1997, and 2000, respectively, all in electronic engineering. He is currently a Professor with the Department of Electronic Engineering and Information Science, USTC. His current research interests include video coding and communication, multimedia search, and image/video analysis. He has authored or co-authored over 100 papers in journals and conferences. He is an Associate Editor of IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY and in the Editorial Board of Journal of Multimedia. He has served on Technical/Program Committees, Organizing Committees, and as Program Cochair, track/session chair for over 10 international conferences. He was a recipient of the Best Paper Award for Visual Communications and Image Processing in 2012, the Best Paper Award for International Conference on Internet Multimedia Computing and Service in 2012, the Best Paper Award for the International Conference on Mobile and Ubiquitous Multimedia from ACM in Feng Zou received the B.S. degree from the Harbin Institute of Technology, Harbin, China, in 2004, and is now taking an internship with Mitsubishi Electric Research Labs, Cambridge, USA, while pursuing the Ph.D. degree in electronic and computer engineering with the Hong Kong University of Science and Technology, Hong Kong, supervised by Prof. O. Au. His current research interests include intra prediction, transform and quantization designs in video compression. He is actively contributing proposals in the standardization of HEVC under the ITU- T/ISO/IEC Joint Collaborative Team on Video Coding (JCT-VC) and JCT3V and holds several video coding related patents.