Application of Contrast Sensitivity Functions in Standard and High Dynamic Range Color Spaces
Total Page：16
File Type：pdf, Size：1020Kb
https://doi.org/10.2352/ISSN.2470-1173.2021.11.HVEI-153 This work is licensed under the Creative Commons Attribution 4.0 International License. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/. Color Threshold Functions: Application of Contrast Sensitivity Functions in Standard and High Dynamic Range Color Spaces Minjung Kim, Maryam Azimi, and Rafał K. Mantiuk Department of Computer Science and Technology, University of Cambridge Abstract vision science. However, it is not obvious as to how contrast Contrast sensitivity functions (CSFs) describe the smallest thresholds in DKL would translate to thresholds in other color visible contrast across a range of stimulus and viewing param- spaces across their color components due to non-linearities such eters. CSFs are useful for imaging and video applications, as as PQ encoding. contrast thresholds describe the maximum of color reproduction In this work, we adapt our spatio-chromatic CSF1 [12] to error that is invisible to the human observer. However, existing predict color threshold functions (CTFs). CTFs describe detec- CSFs are limited. First, they are typically only defined for achro- tion thresholds in color spaces that are more commonly used in matic contrast. Second, even when they are defined for chromatic imaging, video, and color science applications, such as sRGB contrast, the thresholds are described along the cardinal dimen- and YCbCr. The spatio-chromatic CSF model from [12] can sions of linear opponent color spaces, and therefore are difficult predict detection thresholds for any point in the color space and to relate to the dimensions of more commonly used color spaces, for any chromatic and achromatic modulation. In addition, the such as sRGB or CIE L∗a∗b∗. Here, we adapt a recently proposed CSF was fitted to the data that span 0.0002 cd/m2 (scotopic) to CSF to what we call color threshold functions (CTFs), which de- 10,000 cd/m2 (photopic), which makes it appropriate for predict- scribe thresholds for color differences in more commonly used ing thresholds in HDR color spaces. Using this CSF, we numeri- color spaces. We include color spaces with standard dynamic cally solve for detection thresholds in any non-linear color space. ∗ ∗ ∗ ∗ ∗ ∗ range gamut (sRGB, YCbCr, CIE L a b , CIE L u v ) and high Our work lends insight into the coding efficiency and the unifor- dynamic range gamut (PQ-RGB, PQ-YCbCr and ICTCP). Using mity of contrast thresholds in different color spaces. CTFs, we analyze these color spaces in terms of coding efficiency and contrast threshold uniformity. A Device-Independent CSF The spatio-chromatic CSF from [12] is the basis for the CTFs Introduction and Background presented in this work. The CSF was developed to account for Contrast thresholds describe the minimum difference in lu- contrast threshold measurements from 0.125 cycles per degree minance or chromaticity that a human observer can detect. Con- (cpd) to 32 cpd, from 0.0002 cd/m2 to 10,000 cd/m2, and for dif- trast thresholds vary with image and viewing parameters, such as ferent hues [15, 16, 14, 13, 17, 18]. A critical feature of this CSF spatial frequency [1], luminance [2], color [3], and the size of the is that it can accurately describe contrast thresholds for any lumi- stimulus [4]. The dependence of contrast thresholds to such pa- nance and chromaticity combination, meaning we can map con- rameters is described by contrast sensitivity functions (CSFs). trast thresholds from the native color space of the CSF onto color Having an accurate CSF is important for image and video coordinates of other color spaces. coding applications, as contrast thresholds provide a limit to the In Fig. 1, we show the CSF in DKL color space, a linear amount of color reproduction error that is noticeable to a human transformation of LMS color space: observer. For example, CSFs were used to create transfer func- tions for encoding high dynamic range color values [?, 6] in a perceptually uniform manner. In particular, Perceptual Quantizer 2 32 3 1 1 0 DL (PQ) is a non-linear function based on Barten’s CSF [5], used to 6 7 6 L0 76 7 code High Dynamic Range (HDR) content [6], akin to gamma DDKL = 6 1 − 0 76DM7 (1) encoding used in Standard Dynamic Range (SDR) [7]. For fixed 6 M0 76 7 4 L0 + M0 54 5 bit-depth, PQ assigns more code-words to luminance levels with −1 −1 DS S lower thresholds and fewer code-words to levels with high thresh- 0 olds, thus maximizing visual quality while efficiently allocating (L ;M ;S ) is the white point in LMS color space and code-words. 0 0 0 (DL;DM;DS) is the color coordinate with respect to the white However, many existing CSFs only describe achromatic con- point. For the CSFs in Fig. 1 and the CTFs derived from them, we trast, neglecting the detection thresholds in chromatic directions. used CIE 2006 cone fundamentals [19] and assumed D65 white This is true even for Barten’s CSF, meaning that PQ is only appro- point. DDKL is the color coordinate in DKL, representing po- priate for coding luminance, even though the industry standard is sition along achromatic, red-green, and yellow-violet axes. The to use PQ for all color channels [8, 21]. In addition, existing chro- matic CSFs [9, 10] are usually reported in Derrington-Krauskopf- 1The code and more details on the spatio-chromatic CSF can be found Lennie (DKL) color space [11], a linear color opponent space that at https://www.cl.cam.ac.uk/research/rainbow/projects/ is physiologically relevant but has limited application outside of hdr-csf/. IS&T International Symposium on Electronic Imaging 2021 Human Vision and Electronic Imaging 2021 153-1 Achromatic Achromatic ment Dc that yields the detection threshold. For example, con- 1 1 1 1 sider sRGB. We start at (RD65;GD65;BD65) and search for DR such that (RD65 + DR;GD65;BD65) results in the Pdet = 0:5, as- 0.1 10 0.1 10 suming a guess rate of 0. To obtain Pdet, we convert from sRGB to LMS and query the CSF. 0.01 100 0.01 100 Since Dc is an increment in the non-linear target color space, not a contrast, we refer to our thresholds increment thresholds or 1e-3 1e3 1e-3 1e3 color increments thresholds, and the functions as Contrast Thresh- Red-Green Red-Green c > 1 1 Sensitivity (1 / cone contrast) 1 1 Sensitivity (1 / cone contrast) old Functions (CTFs). We show increment thresholds (D 0), but not decrement thresholds (Dc < 0). This is because the under- lying CSF has comparable positive and negative contrast thresh- 0.1 10 0.1 10 olds, resulting in similar increment and decrement thresholds in the target color space. 0.01 100 0.01 100 (cone contrast) (cone contrast) Below, we report increment thresholds for RGB (SDR and ∗ ∗ ∗ ∗ ∗ ∗ 1e-3 1e3 1e-3 1e3 HDR), YCbCr (SDR and HDR), CIE L a b and CIE L u v (SDR only), and IC C (HDR only). For SDR, we used the Yellow-Violet Yellow-Violet T P 1 1 1 1 Threshold Threshold sRGB non-linearity on ITU-R BT.709 color primaries [20], as- suming a display luminance range of 0.1 cd/m2 to 100 cd/m2. For 0.1 10 0.1 10 HDR, we used PQ as the transfer function on ITU-R BT.2020 color primaries [21], assuming a display luminance range be- 2 2 0.01 100 0.01 100 tween 0.005 cd/m and 10,000 cd/m . The values of the increment thresholds between SDR and HDR cannot be directly compared 1e-3 1e3 1e-3 1e3 because the SDR color gamut is smaller than that of HDR, mean- 0.25 1 4 16 0.01 1 100 10k ing an increment threshold of 0.1 corresponds to different amount 0.125Frequency0.5 2 (cpd)8 32 0.001Luminance0.1 10 (cd/m1k 2) of physical change for SDR and HDR. HDR In addition, all plots show the maximum quantization error 0.001 cd/m2 10000 cd/m2 0.125 cpd 4 cpd assuming different bit-depths for each color space (horizontal dot- 0.01 cd/m2 1000 cd/m2 0.25 cpd 8 cpd ted lines). The maximum quantization error is calculated as: SDR 0.5 cpd 16 cpd 0.1 cd/m2 100 cd/m2 1 cpd 32 cpd 2 2 (︃ )︃ 1 cd/m 10 cd/m 2 cpd vmax − vmin maxeq = 0:5 ; (3) Figure 1: Spatio-chromatic contrast threshold functions (CSFs) 2b − 1 along the cardinal directions of the DKL color space [12, 13]. Left column: CSFs as a function of frequency. Right column: where b is the bit-depth, and vmax and vmin are the maximum and CSFs as a function of luminance. Left y-axis: Contrast threshold. minimum values to be encoded, respectively. Note that vmin can Right y-axis: Contrast sensitivity. be negative for some color channels (e.g., a∗ of CIE L∗a∗b∗). When the error is above the threshold, quantization artifacts are contrast thresholds are expressed using cone contrast, a vector likely visible; when the error is below, the artifacts are likely in- length in LMS space: visible, however code-words are wasted. v RGB u "(︃ )︃2 (︃ )︃2 (︃ )︃2# u 1 DL DM DS We show CTFs in RGB with sRGB non-linearity (Fig. 2), Ccone = t + + (2) 3 L0 M0 S0 PQ encoding (PQ-RGB; Fig. 3), and without any non-linearities at HDR luminance levels (linear RGB; Fig.