Received: 29 June 2017 Revised: 7 September 2018 Accepted: 9 September 2018 DOI: 10.1002/col.22291

RESEARCH REVIEW

Review and evaluation of color spaces for image/video compression

Samruddhi Y. Kahu1 | Rajesh B. Raut2 | Kishor M. Bhurchandi1

1Visvesvaraya National Institute of Technology, Nagpur, India Abstract 2Shri Ramdeobaba College of Engineering and A color space plays an important role in color image processing and color vision Management, Nagpur, India applications. While compressing images/videos, properties of the human visual Correspondence system are used to remove image details unperceivable by the human eye, appro- Samruddhi Y. Kahu, Visvesvaraya National priately called psychovisual redundancies. This is where the effect of the color Institute of Technology, South Ambazari Road, Nagpur, 440010, India. spaces' properties on compression efficiency is introduced. In this work, we study Email: [email protected] the suitability of various color spaces for compression of images and videos. This review work is undertaken in two stages. Initially, a comprehensive review of the published color spaces is done. These color spaces are classified and their advan- tages, limitations, and applications are also highlighted. Next, the color spaces are quantitatively analyzed and benchmarked in the perspective of image and video compression algorithms, to identify and evaluate crucial color space parameters for image and video compression algorithms.

KEYWORDS CIE L*a*b*, color spaces, color spaces' review and benchmarking, image and video compression, imaging, quantitative analysis

1 | INTRODUCTION approximately represent the wavelengths of red, green, and blue colors.2,3 A complex color “Ci” perceived as spatial inte- As rightly said, “An image is worth thousand words,” and as gration of incident wavelengths on the three cone sensors by known a video is a long sequence of images. Images and human brain is given by Equation (1)1; videos are used to capture a scene and display it at any later λmaxð instant and place, may be repeatedly, without loss of percep- ¼ ðÞλ ðÞλ λ ¼ ð Þ tual information. Colors are important components of any Ci Si f d i 1,2,3, 1 λ scene. Thus, for efficient and effective capture and display min of color images, a thorough study of the human visual sys- where Si(λ) denotes the sensitivity of the ith type of cones tem (HVS) was carried out and a few color spaces like Red that is, S, M, and L, f(λ) = S(λ), M(λ) or L(λ), and [λmin, 1 1 Green Blue (RGB) and Hue Saturation Intensity (HSI) λmax] denotes the interval of wavelengths outside which the mimicking it were devised. sensitivities are zero. Typically in air or vacuum, visible Human eye contains light sensors spatially and nonuni- region of the electromagnetic spectrum is specified by wave- 3 formly distributed all over the retina. There are mainly two length region between λmin = 360 nm and λmax = 830 nm. types of light sensors; (1) “Rods”—responsible for perception However, some sources state that the effective range is from of gray level or achromatic intensities of light primarily under 400 nm to 700 nm.4 Responses of the S, M, and L cones in very low luminance levels. At very low intensities of light, no this wavelength region are presented in Figure 1. colors can be perceived as only shades of gray are seen. The three primary/fundamental color theory, also known (2) “Cones”—responsible for perception of different colors, as the “trichromatic theory of color vision” was developed are of three types; namely, S, M, and L sensitive to short, based on the works of Maxwell, Young, and Helmholtz.2,5 medium and long wavelengths of light. These wavelengths They proposed the presence of three receptors in the human

8 © 2018 Wiley Periodicals, Inc. wileyonlinelibrary.com/journal/col Color Res Appl. 2019;44:8–33. KAHU ET AL. 9

visual contents using a color space suitable for minimizing psychovisual redundancies. Thus, selection of a color space S M L 1 is a crucial issue for designing compression algorithms. Figure 2 shows flowchart for a generalized compression 0.8 algorithm. Color space conversion is the first step as seen in Figure 2. 0.6 Today, YCbCr color space is widely used by most of the state-of-the-art compression algorithms such as JPEG,9,10 0.4 JPEG 2000,11 H.264,12 HEVC,13 and so forth, as it gives Responsivity high decorrelation. As HVS is more sensitive to luminance, 0.2 separation of luminance, and chrominance components allows use of chrominance subsampling14 resulting in high 0 compression ratios (CR). Most of the research in this field is 400 450 500 550 600 Wavelength (nm) done using the above mentioned decorrelation-based approach, leading to the widespread use of YCbCr. How- 2 FIGURE 1 Responsivity of LMS cones in human retina ever, a small number of compression algorithms have also come up which use the correlation between R, G, and retina approximately sensitive to red, green, and blue wave- – B components effectively.15 17 Hence, these algorithms are lengths of the visible light. Some studies even claim that the called correlation-based approaches. Analysis of the advan- theory of three fundamental colors was developed by tages and disadvantages of the correlation and decorrelation- Thomas Young based on the work of an English glassmaker based approaches is presented by Gershikov and Porat.18 A G. Palmer.5 Thomas Young neglected the fact that fourth few color spaces have also been designed especially for type of sensors are also available in the HVS that are respon- compression algorithms.19,20 sible for gray vision and also affect the saturation of a com- This work initially reviews all the major published color plex color. Accordingly, many other systems use four – spaces, their advantages, limitations, and applications. The components to represent a complete color space.6 8 Hering, main objective of this article is to analyze their suitability for in mid-1870s proposed the idea of four primary colors, compression algorithms. We have also identified and com- namely, yellow, red, blue, and green. This is called Hering's puted nine parameters for benchmarking different color “opponent-colors theory of vision”, which is based on the spaces for compression algorithms. Thus, 38 color spaces observation that certain colors were never perceived to occur are surveyed, analyzed, and benchmarked using the parame- together. He also proposed the presence of three types of receptors in the human eye with bipolar responses to light- ters to establish their suitability for the said purpose. This dark, red-green, and yellow-blue colors.2,5 Aubert first rec- study helps us find the reason behind the widespread use of ognized that Hering's opponent color theory and Young, YCbCr space for compression. Although YCbCr color space Helmholtz, and Maxwell's primary color theory are not con- is being used conventionally for compression, other spaces tradictory.5 Based on the stage theory developed by Jameson may be more suitable due to their properties like linearity and Hurvich, it was established that there are indeed three and perceptual uniformity. Thus, survey of the published types of trichromatic receptors in the human eye as proposed color spaces and their quantitative analysis for suitability for by Young, Maxwell, and Helmholtz. However, the output of compression algorithms are the two goals of this work. the three receptors is not sent as it is to the human brain. It is Rest of the article is organized as follows. Section 2 pre- used to form the opponent red-green and blue-yellow signals sents classification of different color spaces and different as proposed by Hering.2,3 factors affecting their performance in compression applica- In the due course of time, different color spaces were tions. Section 3 reviews all the major published color spaces, introduced for several applications like, color printing and their features, advantages, limitations, and applications. display, image and video transmission, and storage, etc. In Important parameters for quantitative analysis of the color the recent mobile communication and computing revolution, space have been identified and defined in section 4. Quanti- multimedia handling, transmission, and display have gained tative analysis of all these color spaces in terms of the identi- popularity and subsequent prime importance. Obviously, fied parameters has been presented in section 5. Section 6 image and video compression algorithms need to express the concludes the article.

FIGURE 2 Generalized compression scheme 10 KAHU ET AL.

FIGURE 3 Classification of color spaces

2 | FACTORS AFFECTING complex color. New family of calibrated color spaces intro- CLASSIFICATION OF COLOR SPACES duced for such applications is called device independent color spaces.21 Many color spaces have been introduced so far. On the basis of their definition, we classify them in three categories: (1) Fundamental (2) Derived, and (3) Application specific. 3 | COLOR SPACES' REVIEW Most color spaces use three components to represent a complex color. A few spaces can describe a complex color Color spaces can be classified in many ways as already dis- in terms of a chromaticity diagram using only two compo- cussed in section 2. However, to study their attribute impor- nents. Third component can be expressed as one minus addi- tance and characteristics relevant to compression, we group tion of the other two components. A few other color spaces the different color spaces based on the purpose for which use more than three attributes to describe complex colors. they were originally proposed as shown in Figure 3. Thus, Thus, color spaces can also be categorized in two categories; the color spaces are broadly classified into sensor based, dis- (1) Euclidian color spaces with three or less attributes and play based, color matching based, color equivalency based, (2) Riemannian color spaces with more than three attributes. color appearance modeling based and color order systems. Since Riemannian color spaces have higher dimensionality, Display-based color spaces are further classified as Additive, Euclidean color spaces are more suitable for compression Subtractive, and the color spaces which are neither additive algorithms. nor subtractive are grouped under RGB derivatives category. In most of the color spaces, perceptual difference RGB derivatives are the color spaces derived from display- between two complex colors is not proportional to the based (gamma-corrected) RGB space. Since, we are catego- Euclidian distance between them. This is a nonlinearity in rizing color spaces based on the original purpose for which color spaces. Also in many color spaces, a normalized per- they were proposed, we group the color spaces devised for various image processing applications such as segmentation, centage intensity change in all the basic color attributes does compression, and so forth, under RGB derivatives category. not lead to equal perceptual change in a complex color. This defines the nonuniformity of color spaces.1 It should be noted here that nonlinearity and nonuniformity are closely 3.1 | Sensor-based color spaces related. Based on this, color spaces can be classified as; Image data captured by color input devices such as cameras (1) Nonlinear color spaces and (2) Linear color spaces. and scanners generally consists of three channels, namely, Quantization is an important step which directly controls the red (R), green (G), and blue (B). The captured RGB values achieved Peak Signal to Noise Ratio (PSNR) and are proportional to the spectral sensitivities of the color sen- CR. Because quantization error in a linear color space is sors used in these devices. These RGB values are not identi- equal to the perceived color difference at any intensity value, cal to the color matching functions of the HVS.22 Therefore, it is considered more effective than a nonlinear color space these values are transformed from the sensor's spectral space from the quantization point of view. In video compression, to the CIE XYZ color space. The transformation is achieved frame difference is thresholded to remove unperceivable dif- using capture colorimetry model that is unique for different ferences during quantization process. In a nonlinear color image capturing devices.3 The obtained device independent space, a scalar threshold may not yield desirable results. CIE XYZ values or Reference Input Medium Metric (RIMM) In case of generation of colors, it has been observed that RGB values may be used to represent the sensor RGB same color attributes of a complex color produce visibly dif- values. RIMM RGB values are obtained using a color correc- ferent colors on different electronic displays, color printers, tion matrix and the XYZ tristimulus values as shown in and printing surfaces. Thus, need of device independent Ref. 3. color spaces1,21 was felt especially for display and printing RIMM RGB is an unrendered/scene-referred or more applications. In case of such device independent color appropriately, a sensor-based color space, which is not spaces, actual color attributes of a complex color are cali- designed for color output/display devices. Further color trans- brated with the respective color channel transfer functions of formations, such as tone scale/color rendering, gamma cor- the device and calibrated attributes are used to generate a rection, and so forth, are by default applied in digital cameras KAHU ET AL. 11

TABLE 1 Additive color spaces

Sr. no. Color spaces Fundamental/derived Perceptually linear/nonlinear 1 RGB/sRGB Fundamental Nonlinear 2 rgb Derived Nonlinear 3 Adobe RGB Derived Nonlinear 4 Adobe wide gamut RGB Derived Nonlinear 5 ProPhoto RGB Derived Nonlinear to obtain display based RGB values from these RIMM RGB so forth. These spaces are called “additive” as different or XYZ tristimulus values. These transformations are colors can be formed by a combination/mixture of the pri- described in detail in Refs. 3 and 22 Thus, RGB images maries of additive color spaces in various proportions.3 A obtained using digital cameras are by default in the display combination of all the colors results in white. Table 1 shows based RGB space, unless they are raw camera images. some additive color spaces. Because we are evaluating color spaces from the per- spective of image/video compression, we consider only the RGB color space display-based color spaces described further in this article. RGB is a hardware friendly color space due to the believed Therefore, we do not delve deeper into this topic. orthogonality of R, G, and B components. However, its three color planes are highly correlated making it unsuitable for 26 3.2 | Display-based color spaces compression algorithms due to more redundancies. Cross correlation between B and R channels (TBR) is about 0.78, Almost all the digital color output systems like TV, com- – similarly T is 0.98, and T its 0.94.27 29 These high cor- puters, printers, and so forth, use RGB color space for dis- RG GB relation values are not desirable for decorrelation based com- playing images or videos. To ensure interoperability and pression approaches. Its psychological nonintuitivity is device independence of the RGB color space, a standard another problem as human beings can't estimate the R, G, RGB color space, represented by sRGB, is introduced coop- and B attributes by visualizing a complex color. Other prob- eratively by HP and Microsoft for use on monitors, printers, lems of this space include perceptual nonlinearity and and internet.21,23–25 Thus sRGB can be used for exchange of nonuniformity. color information between systems directly. This saves the Still, most of the color space literature is based on the computational effort required for conversion to device inde- – RGB color space.30 39 Huang and Wang32 presented color pendent CIE color spaces for transmission/storage of color facial image compression using a two-stage vector quantiza- information.3 As discussed in section 3.1, the RGB pixel tion in RGB space. Kim et al. presented an efficient video values in digital images should be interpreted as sRGB (non- coding method using interplane prediction that increases linear RGB) unless otherwise stated. The same convention is coding efficiency in the RGB space while avoiding color followed in this work too. In general, (nonprimed) R, G, and distortion.27 B should be interpreted as display based (nonlinear, sRGB) RGB values in machine vision are different from those in RGB values. For conversion to certain color spaces such as human perception because human eyes distinguish objects CIE XYZ and PhotoYCC, linear RGB values represented by 0 with hue, saturation, and intensity of colors as modeled in R’, G’, and B (primed) are required. Linear RGB is obtained 6 HSI space. Therefore, RGB space is not effective for per- by inverting the gamma correction step applied in digital forming color comparisons in natural unconstrained environ- cameras. Conversion from sRGB to linear RGB is achieved 24,25 ment and for machine vision applications where using Equation (2) ; performance is expected to match human vision. Image clas- ( sification and emotion recognition tasks using psychological 0 12:92 × R if R ≤ 0:000304 ¼ : ð Þ R 1 2 and artistic aspects of color images are examples of such 1:055 × R2:4 −0:055 otherwise applications.40,41 However, many applications that need Similar conversion is applied to obtain G and B values human interface or mimic vision capabilities still use RGB from G’ and B0. space. Conversion to/from RGB coordinates from/to any Display-based color spaces are further classified as either other system for processing algorithms and display puts an additive, subtractive, or other spaces derived from RGB undue overhead on the computing machines. Thus, proces- which are neither additive nor subtractive as shown in sing in RGB space will save this overhead but results in the Figure 3. mentioned problems.

3.2.1 | Additive color spaces Normalized RGB color space Additive color spaces are generally used to produce color on The three R, G, and B components are very sensitive to light- a dark background such as CRT monitors, TV screens, and ness and shading.6,42–45 To overcome this problem, the 12 KAHU ET AL.

TABLE 2 Color spaces derived from RGB (RGB derivatives)

Sr. no. Color spaces Fundamental/derived Perceptually linear/nonlinear 1 YUV Derived Nonlinear 2 YIQ Derived Nonlinear 3 YCbCr Derived Nonlinear 4 YCgCr Derived Nonlinear 5 YCgCb Derived Nonlinear 6 YPbPr Derived Nonlinear 7 YDbDr Derived Nonlinear 8 YES Derived Nonlinear

9 I1I2I3 Derived Nonlinear

10 YT1T2 Derived Nonlinear

11 h1h2h3/H1H2H3 Derived Nonlinear 12 Photo YCC Derived Nonlinear 13 Kekre's LUV Derived Linear 14 YSbSr Derived Nonlinear 15 cGST Derived Nonlinear standard RGB color space is usually converted into normal- When working in color spaces with a large gamut if the ized RGB space; rgb.7,8 Normalization of RGB space was gradient steps are large, posterization effects will occur more first performed in by Zhai et al.42 and is defined in frequently in 8-bit/channel resolution. To avoid this, it is Equation (3).39,43,46 recommended to use 16-bit/channel color depth, making 9 these spaces unsuitable for compression. R > r ¼ > R + G + B > => ProPhoto RGB ¼ G : ð Þ ProPhoto RGB color space, also known as ROMM RGB,49 g > 3 R + G + B> > is an output referred RGB color space, developed by Kodak B > b ¼ ; to offer an especially large gamut designed for use with pho- R + G + B tographic output. The ProPhoto RGB color space encom- The main advantage of this normalized color space is passes over 90% of possible surface colors defined in the that individual color components are robust to brightness CIE L*a*b* color space, and 100% of likely occurring real variation in an image. Also, as r + g + b = 1, only two world surface colors. This makes ProPhoto gamut larger components need to be coded. Third component can then be than all other color spaces including the Adobe Wide Gamut determined by subtracting sum of the two from one.43 The RGB space. One of the limitations of this color space is that, disadvantage of RGB color space is that it is very noisy at approximately 13% of the representable colors are imagi- low intensities due to the nonlinear transformation in the nary, i.e., that do not exist and are not visible. This results in range [0, 1].43 This also makes it vulnerable to quantization potential color accuracy being wasted for accommodating and rounding errors, which are inevitable during compres- these unnecessary colors. This wide gamut space is also sion, and decompression. Also, luminance information is not unsuitable for compression due to the required 16-bit color available in the RGB color space; this makes it further unsui- depth to avoid posterization effects. In addition to 16 bits table for compression. per attribute, they also have the disadvantages of high inter- color correlation and perceptual nonlinearity, which again makes them unsuitable for compression. Adobe RGB and adobe wide gamut RGB color spaces The published literature on RGB reveals that almost all 47 Both Adobe RGB and Adobe Wide Gamut RGB color digital devices generate, display, and handle images and 48 spaces are RGB spaces developed by Adobe Systems for videos in RGB format. Thus, processing of images and their commercial interests. Adobe RGB was designed to videos in any other color space requires conversion to/from encompass most of the colors achievable on CMYK printers the new space resulting in increased computations. Many using RGB primary colors on a device such as a monitor. times, RGB derived color spaces such as the normalized Similarly, Adobe Wide Gamut RGB also offers a large gamut RGB space are used for face detection, edge, and texture using pure spectral primary colors.48 A comparison shows detection, where illumination in the image must be normal- the Wide Gamut RGB color space encompasses 77.6% of the ized before processing. However, due to sensitivity to noise visible colors specified by CIE, while the standard Adobe and quantization errors, it is unsuitable for compression. RGB space covers just 50.6% and sRGB only 35.0%. Various RGB derived spaces such as ProPhoto RGB, Adobe KAHU ET AL. 13

RGB, and Wide Gamut RGB are being used commercially 3.2.3 | RGB derivatives for photographic applications. Along with other limitations, Other color spaces derived from RGB which are not used for they require more than 8 bits per attribute; making them either display or printing applications are grouped under this unsuitable for compression. category. Color spaces in this subcategory are obtained by transformation of the display based (ie, gamma-corrected) 3.2.2 | Subtractive color spaces RGB. The transformation is applied to fulfill requirements of Unlike additive color spaces, subtractive color spaces are specific application(s). Their features are presented in used to produce color on transparent or reflective media, that Table 1 followed by their analysis in this section. is, white backgrounds. Subtractive color spaces are used in output devices such as color printers. Different colors are YUV color space produced by these spaces by putting colorants (ink pigments Contemporary CRT devices used in TV boxes and computer in case of printers) on output media such as white paper. monitors comply with the ITU-R recommendation BT.709. These colorants absorb light of certain wavelength while This is required because, if we drive multiple CRT devices transmitting other wavelengths. In other words, color is pro- with the same RGB values, we should yield perceptually duced by removing unwanted spectral components from same colors. As a result of various historical reasons, the 3,50 white light. first TV systems transmitted only the luminance component. A typical example of a subtractive color space is CMY Later, it was decided to add two chrominance components 50,51 color space. CMY color space is quite unintuitive and R-Y and B-Y for color transmission. This system was perceptually nonlinear. The three components, C, M, and Y designed to minimize bandwidth requirement of the compos- represent three reflection filters which absorb red, green, and ite signals. Since HVS is less sensitive to chrominance, chro- 3 blue wavelengths, respectively. There is also a CMYK color minance components can be transmitted using smaller space where the fourth component K represents black color. bandwidth. U and V always have the same bandwidth. CRT There are two types of transformations to the CMY(K) color gamma correction is assumed to be 2.8, but camera correc- space: (1) one minus RGB (2) complicated polynomial arith- tion laws are same as in all other systems (~0.45). metic or 3D interpolations of lookup tables. The first option Conversion from RGB to YUV space is shown in is just a rudimentary conversion between RGB and CMY. Equation (5),43,53 More complex conversions as suggested by the second type 2 3 2 32 3 of transformation are required as the C, M, and Y colorants Y 0:299 0:587 0:114 R do not exactly absorb red, green, and blue lights, respec- 6 7 6 76 7 4U 5 ¼ 4 −0:147 −0:288 0:436 54G5: ð5Þ tively.50 These complex conversions are used in practical : − : − : applications. V 0 615 0 515 0 100 B Conversion from RGB to CMY using first type of trans- Here, R, G, B ϵ [0, 1]. formation is shown in Equation (4). Conversion from RGB The advantage of this color space is; it can be used to effi- to CMY (RGB values from 0 to 255) results in values from ciently encode color information in the TV signals of 0to1.CMY values are further converted to CMYK as pro- European system. It partly gets rid of the correlation between posed in Refs. 50 and 51. 9 RGB components and involves less computation time. YUV C ¼ 1−R=255 > => color space separates luminance and chrominance compo- ¼ − = : ð Þ nents of an image which is advantageous for decorrelation M 1 G 255> 4 ;> based compression approaches. Most color compression sys- ¼ − = Y 1 B 255 tems reduce the redundancies between the RGB components Conventional printing devices with three or more inks by transforming the color primaries into a decorrelated color are termed as multicolor printing devices. Besides C, M, space, such as YIQ or YUV.54 An embedded wavelet coding Y inks, additional inks provided by multicolor printers are algorithm for color image and video compression in YUV often black (K) and/or some light inks like light cyan (Lc), color space is presented by Yuan and Mandal.55 Major draw- light magenta (Lm),52 and so forth. Theoretically, black color back of YUV color space is that correlation still exists due to can be synthesized by C, M, Y inks. But in practice the syn- the linear transformation, though not as high as RGB.43 thesized black color is not satisfactory and also synthesizing Other applications of YUV color space include a new black color using color cartridges is expensive. CMY color YUV color space which has been proposed to improve color space is very useful for printing on white background. For classification and segmentation speed.56 YUV color space is printing on other backgrounds, accurate background sensing rather effective for extraction of color features for a content- and color calibration is required. This space has all the disad- based image retrieval method based on DWT.57 Method of vantages of RGB space and hence, is unsuitable for hiding messages in images based on YUV and its derivatives compression. is also presented.58 14 KAHU ET AL.

YIQ color space YCgCr and YCgCb color space For standardization, calibration, and minimization of band- New color spaces YCgCr and YCgCb were developed for face width, American National Television Systems Committee recognition,65,66 segmentation and analysis applications. The (NTSC, USA) presented the basic YIQ system. In this sys- color spaces YCgCr and YCgCb are based on YCbCr but differ tem, brightness is represented by Y, while chrominance com- on the use of Cg component instead of Cb and Cr components, ponents are I and Q. I and Q carry color information along respectively. The color spaces used in television systems (YIQ, with some luminance and are derived by rotating the UV YCbCr) are transmission oriented. So, to minimize encoding- vector by 33. Y usually has a 4.2 MHz bandwidth in a decoding errors, they use the biggest color differences, (R-Y) 525 line system. Originally, I and Q signals were to have dif- and (B-Y). However, YCgCr and YCgCb spaces use the smal- ferent bandwidths (0.5 and 1.5 MHz) but they now have the lest color differences. This makes the YCgCr and YCgCb color same bandwidth (1 MHz). The Y, I, and Q components are spaces unsuitable for compression applications. Conversion for- defined in Equation (6),43,51 mulae for YCgCr and YCgCb spaces from RGB are defined on 2 3 2 3 2 3 similar lines as that of YCbCr.43,67,68 Fast face segmentation is : : : Y 0 299 0 587 0 114 R also implemented using YCgCr color space.67 4I 5 ¼ 40:595 −0:274 −0:3215 4G5: ð6Þ Q 0:211 −0:522 0:311 B YPbPr color space YIQ space partly gets rid of the correlation between R, YPbPr color space is one of the developments of the NTSC 43 G, and B planes. However, YIQ color space is not image component coding, in which the B primary and white point dependent and as efficient as the K1K2K3 color space in dec- were changed. The conversion equations between RGB and 26 orrelating its color components. This makes it an important YPbPr color spaces are presented in Equation (8),21 choice for compression algorithms. YIQ color space is non- 2 3 2 3 2 3 : : : linear and perceptually nonuniform. A Vector Quantization Y 0 212 0 701 0 0865 R 4Pb5 ¼ 4 −0:116 −0:383 −0:500 5 4G5: ð8Þ (VQ) technique based on YIQ reduces the frame buffer use Pr 0:500 −0:445 −0:055 B of a color image by half compared to the current algorithms with comparable image quality.59 Y component of YIQ color space is also good for edge detection.43 YDbDr color space YDbDr is a color space used in the SÉCAM color television standard, used mainly in France. YDbDr has three YCbCr color space components—Y, Db, and Dr. Y is the luminance, Db (blue YCbCr color space is an international standard for digital color difference), and Dr (red color difference) are the chro- coding of TV pictures for 525 and 625 lines. It is indepen- minance components. Conversion between YDbDr and RGB dent of the scanning standard and system primaries. There- color spaces is presented in Equation (9),43,69 fore, there are no chromaticity coordinates, no CIE XYZ 2 3 2 3 2 3 Y 0:299 0:587 0:114 R matrices, and no assumptions about white point or CRT 4Db5 ¼ 4 −0:450 −0:883 1:133 5 4G5: ð9Þ gamma. It provides digital representation of RGB signals in Dr −1:333 1:116 0:217 B YCbCr form. Linear conversion from RGB to YCbCr is given in Equation (7),54 YES color space 2 3 2 32 3 2 3 Y 65:48 128:55 24:97 R 16 YES color space is another luminance-chrominance color 70 4Cb5 ¼ 4 −37:78 −74:16 111:93 54G5 + 41285: space proposed by Saber and Tekalp for face detection. In Cr 111:96 −93:75 −18:21 B 128 this color space, Y represents the luminance component and ð7Þ E and S denote the chrominance components. The luminance component (Y) is a weighted sum of the RGB values, while R, G, B ϵ Here [0, 1]. the chrominance factors are spectral differences. E is propor- YCbCr color space is widely used for image and video tional to the difference of the red and green color channels, compression. Almost all the existing compression standards while S is proportional to difference of yellow and blue. The such as JPEG,10 JPEG2000,11 H.264,12 HEVC,13 and so conversion from RGB to YES is presented in Equation (10), forth, use YCbCr color space. Apart from compression, 2 3 2 32 3 YCbCr space has been used for many other applications. Y 0:253 0:684 0:063 R 6 7 6 76 7 The Cr and Cb components of YCbCr have been used for 6 7 6 76 7 4E 5 ¼ 40:500 −0:500 0:000 54G5: ð10Þ skin color extraction and skin-pixel classification.60–62 Noda, : : − : H. presented a simple and efficient colorization method in S 0 250 0 250 0 500 B YCbCr space that requires only about one fourth of computa- In other researchers used YES color space for face tion time in RGB space.63,64 detection.71,72 YES space was chosen by the author KAHU ET AL. 15 9 because of the following reasons: (1) The luminance h1 ¼ R−G=> component (Y) is independent of chrominance compo- ¼ − : ð Þ h2 G B> 13 nents (E and S). (2) It is computationally efficient and ; h3 ¼ B−R (3) its attributes result in minimum overlap between skin and nonskin data. YES color space has never been A comprehensive color edge detection comparison in a used for compression. However, being a luminance- variety of color spaces including h1h2h3 is presented in chrominance space, we analyze it in section 5. Refs. 42 and 43. In spite of the widespread use of h1h2h3 color space for edge detection,75,78 it cannot be used for

I1I2I3 or K1K2K3 color space compression as it is a noninvertible space. This fact is

A new color space called I1I2I3 color space or the Ohta color proven by Equation (14), which shows that RGB to h1h2h3 space was proposed to overcome the shortcomings of the transformation matrix is singular. 73 RGB color space. To decorrelate the RGB components, 1 −10 Ohta applied K-L transformation to the RGB space and 01−1 ¼ 0: ð14Þ derived the new color space named, I1I2I3 or a K-L transform −10 1 43,51 space K1K2K3, finally represented by Equation (11). 79 Braquelaire proposed another color space H1H2H3 in 2 3 2 32 3 Equation (15). : : : K1 0 333 0 333 0 333 R 9 4 5 ¼ 4 : − : 54 5: ð Þ K2 0 500 0 0 500 G 11 H1 ¼ R + G => − : : − : K3 0 2500 0 500 0 2500 B H ¼ R−G : ð15Þ 2 ;> H3 ¼ B−ðÞR + G =2 It may be noted that the first component of this space is This space was used for top-down quantization of color intensity, which indeed is an important feature. The other images. This method splits each color cluster along the axis two axes are statistically uncorrelated with intensity due to H , H , or H of greatest variance as in Equation (15). This properties of the K-L transform. Correlation of RGB compo- 1 2 3 model is derived from Faugeras theory of human vision, nents is partially lowered down by this color space, which is which incorporates the widely established notions of tri- important for success of a decorrelation based compression 80,81 algorithm. This color space is computationally less complex, chromaticity and color opponency theory. Achromatic and is also useful for various image processing applications, channel, H1 corresponds to the sum of long and medium such as edge detection and texture classification.43,74–76 wavelengths, that is, R and G. The two other channels are Quantitative analysis of this color space from compression called the red-green and blue-yellow channels respectively, perspective is presented in section 5. approximated by H2 and H3. This color space is further quantitatively analyzed in detail from the perspective of

YT1T2 color space compression algorithms in section 5.

YT1T2 is another normalized color space derived from Photo YCC color space RGB.43,77 Y component of this color space is presented in Equation (12). This color space was defined by Kodak in 1992 for the stor- age of digital color images on Photo CDs. Transformation ¼ ð Þ Y c1R + c2G + c3G, 12 from RGB to Photo YCC is done in three steps: (1) Gamma where, c1, c2, and c3 are constants and are linearly combined correction from R'G'B' values to RGB. (2) Linear transforma- with R, G, and B to produce luminance (Y) component. T1 tion from RGB to Y'C'C' and (3) Quantization of Y'C'C' to 82 and T2 are the r and g components computed using absolute 8-bit YCC data. Photo YCC space is designed for display brightness (c1 = c2 = c3 = 1) as in Equation (3). Addition- on both television and computer graphics systems. To ally, they are independent of Y; which is one of the advan- encode data, a transfer function that is, gamma correction of tages of this color space.46 Thus, this color space removes 0.018 is first applied to obtain the R, G, and B values.21 an important shortcoming that is, lack of luminance in the RGB data are further transformed into Photo YCC data. For rgb space, from compression point of view. Unlike rgb, display, Photo YCC values are converted back to RGB (and 21,82 YT1T2 is an invertible color space. However, like rgb, this later to R'G'B'). space is also noisy at low color intensities, as the intensity Gamma-correction is applied to RGB signals before con- range of T1 and T2 is [0, 1] and consequently, it is unsuitable version to Photo YCC color space unlike most of the other for compression algorithms. color spaces. This is because Photo YCC is designed for encoding the RGB signals obtained from a capturing device h1h2h3/H1H2H3 color space such as a camera. An important point to note here is that the h1h2h3 is a differential color space defined using R, G, and RGB values obtained back from Photo YCC can have values B coordinates and presented in Equation (13). in the range 0-342.21 If clipped or normalized to the desired 16 KAHU ET AL. range of 0-255 values, the output image may have consider- color plane. Conversion formulae from RGB to cGST are able distortions.82 Therefore, though Photo YCC is a presented by Porwal.19 luminance-chrominance color space, it is unsuitable for use The luminance (brightness)-chrominance color spaces like in compression algorithms. YUV, YIQ, YCbCr, and so forth, were introduced for encoding and transmission of TV signals to minimize bandwidth Kekre's luv color space requirements. YCbCr color space is widely used for the com- Kekre's Luv is a reversible RGB transform used for color pression of images and videos (JPEG, JPEG2000, H.264, blending. Conversion formulae for conversion from RGB to HEVC, etc.). This is because separation of luminance and Luv are presented in Refs. 83 and 84. From the conversion chrominance information into different color channels offers formulae, it is clear that L, u and v components have pixel the advantage of low HVS sensitivity to chrominance infor- intensity values in the range 0-765, −510-510, and mation. Thus, chrominance subsampling can be effectively −255-255. So, each plane requires more than 8 bits/pixel, utilized. Also, due to high decorrelation between luminance which is quite unsuitable for any compression algorithm. and chrominance channels, spectral redundancies are reduced. Therefore, we further analyze and compare properties of these YSbSr color space color spaces (YUV, YIQ, YCbCr, YPbPr, YDbDr, and YES) from perspective of compression in section 5. Apart from A new color space named YSbSr is proposed to reduce the these transmission oriented spaces, luminance-chrominance conversion and coding error.20 Color transforms such as color spaces such as I I I and H H H may also be used for YFbFr,85 YCoCg-R,86 and so forth, have already been pro- 1 2 3 1 2 3 image/video compression. These two color spaces are also posed for the professional extensions of the MPEG-4 coding further analyzed in detail in section 5. standard. Main aim of these color transforms is to ensure reversibility using lifting scheme for integer implementation 3.3 | Color matching based spaces and high decorrelation among the color planes. However, conversion error introduced due to the conversion from/to The color matching based spaces are related to the cone sen- RGB space is also an important factor which determines the sitivities in the human retina and hence, they are perception maximum reconstruction PSNR achieved. In this work, based. As sensitivities of the cones were not known, color authors have used K-L transform to devise YSbSr color matching experiments were performed to determine them. space with a view to reduce the conversion error. Although, The physical properties of these experiments were ultimately YSbSr has lower conversion error and high decorrelation like used to derive Color Matching Functions (CMFs), which in 2 YFbFr and YCoCg-R, it increases the bit-depth of the origi- turn were used to derive the tristimulus values. — nal image after transformation resulting in accurate recon- The International Commission on Illumination struction but lower CR. Therefore, we do not consider abbreviated as CIE (from its French name Commission — YSbSr for further quantitative analysis in section 5. Conver- International de I'Eclairage) is an organization devoted to sion formulae from RGB to YSbSr and back are proposed in international cooperation and exchange of information among its member countries on all matters related to the sci- Ref. 20. ence of colors and lighting. In 1931, CIE laid down the CIE 4,87 cGST color space 1931 standard colorimetric observer, which consists of three CMFs known as CIE XYZ CMFs. These CMFs were cGST color space can be efficiently used for lossy as well as used for the development of the CIE XYZ color space lossless compression.19 cGST has high decorrelation among (or more appropriately a colorimetric system of tristimulus color planes which is a must for any decorrelation based values) which can predict when two stimuli match in color compression approach. Real-life images mostly contain for an average human observer.2 Other color spaces in this regions having uniform colors, which require fewer bits for group such as CIE L*a*b*, CIE L*u*v*, and so forth, are encoding and regions containing edges which require more derived from the CIE XYZ color space which itself is devel- bits for representation. Sometimes different illumination oped from the CMFs. Hence, the name, “Color Matching conditions, shading and reflections give rise to perceived dif- based Spaces.” CIE is responsible for the formulation of the ferent color regions, although they actually have same color. CIE 1931 standard colorimetric observer which eventually cGST color space separates such regions, thus increasing the led to the development of XYZ color space and its deriva- CR. Major advantages of cGST color space are; (1) it is tives. Hence, these color spaces are also called the “CIE reversible, (2) its color planes have less bit-depth unlike Color Spaces.” The reviewed CIE spaces are tabulated in YUV color space, and (3) it provides high decorrelation. Table 3. Thus, using cGST space, we get 35% better CRs than JPEG- LS and 2-5 times less runtime compared to JPEG2000 with 3.3.1 | CIE XYZ and xyz color space comparable CRs. However, it should be noted that cGST CIE standardized the X, Y, and Z values as tristimulus values color space adds to redundancies in the form of a fourth which can describe any color perceived by an average KAHU ET AL. 17

TABLE 3 Color matching-based spaces values which are obtained from Equation (16), using the ref- Fundamental/ Perceptually erence white stimulus, that is, R' = G' = B' =1. Sr. no. Color spaces derived linear/nonlinear 8 1 CIE XYZ Fundamental Linear > 1=3 Y <>116 Y −16 if >0:008856 2 CIE xyz Derived Nonlinear Y0 Y W* ¼ 0 , ð17Þ 3 CIE W*u*v* Derived Linear > Y :> 903:3 otherwise 4 CIE L*a*b* Derived Linear Y0 5 Hunters' Lab Derived Linear 6 CIE L*u*v* Derived Linear * ¼ *ðÞ− ð Þ 7 CIE L*C*h* Derived Linear u 13W u u0 , 18 8 CIE LAR Derived Linear 9 IPT Derived Linear * ¼ *ðÞ− ð Þ 10 DIN99 Derived Linear v 13W v v0 , 19 human observer (CIE 1931 standard colorimetric observer). 4x 4X where u ¼ or These primaries are nonreal, that is, all of them cannot be ðÞ−2x +12y +3 ðÞX +15Y +3Z realized by actual color stimuli. In this color space, every 6y 6Y perceptible visual stimulus is described with positive X, Y, v ¼ or : ðÞ−2x +12y +3 ðÞX +15Y +3Z and Z values while those with negative values can not be perceived. XYZ color space is generally used as a reference color space for device calibration as it is considered an inter- 3.3.3 | CIE L*a*b* and Hunter's lab color space mediate device-independent color space. Linear R'G'B' and L*a*b* color space is a color-opponent space with L* for XYZ are related by the linear transformation shown in lightness (luminance) and a* and b* for the color-opponents Equation (16).29,43,87 (chrominance), based on nonlinearly compressed CIE XYZ color space. The stars (*) are marked to discriminate CIE 2 3 2 3 2 0 3 X 0:4124 0:3576 0:1805 R Lab coefficients from the Hunter's Lab coefficients. 4 5 4 5 4 0 5 Y ¼ 0:2126 0:7152 0:0722 G : ð16Þ The L*a*b* system is a good decoupler of luminance 0 Z 0:0193 0:1192 0:9505 B and chrominance. The L* component ranges from 0 (black) to 100 (white). a* component approximately measures the The above transformation matrix defines the relationship color redness (positive value) or greenness (negative value). between sRGB (converted to linear R'G'B') and XYZ. Differ- b* component approximately measures the color yellowness ent display based RGB color spaces such as Adobe RGB, (positive value) or blueness (negative value). L*a*b* com- Adobe Wide Gamut RGB will be related to XYZ using differ- ponents are obtained from XYZ tristimulus values using ent transformation matrices.3 Equations (20)-(22).43,87 Every color space obtained using CIE XYZ space is also regarded as device independent. It is a common practice to Y L* ¼ 116f −16 ð20Þ express the chromatic property of a color using only the rela- Y0 tive chromaticity coordinates of a CIE XYZ tristimulus.1 X, Y, and Z denote the amount of red, green, and blue color * ¼ X − Y ð Þ needed to form a complex color, respectively. The CIE uni- a 500 f f 21 X0 Y0 form color spaces W*u*v*, L*u*v*, L*a*b*, and LCh are developed using the XYZ space.4,30,43,87 CIE xyz space is the Y Z normalized form of XYZ space obtained by dividing each of b* ¼ 200 f −f ð22Þ the X, Y, Z tristimulus values by the summation (X+Y+Z). Y0 Z0 Only two components are enough to describe a color in 8 xyz space as third component is given by subtraction of > 1 < , if x >0:008856 1 and the other two components; for example, z =1-x-y or ðÞ¼ x3 wheref x > 16 , x =1-z-y. :7:787x + , otherwise 116

3.3.2 | CIE W*u*v* color space where X0, Y0, and Z0 are derived from the reference white W* represents luminance and u*, v* represent chrominance. stimulus, that is, R=G=B= 1 using Equation (16). The W*u*v* color space is represented by Equations (17)- The Hunter's Lab coefficients L, a and b are defined 43,87 (19). Y0,u0, and v0 are derived from the X0, Y0, and Z0 using Equations (23)-(25) 18 KAHU ET AL.

rffiffiffiffiffi Y space. Thus, L* and u* components of L*u*v* are equal to L ¼ 100 , ð23Þ Y0 W* and u* component of W*u*v* defined in Equations (17) and (18). Similarly, v*ofL*u*v* is obtained by replacing where Y is the Y tristimulus value of a specified white 0 v* in Equation (19) by (3/2)v. object. “a” and “b” are opponent color axes. “a” represents L*u*v* color space is used for estimation of restoration redness (positive) and greenness (negative) and is computed error for fast detection and impulsive noise removal in real- as in Equation (24). 93 ! time color imaging. Histogram explosion algorithm for = − = segmentation of color images was implemented in CIE ¼ X pX0ffiffiffiffiffiffiffiffiffiffiffiY Y0 ð Þ a Ka , 24 L*u*v* color space to produce perceptually improved results Y=Y 0 compared to that in the RGB space.94 where Ka is a coefficient which depends upon the illuminant (for D65, Ka is 172.30) and X is the X tristimulus value of 3.3.5 | CIE L*C*h* color space the specified white object. D65 is the standard CIE day-light L*C*h* color space is a polar color space defined using illuminant. Other opponent color axis, b, is positive for yel- L*a*b*.Lattribute is the same as L*ofL*a*b* given by low colors and negative for blue colors. It is computed as in Equation (20). C* (chroma) and h* (hue angle) are calcu- Equation (25). lated from the a*&b* coordinates as presented in Equa- ! tions (26) and (27).43 Y=Y0 −Z=Z0 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi b ¼ K pffiffiffiffiffiffiffiffiffiffiffi , ð25Þ b * ¼ 2 þ 2 ð Þ Y=Y0 C a* b* 26 where K is a coefficient which depends upon the illuminant * b * ¼ b : ð Þ h arctan * 27 (for D65, Kb is 67.20) and Z0 is the Z tristimulus value of a the reference white. Both a and b will be zero for objects Because h component of L*C*h* color space contains with the same chromaticities as the reference white (ie, ach- angular information which is very sensitive to noise and romatic, gray objects). Hunters' Lab space approximates the quantization this space is not very suitable for compression color variations using square root of the ratio of X, Y and algorithms. Z tristimuli with their reference white stimuli X0, Y0, and Z0, respectively. On the other hand, CIE L*a*b* space approxi- 3.3.6 | CIE LAR color space mates the color variations f(x) using piece-wise linear and CIE LAR space is derived from L*a*b* by rotating it by an cube-root approximations. It is experimentally observed by angle π/4. LAR space performs better than RGB, YCbCr, us that Hunters' Lab space is not exactly invertible using YUV, HSV, and L*a*b* for color-based image retrieval.95 It Equations (23)-(25), as the CIE L*a*b* space is. Thus, we is also a luminance-chrominance color space. However, due consider only L*a*b* for quantitative analysis in section 5. to lack of information about its accurate conversion formulae CIE L*a*b* space is an approximately linear color to/from RGB (or XYZ), it is not used for compression algo- space, that is, Euclidean distance between two colors is rithms. Therefore, it is not included in the quantitative analy- approximately proportional to their perceived difference by sis in section 5. humans. It is also a uniform space and thus colors in the CIE L*a*b* space are perceptually more uniformly spaced than 3.3.7 | IPT color space 88 those in RGB or HSI spaces. This enables use of a fixed As discussed earlier, CIE L*a*b* is only an approximately color distance like Just Noticeable difference (JND) in linear and perceptually uniform color space. This means that decision-making over a wide range of colors. CIE L*a*b*is Euclidean distance between two colors in CIE L*a*b* color very efficient in representing even small color differ- space poorly relates to the perceptual difference between 43,89,90 ences. Luminance and chrominance components of them. Thus, many color-difference formulae such as CMC, L*a*b* space are independently found in face and fire detec- CIE94 and CIEDE2000 were developed based on CIE 91,92 tion applications, respectively. L*a*b* color space. CIEDE2000 being the most recent one, performs better than the former two. However, it is computa- 3.3.4 | CIE L*u*v* color space tionally very complex. Besides this, CIE L*a*b* has rela- CIE L*u*v* is also a perceptually uniform luminance- tively poor hue constancy, a desirable property for uniform chrominance color space proposed by CIE in 1976 along color spaces.96 To overcome these problems, a new color with CIE L*a*b*. Major difference between the two color space, namely, IPT was proposed by Ebner and Fairchild in spaces is in the chromatic adaptation model implemented. 1998.97 IPT has better hue constancy compared to CIE L*a*b* color space normalizes its values by division by the L*a*b*. Computational complexity of conversion from CIE white point, while L*u*v* normalizes its values by subtrac- XYZ to IPT and to CIE L*a*b* is approximately equal. tion of the white point. L*u*v* is called Adams Chromatic In IPT color space, I, P, and T coordinates represent Valence Color Space and is an update of CIE W*U*V* color lightness (intensity), red-green (protan), and blue-yellow KAHU ET AL. 19

(tritan) dimensions, respectively.98 Thus, IPT is also a TABLE 4 Color equivalency-based spaces luminance-chrominance color space like CIE L*a*b*. IPT is Perceptually said to be an abbreviated form of “Image Processing Trans- Fundamental/ linear/ Sr. no. Color spaces derived nonlinear form” due to its use in gamut mapping transformations. Con- 1 Waypoint (Wpt) Derived Nonlinear version formulae for conversion from CIE XYZ to IPT are as 2 WLab Derived Linear shown in Equations (28)-(31)98; 2 3 2 32 3 L 0:4002 0:7075 −0:0807 X have three major advantages. (1) Since they are obtained 4M 5 ¼ 4 −0:228 1:15 0:0612 54Y 5, ð28Þ from CIE XYZ, they are device independent. (2) They are S 000:9184 Z mostly luminance-chrominance color spaces. Thus, they have very less correlation among their color components which is beneficial for their use in any compression algo- 0 0 0 L ¼ fLðÞ,M ¼ fMðÞ,S ¼ fSðÞ: ð29Þ rithm. (3) They are linear and perceptually uniform. This helps to minimize perceived error in the reconstructed Function f in the above equations is defined as; image/video due to quantization. The hardware implemen- x0:43 if x ≥ 0 tation also becomes simpler. The luminance-chrominance fxðÞ¼ : ; ð30Þ −−ðÞx 0 43 otherwise color matching based spaces such as L*a*b*, W*u*v*,

2 3 2 32 0 3 L*u*v*,IPTare found appropriate for image/video com- I 0:40:40:2 L 4 5 4 54 0 5 pression. Hence, they are further analyzed from compres- P ¼ 4:455 −4:851 0:396 M : ð31Þ 0 sion perspective in section 5. T 0:8056 0:3572 −1:1628 S

The range of variable (ie, axis) I in Equation (31) is from 3.4 | Color equivalency based spaces 0 to 1, whereas the range of variables (ie, axes) P and T is from −1 to 1. However, I, P, and T can have the range of Tristimulus values (or color coordinates derived from them) values as that of L*, a*, and b*, respectively, if I, P, and for an object color change due to changes in observer or T are multiplied by 100 each, respectively.99 IPT is an lighting conditions. Hence, colorimetric values always invertible color space. Thus, XYZ (or RGB) values can be assume a standard observer and illumination while describ- recovered from I, P, and T by exactly reversing the transfor- ing color. Color equivalency based spaces transform the tri- mations given in Equations (28)-(31). stimulus values to obtain color spaces whose coordinates represent actual “material color.” In other words, coordinates 3.3.8 | DIN99 color space of such a color space are approximately identical even if DIN99 is a color-difference formula developed in 1999 which there are changes in the color of the object due to change in 100 also consists of an associated color space unlike the CIE94 observer and/or lighting conditions. Color equivalency and CIEDE2000 color-difference formulae defined on the CIE based spaces are generally derived from XYZ tristimulus 100,101 L*a*b* color space. DIN99 provides better perceptual unifor- values by a Material Adjustment Transform (MAT) mity compared to CIE L*a*b* color space. DIN99 involves a with further amendments (optional). MATs are used to pre- logarithmic transformation and rescaling of the CIE L*a*b* dict how the sensor excitations (cone responses in the human eye) due to a material object (or material color) change when attributes L*andC* .96 It also modifies the original a* and ab the observing conditions change.100 Hence, these color b* coordinates of CIE L*a*b* based on the hue angle h .99 ab spaces can also be called as “MAT based color spaces.” New chromatic attributes e and f are calculated by applying Two color equivalency-based spaces proposed in Refs. rotation and stretch on the chrominance planes.96 Because 101,102 are shown in Table 4. DIN99 color planes are derived from the CIE L*a*b* color planes, the computational complexity of DIN99 is much 3.4.1 | Wpt color space higher compared to CIE L*a*b*. Therefore, though DIN99 Wpt (waypoint) is actually a normalization methodology is also a luminance-chrominance color space and provides used for defining Material Adjustment Transforms (MATs). better perceptual uniformity than CIE L*a*b*, we do not It linearly transforms sensor excitations into a material color consider DIN99 for further quantitative analysis from the equivalency representation. There is very little change in perspective of image/video compression. Color transforma- Wpt coordinate values due to changes in observer and/or tion equations from CIE L*a*b*toDIN99 and the DIN99 illuminating conditions. Wpt coordinate values are obtained color-difference equation can be found in Refs. 96 and 99. from the XYZ tristimulus values using Equation (32)101; X, Y, and Z tristimulus values can describe any color which can be perceived by an average human observer and 2 3 2 3 hence CIE XYZ is generally used as a reference color space W X 4 5 4 5 for display device calibration and is considered an interme- p ¼ A Y : ð32Þ diate device-independent color space. CIE color spaces t Z 20 KAHU ET AL.

Here, W axis represents perceptive lightness and axes This can be proved using a very simple example. If two p and t represent perceptive chromaticness. In Equation (32), identical gray colored square patches are placed on two big- A is a linear transformation matrix which is different for dif- ger square patches, one white and one black, they do not ferent observer and illuminating conditions. Transformation “appear” to have same color. This is called the “color matrix100 for D65 illuminant and 20 standard observer is; appearance phenomena.”2 In some practical situations/appli- 2 3 cations, color appearance may be more important than the : : − : 0 02964 0 97487 0 00280 actual color of objects/stimuli. Thus, viewing conditions A ¼ 4 : − : : 5: 4 83916 4 73122 0 12117 should also be taken in consideration while describing color. 0:54248 1:30671 −1:67368 Such enhancements to the existing colorimetric system are For further quantitative analysis of Wpt from the per- incorporated by “Color Appearance Model based Color spective of compression, we assume A for D65 illuminant Spaces,” simply called as “Color Appearance Models.” and 20 standard observer as a standard transformation matrix HVS's perception of the color of an object under different for conversion from XYZ to Wpt color space. viewing conditions and with different backgrounds is mod- eled by the Color Appearance Models. 3.4.2 | WLab color space All the Color Appearance Models are derived from CIE WLab (Waypoint Lab) is obtained from the Wpt color space XYZ color space due to its popularity and availability. CIE by nonlinear transformations on the polar coordinates of Wpt L*a*b* color space is a very simple form of color appear- 102 (Cpt and hpt). Unlike Wpt, WLab is a perceptually uniform ance model as it has some predictors of the color appearance color space and hence can be used in situations where other attributes. Although CIE L*a*b* color space performs well standard color spaces are applicable. Like CIE L*a*b*, as a color appearance model in most cases, it has many limi- L coordinate in WLab indicates lightness (luminance), while tations. It fails to predict many important color appearance a and b components represent chrominance. Conversion phenomena like the Hunt effect and the Stevens effect.2,99 from Wpt to WLab is described in detail by Derhak and Therefore, a few other color appearance models have been Berns100,102. The computational complexity for conversion devised such as the Nayatani et al. model, the Hunt model, from Wpt to WLab (and hence from XYZ/RGB to WLab)is the RLAB model, the ATD model, the LLAB model, the much higher compared to other conventional color spaces CIECAM97s, and the CIECAM02. The CIE L*a*b* color such as CIE L*a*b*. Hence, we do not consider WLab color space is already found to be appropriate for image/video space for further quantitative analysis in section 5. compression as per the discussion in section 3.3. The other Color equivalency-based spaces are useful for solving color appearance models are computationally too complex the problem of color constancy to some extent. A lot of and consist of more than 3 (lot more than 3) color (appear- research effort has already been done in the past to achieve ance) attributes. Hence, they are totally unsuitable for com- color constancy (similar color coordinate values irrespective pression and not included in further analysis in section 5. of changes in illumination and/or observer).103,104 Color More details about color appearance modeling and models constancy is useful in many machine vision applications can be found in Ref. 2. such as object recognition and tracking, digital photography, 105 etc. However, it has not been explored so far for image/ 3.6 | Color order systems video compression to the best of author's knowledge. There- In this section, we briefly review a few color order systems. fore, we analyze only two important color spaces in this arti- A color order system is approximately a conceptually orga- cle. We further quantitatively analyze only Wpt color space nized collection of visual stimuli that can be described using from the perspective of image/video compression in perceptual color attributes.2,3 Color Order systems overcome section 5. an important limitation of the color spaces discussed so far. Color spaces, in general, are nonintuitive and do not have a 3.5 | Color appearance modeling based spaces clear relation to the commonly understood perceptual attri- Color spaces discussed so far can efficiently describe color butes such as hue, saturation, brightness, and so forth. Color of a stimuli/object. As discussed earlier, tristimulus values images, which have spatially and typically continuously developed by the CIE system of colorimetry (CIE XYZ color varying colors cannot be defined/specified using these sys- space) can predict if the given two stimuli match in color. tems. Most of the color order systems have not even been However, they can do so under certain constraints or specific proposed by researchers, scientific or engineering communi- viewing conditions. Two stimuli/objects having identical ties. However, they satisfactorily serve the purpose and color may or may not “appear” same because of various fac- applications for which they are proposed. A few of them are tors (viewing conditions) such as surrounds, backgrounds, just used for visual calibration and matching of colors using size, shape, surface characteristics, illumination geometry, physical printed color pallets, while some of them have been etc.2 Although these factors do not affect (change) the color used to design a visual standard colors library. Exact mathe- of a stimuli/object, they do affect their color appearance. matical formulations for them are not available or even they KAHU ET AL. 21

TABLE 5 Color order systems

Sr. no. Color order systems Fundamental/derived Perceptually linear/nonlinear 1 Munsell106,107 Fundamental Linear 2 NCS108 Fundamental Linear 3 PMS109 Fundamental Linear 4 RAL110 Fundamental Nonlinear 5 OSA-UCS111,112 Fundamental Linear do not require such mathematical representations for their Color Order Systems have been designed simply for visual specific applications. Thus, Color Order Systems may not calibration of colors and/or physical color matching. They even be considered as color spaces. But, we describe a few may not even be considered as color spaces in most cases. important color order systems here in brief for the sake of Therefore, Color Appearance Modeling based Spaces and completeness. A few such important color order systems are Color Order Systems are not included in Table 6. It must be presented in Table 5. noted that advantages and limitations, if any, of the color Most of the color order systems are independently spaces in Table 6 are stated from the perspective of compres- defined and hence are fundamental, while a few may be sion in general. defined using the other spaces like RGB, XYZ, CIE L*a*b*, etc. Munsell book of colors is a set of basic color (R, G, and B) palettes with 24 color shades in each basic color and is 4 | PARAMETERS FOR ANALYSIS OF used for visual color matching. Natural Color System (NCS) COLOR SPACES is a logical color notation system that describes colors In the preceding section, a brief review of 38 important pub- exactly as we see them. NCS starts with six elementary lished color spaces is presented. A few of their applications colors (Red, Green, Blue, Yellow, Black, and White) consid- are also cited. Color spaces such as I I I , H H H , YIQ, ered as “pure.” It describes all other colors in terms of their 1 2 3 1 2 3 YUV, YCbCr, YPbPr, YDbDr, YES, CIE L*u*v*, CIE degree of visual resemblance to the elementary colors. Pan- L*a*b*, CIE W*u*v*, IPT, and Wpt have been found suit- tone Matching System (PMS) is a popular color matching able for compression. In this section, we identify the parame- system used by the printing industry for color printing. PMS ters required for their quantitative analysis for use in is used to specify colors for printing applications using their compression algorithms. The parameters identified for in- Pantone name or number. This ensures that correct colors depth analysis of these color spaces are decorrelation among are selected for printing though the color may not look visu- color planes, entropy of a color space, energy compaction of ally appropriate when displayed on the available monitor. individual color planes, conversion loss, number of compu- RAL color system was developed in 1927 to assist the paint tations required, number of bits required to represent a single and pigment industry in streamlining their products and to intensity value in each color plane, and range of intensity avoid coordination problems in the use of colors. Originally values for each color plane. Some of the abovementioned RAL started with only 40 colors. However, the palate of parameters are image independent such as number of com- colors grew to 2325 over the decades due to increased putations required, number of bits per pixel for each color demand. Optical Society America's Uniform Color System, plane, and range of intensity values for each color plane. popularly known as OSA-UCS is a color appearance system, Parameters such as decorrelation among color planes, based on a 3D Euclidean geometry with axes L, j, and entropy of a color space, energy compaction, and conversion g representing lightness, yellow-blueness and red-greenness. loss directly depend on the image or video data being com- One important property of OSA-UCS is that a given color pressed. For example, a certain color space may have differ- sample is equal in perceptual color difference form each of ent energy compaction for different images. Therefore, the its neighbors in three the dimensional space which also image dependent parameters are compared by taking their makes it very complex. This limits the popularity and practi- average on a set of standard database test images. cal utility of OSA-UCS. These application specific color order systems are not suitable for image/video compression as they are very narrow gamut spaces. 4.1 | Image independent parameters Table 6 summarizes the classification, characteristics, A brief description of the image independent parameters and advantages, disadvantages, and applications of the 33 pub- their effect on compression is presented in this section. lished color spaces discussed briefly in this section. As dis- cussed above, Color Appearance Modeling based spaces are 4.1.1 | Number of computations for color space conversion developed with a purpose altogether different from compres- This parameter is directly related with the speed of compres- sion. These color spaces are computationally very complex sion algorithms and the hardware cost. Thus, number of and contain a lot more than three color attributes (planes). computations/pixel required for conversion from RGB is 22

TABLE 6 Summary of color spaces

Suitable/not Sr. Category/ Fundamental/ Perceptually linear/ suitable for no. Color space subcategory derived nonlinear Advantages Limitations Applications compression 1. RGB/sRGB Display based/ Fundamental Nonlinear (1) Orthogonal and hardware (1) High correlation. (1) Color facial image Not suitable additive color friendly. (2) Psychologically compression using 2-stage space (2) No need of color space nonintuitive. VQ.32 conversion. (3) Very sensitive to lightness (2) Efficient video coding using and shading. interplane prediction.27 2. Rgb Display based/ Derived Nonlinear (1) Robustness to brightness (1) Very noisy at lower Not suitable additive color variations. intensities. space (2) As r+g+b= 1, only two (2) Vulnerable to quantization components need to be and rounding errors. coded. (3) Does not contain luminance information. 3. Adobe RGB Display based/ Derived Nonlinear (1) Wider gamut than sRGB. (1) Requires more bit-depth, Not suitable additive color (2) Encompasses 50.6% of that is, 16 bits/color plane. space visible colors specified by CIE. 4. Adobe wide Display based/ Derived Nonlinear (1) Wider gamut than sRGB. (1) Requires more bit-depth, Not suitable gamut RGB additive color (2) Encompasses 77.6% of that is, 16 bits/color plane. space visible colors specified by CIE. 5. ProPhoto RGB Display based/ Derived Nonlinear (1) Wider gamut than sRGB. (1) Requires more bit-depth that Not suitable additive color (2) Encompasses 100% of is, 16 bits/color plane. space likely occurring real world (2) Approximately 13% of surface colors. representable colors are imaginary. 6. CMY/ CMYK Display based/ Derived Nonlinear (1) Used for printing on white (1) High correlation. Not suitable subtractive color backgrounds. (2) Psychologically space nonintuitive. (3) Very sensitive to lightness and shading. 7. YUV Display based color Derived Nonlinear (1) Efficient color encoding in (1) Correlation still exists due (1) Color classification and Suitable space derived European TV systems. to linear transformation. segmentation using new from RGB (2) Separates luminance and YUV color space.56 chrominance components. (2) Content-based image retrieval using DWT.57 (3) message hiding using YUV and its derivatives.58 8. YIQ Display based color Derived Nonlinear (1) Efficient color encoding in (1) Correlation still exists due (1) YIQ based VQ.59 Suitable space derived European TV systems. to linear transformation. (2) Edge detection.43 from RGB (2) Separates luminance and chrominance components. KAHU TAL ET . TABLE 6 (Continued) KAHU

Suitable/not AL ET Sr. Category/ Fundamental/ Perceptually linear/ suitable for

no. Color space subcategory derived nonlinear Advantages Limitations Applications compression . 9. YCbCr Display based color Derived Nonlinear (1) Separates luminance and (1) Widely used for image and Suitable – space derived chrominance components. video compression.10 13 from RGB (2) Skin color extraction and classification using Cb and – Cr.60 62 (3) Efficient colorization method that saves computation time.63,64 10. YCgCr Display based color Derived Nonlinear (1) Separates luminance and (1) Prone to encoding and (1) Face recognition, Not suitable space derived chrominance components. decoding errors. segmentation and – from RGB analysis.65 67 11. YCgCb Display based color Derived Nonlinear (1) Separates luminance and (1) Prone to encoding and (1) Face recognition, Not suitable space derived chrominance components. decoding errors. segmentation and – from RGB analysis.65 67 12. YPbPr Display based color Derived Nonlinear (1) Separates luminance and Suitable space derived chrominance components. from RGB 13. YDbDr Display based color Derived Nonlinear (1) Separates luminance and (1) Used in SECAM color TV Suitable space derived chrominance components. system. from RGB 14. YES Display based color Derived Nonlinear (1) Separates luminance and (1) Face detection.71,72 Suitable space derived chrominance components. from RGB (2) Its attributes result in minimum overlap between skin and nonskin data.

15. I1I2I3/K1K2K3 Display based color Derived Nonlinear (1) Separates luminance and (1) Edge detection and texture Suitable – space derived chrominance components. classification.74 76 from RGB (2) Computationally less complex.

16. YT1T2 Display based color Derived Nonlinear (1) Separates luminance and (1) Very noisy at lower Not suitable space derived chrominance components. intensities. from RGB

17. h1h2h3 /H1H2H3 Display based color Derived Nonlinear (1) Separates luminance and (1) h1h2h3 is noninvertible. (1) Widely used for color edge H1H2H3 is suitable space derived chrominance components. detection.75,78 from RGB (2) Top-down quantization. 18. Photo YCC Display based color Derived Nonlinear (1) Separates luminance and (1) Noninvertible color space. (1) Storage of color images on Not suitable space derived chrominance components. CDs, display on TV and from RGB computer graphic systems.21 19. Kekre's LUV Display based color Derived Linear (1) Separates luminance and (1) Needs ~10 bits/pixel per (1) Used for color blending.83,84 Not suitable space derived chrominance components. color plane. from RGB 20. YSbSr Display based color Derived Nonlinear (1) Separates luminance and (1) Needs more than 8 bits/ (1) Reduced conversion and Not suitable space derived chrominance components. pixel per color plane. coding errors in image and from RGB (2) Reduced conversion and video compression.20 coding errors. 23 (Continues) TABLE 6 (Continued) 24

Suitable/not Sr. Category/ Fundamental/ Perceptually linear/ suitable for no. Color space subcategory derived nonlinear Advantages Limitations Applications compression 21. cGST Display based color Derived Nonlinear (1) Separates luminance and (1) Adds redundancies in the (1) Lossy as well as lossless Not suitable space derived chrominance components. form of a 4th color plane. compression.19 from RGB (2) Lower bits/pixel per color plane. 22. CIE XYZ Color matching Fundamental Linear (1) Device independent color (1) all the colors defined by X, (1) Reference color space for Not suitable based space space. Y and Z cannot be realized device calibration. by actual color stimuli. 23. CIE xyz Color matching Derived Nonlinear (1) Device independent color (1) All the colors defined by x, (1) xy chromaticity diagram. Not suitable based space space. y, and z cannot be realized (2) Only two components are by actual color stimuli. enough to describe color as (2) Does not contain luminance x + y + z =1. information. 24. CIE W*u*v* Color matching Derived Linear (1) Device independent color Suitable based space space. (2) Separates luminance and chrominance components. 25. CIE L*a*b* Color matching Derived Linear (1) Device independent color (1) Used for calculation of JND Suitable based space space. and efficient in (2) Separates luminance and representing small color chrominance components. difference.43,89,90 (2) Face and fire detection.91,92 26. Hunters' Lab Color matching Derived Linear (1) Device independent color (1) Noninvertible. Not suitable based space space. (2) Separates luminance and chrominance components. 27. CIE L*u*v* Color matching Derived Linear (1) Device independent color (1) Restoration error estimation Suitable based space space. for fast detection and (2) Separates luminance and impulsive noise removal in chrominance components. real-time color imaging.93 (2) Histogram explosion algorithm for color images segmentation.94 28. CIE L*C*h* Color matching Derived Linear (1) Device independent color (1) h* component of L*C*h*is Not suitable based space space. very sensitive to noise and (2) Separates luminance and quantization errors. chrominance components. 29. CIE LAR Color matching Derived Linear (1) Device independent color (1) Accurate color conversion (1) Color-based image Not suitable based space space. formulae to/from RGB retrieval.95 (2) Separates luminance and (or XYZ) not available. chrominance components. 30. IPT Color matching Derived Linear (1) Device independent color (1) Used in gamut mapping Suitable based space space. applications.98

(2) Separates luminance and KAHU chrominance components. (3) Better hue constancy compared to CIE L*a*b*. AL ET . KAHU TAL ET .

TABLE 6 (Continued)

Suitable/not Sr. Category/ Fundamental/ Perceptually linear/ suitable for no. Color space subcategory derived nonlinear Advantages Limitations Applications compression 31. DIN99 Color matching Derived Linear (1) Device independent color (1) Very high computational Not suitable based space space. complexity for conversion (2) separates luminance and to/from RGB. chrominance components. (3) better perceptual uniformity than CIE L*a*b*. 32. Waypoint (Wpt) Color equivalency Derived Nonlinear (1) Device independent color (1) Different machine vision Suitable based space space. applications such as object (2) Separates luminance and recognition and tracking, chrominance components. digital photography, etc.105 (3) Achieves color constancy to some extent. 33. WLab Color equivalency Derived Linear (1) Device independent color (1) Very high computational (1) Different machine vision Not suitable based space space. complexity for conversion applications such as object (2) Separates luminance and to/from RGB. recognition and tracking, chrominance components. digital photography, etc.105 (3) Achieves color constancy to some extent. 25 26 KAHU ET AL.

calculated for each color space. Note that additions and sub- absolute intensity (Iaa) of all the three color planes of an tractions are counted as same operations. This is because an image. It is given as an example for RGB space in adder hardware can also be used to perform subtractions. Equation (34). Same is the case for multiplications and divisions. Nature of Iai computations, that is, integer or floating point is also an Iapi ¼ × 100, i ¼ 1,2,3 for R,G,and B, respectively: Iaa important issue. However, we are considering only lossy ð34Þ compression and consequently floating point computations. If high amount of intensity is contained in only one color 4.1.2 | Number of bits per pixel per color plane plane, it is represented using more bits while a few bits This is the number of bits required per pixel to represent the would suffice for the other planes. Thus, an image in maximum intensity range in each color plane. More number of luminance-chrominance space typically has higher value of bits required to represent a pixel in a color plane results in lower Iap1 and lower values of Iap2 and Iap3, resulting in efficient CRs. But, they may require less time for coding and compres- compression. sion. Range of intensity values for every color plane may be dif- 4.2.4 | Entropy and entropy compaction ferent. If the intensity range of a color plane is higher, number Entropy of an image is a measure of its information content. of bits/pixel required by that color plane may be higher. How- Thus, an image with large textured regions or more edges is ever, in some cases, even if bits/pixel required by two color said to have higher entropy. Such an image requires more planes is the same, their intensity ranges may be different. bits for its representation in compressed form compared to an image having smoother regions or lower entropy. Entropy 4.2 | Image-dependent parameters of an image is given by Equation (35). A brief description of the image dependent parameters and XNn their effect on compression is presented in this section. ¼ − ðÞ ðÞðÞ ð Þ E pi × log2 pi , 35 i¼1 4.2.1 | Decorrelation where Nn is the number of different intensity values present Decorrelation among color planes ensures that image spec- in a color plane of an image and p(i) is the probability of the tral redundancies are reduced. If a highly correlated color ith intensity value. Total entropy of a color image is obtained space like RGB is used for compression, we end up encoding by adding the entropies of all the color planes. On the lines the same information repeatedly. Correlation among color of intensity compaction, we compute percentage of the total planes has also been exploited in the past for compression of entropy contained in each plane of a color space, called images in the Correlation Based Approaches (CBA). How- entropy compaction. Entropy compaction dictates the effec- ever, in this article, we restrict ourselves to the study of the tiveness of chrominance subsampling in a color space for an popular decorrelation based compression approaches. We image. measure decorrelation between color planes using Pearson's Correlation Coefficient (PCC). Low values of PCC between 4.2.5 | Average absolute high pass intensity and its compaction two color planes denote high decorrelation among them. High decorrelation results in higher CR and PSNR. Like entropy, average absolute high pass intensity of an image is also used to measure the amount of information 4.2.2 | Average absolute intensity (edge or texture) content in an image. For computing this Given an M × N color image I with three color planes, its aver- parameter, Laplacian Pyramid (LP) decomposition of an image plane is obtained initially. This yields a low pass com- age absolute intensity Iaa is calculated using Equation (33). ponent of the image plane which is half its dimension and a X3 XM XN high pass component having same size as the image plane ¼ 1 jjðÞ: ð Þ Iaa Ii,j,k 33 itself. Average absolute intensity of the high pass component M × N ×3k¼1 i¼1 j¼1 of an image plane is calculated using Equation (33). Total This parameter is one of the broad measures of the mag- average absolute high pass intensity of the image is obtained nitude of information content in an image. Although this by adding the values for its component color planes. As ear- parameter itself tells little about the nature (or characteristics) lier, percentage of high pass information contained in each of a color space, it is used in calculation of intensity compac- of the individual color planes yields its compaction on the tion across the component color planes of any color space. lines of Equation (34).

4.2.3 | Intensity compaction 4.2.6 | Conversion loss

Intensity compaction (Iapi) is the ratio of average absolute At the encoder, RGB images are converted to a suitable color intensity of the ith color plane (Iai) to the total average space for compression. Decoded images after decompression KAHU ET AL. 27

TABLE 7 Comparison of image independent parameters

Range of intensity values No. of bits per pixel Perceptual uniformity/ Color spaces No. of computations per color plane per color plane linearity RGB - R, G, B: 0 to 255 8 Nonlinear

I1I2I3 5 a,4mI1, I2, I3: 0 to 255 8 Nonlinear

H1H2H3 4 a,1mH1: 0 to 510; 9 Nonlinear H2: −255 to 255; H3: −255 to 255 YIQ 6 a,9mY, I, Q: 0 to 255 8 Nonlinear YUV 6 a,9mY, U, V: 0 to 255 8 Nonlinear YCbCr 9 a,9mY: 16 to 235; Cb, Cr: 16 to 240 8 Nonlinear YPbPr 6 a,9mY, Pb, Pr: 0-255 8 Nonlinear YDbDr 6 a,9mY, Db, Dr:0–255 8 Nonlinear YES 6 a,9mY, E, S: 0 to 255 8 Nonlinear L*u*v*11a,22m,1cL*: 0 to 100; L*-7, u*-9 and v*-8 Perceptually uniform & linear u*: −83 to 175; v*: −134 to 108 L*a*b*12a,18m,3cL*: 0 to 100; L*-7, a*- 8 and b*-8 Perceptually uniform & linear a*, b*: −100 to +100 W*u*v*11a,22m,1cW*: 0 to 100; W*-7, u*-9 and v*- 8 Perceptually uniform & linear u*: −83 to 175; v*: −90 to 72 IPT 18 a,30m,3cI: 0 to 100; I – 7, p - 8 and t - 8 Perceptually uniform & linear p, t: −100 to +100 Wpt 12 a,18mW: 0 to 255; p, t: −400 to 400 W -8,p - 10 and t - 10 Nonlinear

Abbreviations: a, number of additions or subtractions; m, number of multiplications or divisions; c, number of test checks. are converted back to RGB for display. This process intro- based MATLAB running on Intel core-i7 PC with acceler- duces a computational error due to rounding of the image ated Graphics card and 32 GB RAM is used. coefficients. Thus, MSE between the original RGB image Most of the color spaces in Table 7 are nonlinear color and the reconstructed RGB image is a “conversion loss” due spaces. YCbCr is the most widely used color space for com- to color space conversions. Conversion loss represents the pression of color images. As in Table 7, almost all the non- minimum amount of error which is inevitable due to encod- linear color spaces are similar. Although H1H2H3 requires ing and decoding. If the value of conversion loss is very least computations, it also requires 9 bits/pixel for each color high, maximum value of PSNR that can be achieved is less, plane and hence is not useful for compression. Among all resulting in less accurate reconstruction. For lossless com- the color spaces considered, Wpt requires the highest number pression algorithms, specially designed reversible color of bits/pixel per color plane. The color matching based 113–116 transforms (RCTs) are used which do not introduce spaces in Table 4 are perceptually uniform and linear. As any error in the color space conversion process. discussed in section 2, quantization is very effective in a lin- ear color space as perceived error is proportional to the quantization error at any intensity value. Out of the color 5 | QUANTITATIVE ANALYSIS AND matching-based spaces, L*u*v*and W*u*v* require more COMPARISON OF COLOR SPACES bits/pixel for chrominance planes. L*a*b* and IPT has an In this section, we compute the above discussed parameters advantage over them as they require only 8 bits/pixel for rep- for the shortlisted color spaces for possible use in compres- resentation of chrominance planes and only 7 bits/pixel for sion algorithms and benchmark them. Image independent the luminance (L* and I) planes. The computational com- parameters for all the spaces are computed and presented in plexity of IPT to/from RGB color space is higher than rest of Table 7. For image-dependent parameters, we analyze differ- the color spaces followed by L*a*b*. However, computation ent color spaces using standard image databases, such as time for color space conversion is not a major overhead LIVE database117 (29 images) and UCID database118 (1338 compared to the total time required for compression. Tables 8 images). We compute average of each image dependent and 9 show the average values of image-dependent parame- parameter on the complete databases. Use of a large number ters calculated by us for 14 different color spaces on the said of images for the calculation of image-dependent parameters databases, respectively. ensures robustness of the estimates. Image-dependent param- Certain observations from Tables 8 and 9 are presented eters are computed and presented for the two databases in as under: Tables 8 and 9, respectively. These parameters are also cal- culated for the RGB space along with the other color spaces. 1. RGB is a highly correlated color space. This is evident For computation of the parameters, a machine with windows from the high correlation values among its component 28 KAHU ET AL.

TABLE 8 Average image dependent parameters computed on LIVE database

Color Average absolute Intensity Entropy High pass intensity Conversion Maximum spaces Decorrelation intensity compaction (%) compaction (%) compaction (%) loss PSNR (dB) RGB 0.9018, 0.8932, 341.3 36.2, 34.4, 29.5 33.3, 33.5, 33.2 32.9,33.3, 33.8 - - 0.8084

I1I2I3 0.2402, 0.2832, 134.7 84.2, 11.1, 4.7 50.1, 29.3, 20.6 82.9, 9.8, 7.3 0.3280 52.9 0.2864

H1H2H3 0.1636, 0.5225, 280.4 85.4, 5.9, 8.7 51.0, 32.6, 16.4 83.3, 8.3, 8.4 0.0392 62.2 0.2929 YIQ 0.2163, 0.3323, 136.8 85.3, 9.6, 5.1 57.7, 31.0, 11.3 82.8, 9.5, 7.7 0.3127 53.2 0.3208 YUV 0.2805, 0.5499, 137.6 84.8, 7.3, 7.9 61.3, 10.8, 27.9 83.0, 7.3, 9.7 0.3141 53.2 0.1826 YCbCr 0.2811, 0.5497, 368.4 31.4, 32.8, 35.8 42.8, 29.4, 27.8 81.5, 9.5, 9.0 0.3885 52.2 0.1826 YPbPr 0.2854, 0.4691, 137.0 85.5, 8.1, 6.4 62.4, 11.2, 26.4 83.6, 8.4, 8.0 0.3132 53.4 0.1949 YDbDr 0.2805, 0.5503, 170.5 69.0, 17.6, 13.4 66.1, 13.0, 20.9 66.2,17.4, 16.4 0.1143 57.6 0.1826 YES 0.1936, 0.5225, 137.9 85.0, 6.1, 8.9 51.2, 19.7, 29.1 83.3, 8.3, 8.4 0.3245 53.0 0.2802 L*u*v* 0.1649, 0.4078, 73.5 66.8, 13.8, 19.4 44.4, 23.3, 32.3 60.4,17.6, 22.0 0.8298 48.9 0.3146 L*a*b* 0.2519, 0.2729, 68.2 71.7, 10.4, 17.9 48.6, 17.4, 34.0 65.0,16.5, 18.5 0.7996 49.1 0.2809 W*u*v* 0.1649, 0.4078, 68.7 71.4, 14.7, 13.9 45.9, 24.1, 30.0 65.1,19.0, 15.9 0.947 48.4 0.3146 IPT 0.7468, 0.2147, 80.62 15.5, 54.4, 30.1 49.2, 16.1, 34.7 22.3, 50.2, 27.5 5.0601 41.09 0.3563 Wpt 0.3465, 0.3379, 105.52 58.9, 16.0, 25.1 47.6, 17.1, 35.3 54.0, 20.6, 25.4 1.0754 47.82 0.4639

color planes as seen in Tables 8 and 9. The other color have equal amounts of the total average absolute inten- spaces have correlation values much lower than that of sity whereas IPT color space has more amount of the RGB. Because most of the color spaces (except RGB) total average absolute intensity in the P chrominance considered here are luminance-chrominance color spaces, plane. correlation between the luminance plane, and each of the 4. Since luminance plane is said to have more information chrominance planes is lower. However, in case of IPT than chrominance planes, entropy of luminance planes color space, correlation between I (intensity/luminance) should be typically higher than the chrominance planes. and P (one of the chrominance) planes is quite high. The This is crucial for the success of chrominance subsam- correlation among the chrominance planes can be higher pling implemented in any compression algorithm. Since as represented by the second correlation value. Color only luminance-chrominance color spaces are discussed

spaces such as I1I2I3, YIQ,CIEL*a*b*, and Wpt are here, all of them offer good entropy compaction as seen exception to this. Thus, it can be said that they have least from Tables 8 and 9. YIQ, YUV, YPbPr, and YDbDr dependency between their chrominance planes. color planes offer high entropy compaction whereas the 2. It is clear that if the value of average absolute intensity color matching-based spaces and Wpt space offer is lower, very less number of bits/pixel are required for slightly lower entropy compaction. YCbCr has the low- the color plane. Average absolute intensity is the lowest est entropy compaction with its luminance plane con- for L*a*b* and the highest for YCbCr color space. taining only 42.78% of the total entropy. L*a*b* color

3. If higher percentage of Iaa is present in the luminance space has average performance in terms of this plane than chrominance planes, less information would parameter. be lost during chrominance subsampling. Color spaces 5. Entropy measures the information whereas average abso-

such as I1I2I3, H1H2H3, YIQ, YUV, YPbPr, and YES have lute high pass intensity measures the magnitude of high very good intensity compaction. L*u*v*, L*a*b*, pass content in an image. As seen in Table 8, most of the W*u*v*, YDbDr, and Wpt color spaces have moderate high pass intensity is contained in luminance plane of all levels of intensity compaction. As seen in Table 8, the color spaces derived from RGB except YDbDr. More YCbCr and IPT color spaces have poor intensity com- amount of high pass intensity is contained in the luminance paction. The three color planes of YCbCr approximately plane of color matching-based spaces also except IPT; KAHU ET AL. 29

TABLE 9 Average image-dependent parameters computed on UCID database

Color Average absolute Intensity Entropy High pass intensity Conversion Maximum spaces Decorrelation intensity compaction (%) compaction (%) compaction (%) loss PSNR (dB) RGB 0.9252, 0.9459, 327.3 35.9, 33.1, 31.0 33.8, 33.4, 32.8 35.5, 31.2, 33.3 - - 0.8438

I1I2I3 0.2882, 0.3569, 130.5 83.4, 12.8, 3.8 56.0, 26.5, 17.5 65.0, 22.4, 12.6 0.3327 52.9 0.2572

H1H2H3 0.2342, 0.6808, 268.9 83.6, 7.3, 9.1 46.9, 29.8, 23.3 65.4, 18.0, 16.6 0.0652 60.0 0.3238 YIQ 0.2708, 0.3683, 131.0 84.0, 12.0, 4.0 58.1, 26.0, 15.9 64.4, 22.7, 12.9 0.3164 53.1 0.2856 YUV 0.3157, 0.6969, 133.0 82.9, 7.2, 9.9 58.6, 16.4, 25.0 64.2, 13.7, 22.1 0.3114 53.2 0.2483 YCbCr 0.3153, 0.6960, 366.2 29.9, 34.1, 36.0 42.9, 28.8, 28.3 64.6, 16.5, 18.9 0.3924 52.2 0.2480 YPbPr 0.3156, 0.6971, 132.0 83.4, 8.4, 8.2 59.0, 17.2, 23.8 65.5, 16.0, 18.5 0.2986 53.4 0.2483 YDbDr 0.3156, 0.6973, 168.2 67.0, 16.9, 16.1 59.5, 19.1, 21.4 42.5, 26.7, 30.8 0.1177 57.4 0.2483 YES 0.2397, 0.6808, 132.5 83.1, 7.6, 9.3 55.9, 20.1, 24.0 64.7, 18.3, 17.0 0.3284 52.9 0.3074 L*u*v* 0.2468, 0.5777, 71.1 65.8, 15.9, 18.3 50.2, 23.2, 26.6 39.8, 30.5, 29.7 0.8895 48.6 0.3417 L*a*b* 0.2347, 0.3925, 64.4 71.6, 10.6, 17.8 51.7, 20.9, 27.4 44.9, 26.8, 28.3 0.8463 48.9 0.3060 W*u*v* 0.2468, 0.5777, 66.6 69.9, 17.0, 13.1 51.6, 23.9, 24.5 44.0, 33.9, 22.1 1.0209 48.0 0.3417 IPT 0.6447, 0.4041, 78.95 18.1, 49.1, 32.8 45.3, 20.9, 33.8 27.9, 46.4, 25.7 6.3798 40.1 0.3821 Wpt 0.2819, 0.3650, 106.49 61.6, 14.1, 24.3 49.6, 22.6, 27.8 41.9, 28.4, 29.7 1.9222 45.3 0.3833

though it is lower compared to the RGB derivatives. Per- correlation among its color planes alongside I1I2I3 and YIQ. formance of Wpt is even lower than color matching based However, I1I2I3 and YIQ are nonlinear color spaces. Among spaces, though it is slightly better than IPT and YDbDr. the nonlinear color spaces, YCbCr has been widely used for 6. All the color matching based spaces along with the Wpt compression. But CIE L*a*b* outperforms YCbCr in many color space have high values of conversion loss as important parameters except high pass intensity compaction shown in Tables 8 and 9. One possible reason for this is and conversion loss. that direct conversion formulae from RGB to these color To compare the 13 shortlisted color spaces visually, spaces are not available. RGB values are first converted three component color planes of each of these color spaces XYZ and then to the desired CIE color space. This is one for “bikes” image from the LIVE database are shown in of the major drawbacks of the color matching-based Figure 4. Figure 4 (i) shows the original “bikes” RGB image spaces. Highest value of conversion loss is obtained for followed by the individual R, G, and B planes shown in IPT color space. Last column of Tables 8 and 9 show Figure 4 (ii), (iii), and (iv), respectively. As seen, all the the maximum PSNR corresponding to each of the con- three R, G, and B planes contain similar information. It is version loss shown. It is clearly seen that PSNR values apparent that the R, G, and B planes are highly correlated. range from approximately 40 dB to 52 dB. As for all The “bikes” image shown in Figure 4 (i) is converted into the color spaces, PSNR values are more than 40 dB, each of these 13 color spaces. The converted images, along reconstructed image quality would be visually same. with their component color planes, are shown in Figure 4 (v)-(lvi). Although these figures are converted to the above It may be noted that, quantitative values of the parame- mentioned color spaces, they are displayed using the RGB ters for the same color space are different for the two data- framework. In other words, to display the “bikes” image in bases. However, the relation of the values among the color each of these 13 color spaces, intensity values in the first spaces is same. For example, in Tables 8 and 9 both, L*a*b* (luminance) plane of these images is considered as the inten- has the lowest average absolute intensity value and YCbCr sity values in the R plane of an RGB image. Similarly, inten- has the highest value. L*a*b* has the lowest average abso- sity values in the second and third (chrominance) planes are lute intensity value among all the color spaces studied. considered as the intensity values in the G and B planes,

Therefore, L*a*b* inherently requires less number of bits/ respectively. Color spaces such as I1I2I3, H1H2H3, YIQ, pixel for each of its color planes. It also has very low YUV, YPbPr, YDbDr, YES, and L*a*b* contain most of their 30 KAHU ET AL.

information in the luminance planes. These color spaces have higher average absolute intensity values for luminance planes compared to their chrominance planes unlike RGB, YCbCr, Ipt, and Wpt as evident from Tables 8 and 9. There- fore, they appear red in color compared to RGB, YCbCr, Ipt, and Wpt color spaces as evident from Figure 4 (i), (v), (ix), (xiii), (xvii), (xxi), (xxv), (xxix), (xxxiii), (xlix), and (liii), respectively. In fact, chrominance t plane in the Ipt color space has more information content than the other two planes as seen in Figure 4 (lii). It can be concluded that the visual results are conforming to the quantitative results tabu- lated in Tables 7–9.

6 | CONCLUSION

In this article, we have presented a survey of 38 major color spaces starting from their definitions followed by their mathe- matical formulation, advantages, disadvantages, applications and their suitability for compression. Initially, the color spaces are categorized into different groups based on their origin and have been described using their properties such as number of components required to define the color space, perceptual linearity, and uniformity, and so forth. Suitability of these color spaces for image/video compression is ascer- tained using basic properties such as their invertibility, avail- ability of accurate conversion formulae to/from the RGB color space, sensitivity to noise, and so forth. On the basis of this preliminary analysis, 13 color spaces were shortlisted as suitable for image/video compression. These 13 color spaces were further quantitatively analyzed in detail using two image-dependent and six image-independent parameters. It has been found that L*a*b* is a good choice for compres- sion of images/videos due to its advantages such as percep- tual linearity and uniformity, high entropy compaction, high decorrelation among its component color planes, and so forth. However, it has two major shortcomings. First, its conversion computational complexity to/from RGB space is high. But, compared to the total computational complexity of any compression algorithm, this overhead is negligible. Second, it has high value of conversion loss. Conversion loss dictates the achievable maximum value of PSNR for an image using a lossy compression technique. However, maxi- mum PSNR values even for CIE spaces for the given con- version loss are greater than 48 dB. Thus, L*a*b*is definitely a viable alternative to YCbCr for acceptably accu- rate and fast compression at least for mobile computing and wireless platforms.

CONFLICTS OF INTEREST FIGURE 4 “Bikes” image represented using 13 shortlisted color spaces and the component color planes of these spaces The authors of this manuscript have no relevant financial interests in the manuscript and no other potential conflicts of interest to disclose. KAHU ET AL. 31

ORCID [30] Dougherty ER. Electronic Imaging Technology. Vol 60. Bellingham, Washington: SPIE Press; 1999. Samruddhi Y. Kahu https://orcid.org/0000-0003-2761-8313 [31] Orchard M, Bouman C. Color quantization of images. IEEE Trans Signal Process. 1991;39:2677-2690. [32] Huang J, Wang Y. Compression of color facial images using feature cor- REFERENCES rection two-stage vector quantization. IEEE Trans Image Process. 1999;8: [1] Sharma G, Trussell H. Digital color imaging. IEEE Trans Image Process. 102-109. 1997;6:901-932. [33] Littmann E, Ritter H. Adaptive color segmentation-a comparison of neural [2] Fairchild M. Color Appearance Models. 2nd ed. USA: Wiley; 2013. and statistical methods. IEEE Trans Neural Netw. 1997;8:175-185. [3] Sharma G, ed. Digital Color Imaging Handbook. Webster, NY: CRC [34] Lalanne T, Lempereur C. Color recognition with a camera: a supervised Press; 2003. algorithm for classification. IEEE Southwest Symp Image Anal Interpret. [4] Gonzalez R, Woods R. Digital image processing. 3rd ed. Upper saddle 1998;198-204. river, NJ: Prentice Hall; 2008. [35] Stachowicz M, Lemke D. Color recognition. Paper presented at: 22nd Intl. [5] Kuehni R. Color Space and its Divisions: Color Order from Antiquity to Conf Info Technol Interfaces2000; 329–334. the Present. Hoboken, NJ: J. Wiley & Sons; 2003. [36] Xu J, Li S, Chen Z. Color analysis for Chinese car plate recognition. IEEE [6] A. W. M. S, Gevers T. Color-based object recognition. Pattern Recogn. Intl Conf Robotics Intell Syst Signal Process. 2003;2:1312-1316. 1999;32:453-464. [37] Yang Y, Bai J, Tian R, Liu N. A vehicle license plate recognition system [7] Lin Z, Wang J, Ma K. Using eigencolor normalization for based on fixed color collocation. Intl Conf Mach Learning Cybern. 2005; illumination-invariant color object recognition. Pattern Recogn. 2002;35: 9:5394-5397. 2629-2642. [38] Kauppinen H, Silven O. The effect of illumination variations on [8] Tao L, Xu G. Color in machine vision and its application. Chin Sci Bull. color-based wood defect classification. Paper presented at: 13th IEEE Intl. 2001;46:1411-1421. Conf. Pattern Recognition. 1996; 828–832. [9] Wallace GK. The JPEG still picture compression standard. IEEE Trans [39] Golland P, Bruckstein A. Why RGB? Or how to design color displays for Consumer Electronics. 1992;38:xviii-xxxiv. Martians. Graph Models Image Process. 1996;58:405-412. [10] CCITT 1991Digital compression and coding of continuous-tone still [40] Machajdik J, Hanbury A. Affective image classification using features – images, JTC1/SC29/WG1 10918 1, CCITT Recommendation T.81. inspired by psychology and art theory. Paper presented at: 18th ACM Intl [11] Skodras A, Christopoulos C, Ebrahimi T. The JPEG 2000 still image com- Conf Multimedia 2010; 83–92. pression standard. IEEE Signal Process Mag. 2001;18:36-58. [41] Zhao S, Gao Y, Jiang X, Yao H, Chua T, Sun X. Exploring [12] Wiegand T, Sullivan G, Bjontegaard G, Luthra A. Overview of the principles-of-art features for image emotion recognition. Paper presented H. 264/AVC video coding standard. IEEE Trans Circuits Syst Video Tech- at: 22nd ACM Intl Conf Multimedia2014; 47–56. nol. 2003;13:560-576. [42] Zhai H, Chavel P, Wang Y, Zhang S, Liang Y. Weighted fuzzy correlation [13] Sullivan G, Ohm J, Han W, Wiegand T. Overview of the high efficiency for similarity measure of color-histograms. Opt Commun. 2005;247: video coding (HEVC) standard. IEEE Trans Circuits Syst Video Technol. 49-55. 2012;22:1649-1668. [43] Cheng H, Jiang X, Sun Y, Wang J. Color image segmentation: advances [14] Dumic E, Mustra M, Grgic S, Gvozden G. Image quality of 4: 2: 2 and 4: and prospects. Pattern Recogn. 2001;34:2259-2281. 2: 0 chroma subsampling formats. Intl Sym ELMAR. 2009;19-24. [44] Woebbecke D, Meyer G, Von Bargen K, Mortensen D. Color indices for [15] Gershikov E, Lavi-Burlak E, Porat M. Correlation-based approach to color weed identification under various soil, residue, and lighting conditions. image compression. Signal Process Image Commun. 2007;22:719-733. Trans ASAE. 1995;38:259-269. [16] Goffman-Vinopal L, Porat M. Color image compression using inter-color [45] Liu Y, Yang F, Yang R, Jia K, Zhang H. Research on segmentation of correlation. Intl Conf Image Process. 2002;2:353-356. weed images based on computer vision. J Electronics (China). 2007;24: [17] Roterman Y, Porat M. Progressive image coding using regional color cor- 285-288. relation. Paper presented at: 4th EURASIP Conf. Video/Image Process. [46] Terrillon J, David M, Akamatsu S. Detection of human faces in complex Multimedia Communications. 2003; 65–70. scene images by use of a skin color model and of invariant Fourier-Mellin [18] Gershikov E, Porat M. Correlation VS. decorrelation of color components moments. Paper presented at: 14th IEEE Intl. Conf. Pattern Recogni- in image compression—Which is preferred? Paper presented at: 15th tion1988; 1350–1355. European Conf. Signal Process. 2007; 985–989. [47] Adobe RGB. http://www.adobe.com/digitalimag/pdfs/AdobeRGB1998.pdf [19] Porwal J. A 3D! 4D color space transform for efficient lossless image [48] Adobe wide Gamut RGB. http://www.adobe.com/products/adobemag/ compression. IEEE Vis Commun Image Process (VCIP). 2013;1-6. [20] Kim H, Kim WS, Cho DS. A new color transform for RGB coding. IEEE archive/pdfs/98auhtbf.pdf Intl Conf Image Process. 2004;1:107-110. [49] Pro Photo RGB. http://en.wikipedia.org/wiki/ProPhoto_RGB_colour_ [21] A. Ford and A. Roberts, 2011"Colour space conversions, 1998," http:// space www.poynton.com/PDFs/coloureq.pdf [50] Rodriguez M. A graphic arts perspective on RGB-to-CMYK conversion. [22] Ramanath R, Snyder WE, Yoo Y, Drew MSJISPM. Color image proces- IEEE Intl Conf Image Process. 1995;2:319-322. sing pipeline. IEEE Signal Process Mag. 2005;22:34-43. [51] Tkalcic M, Tasic J. Colour spaces: perceptual, historical and applicational [23] Poynton C. Frequently-asked questions about gamma, (2009b). http:// background. Eur Secur. 2003;1:304-308. www.poynton.com/PDFs/GammaFAQ.pdf. [52] Wang X, Xiu X, Zhu W, Tang H. Color control of the multi-color printing [24] Multimedia systems and equipment–colour measurement and device. J Zhejiang Univ Sci A. 2006;7:1187-1192. management–part 2-1: Colour management–default RGB colour space– [53] Shih P, Liu C. Comparative assessment of content-based face image sRGB, IEC 61966–2-1; 1999. retrieval in different color spaces. Int J Pattern Recognit Artif Intell. 2005; [25] Stokes M, Anderson M, Chandrasekar S, Motta R. A standard default 19:873-893. color space for the internet—sRGB, 1996. 2012 http://www.w3.org/ [54] Roterman Y, Porat M. Progressive image coding using regional color cor- Graphics/Color/sRGB. relation. Paper presented at: 4th EURASIP Conf. Video/Image Process. [26] Pratt W. Spatial transform coding of color images. IEEE Trans Commun Multimedia Communications2003; 65–70. Technol. 1971;19:980-992. [55] Yuan Y, Mandal M. Embedded color image coding using context- [27] Kim W, Cho D, Kim H. Interplane prediction for RGB video coding. modeled wavelet difference reduction. IEEE Intl Conf Acoustics, Speech, IEEE Intl Conf Image Process. 2004;2:785-788. Signal Process. 2004;3:61-64. [28] Sangwine S, Horne R, eds. The Colour Image Processing Handbook. Dor- [56] Gupta G, Bailey D. Discrete YUV look-up tables for fast colour segmenta- drecht, Netherlands: Springer Sci & Business Media; 2012. tion for robotic applications. Canadian Conf Electrical Comput Engg. [29] Pietikainen M, Nieminen S, Marszalec E, Ojala T. Accurate color discrim- 2008;963-968. ination with classification based on feature distributions. IEEE Intl. Conf. [57] Chiang T, Tsai T, Huang Y. Image retrieval based on the wavelet features Pattern Recognit. 1996;3:833-838. of interest. IEEE Intl Conf Syst Man Cybern. 2006;4:3324-3329. 32 KAHU ET AL.

[58] Stanescu D, Stratulat M, Groza V, Ghergulescu I, Borca D. Steganography [86] Malvar H, Sullivan G. YCoCg-R: A color space with RGB reversibility in YUV color space. Intl Workshop on Robotic Sensors Environments. and low dynamic range. ISO/IEC JTC1/SC29/WG11 and ITU-T SG16 2007;1-4. Q2003; 6. [59] Wu X. YIQ vector quantization in a new color palette architecture. IEEE [87] CIE 15 Supp. 2:1978 Recommendations on uniform color spaces. Paris: Trans Image Process. 1996;5:321-329. International Commission on Illumination; 1978. [60] Im J, Lee S, Song J, Kim T, Kang B. Implementation of memory-free [88] Kasson J, Plouffe W. An analysis of selected computer interchange color skin-detection system for mobile display device. Paper presented at: 9th spaces. ACM Trans Graph (TOG). 1992;11:373-405. Intl. Symp Signal Process Appl2007; 1–4. [89] Hadhoud M, Fouad M, Hamdi A. Performance study for color filter Array [61] Mohabbati B, Kasaei S. An efficient wavelet/neural networks-based face Demosaicking methods. Nat Radio Sci Conf. 2007;1-10. detection algorithm. Paper presented at: 1st IEEE and IFIP Intl. Conf Cen- [90] Alata O, Quintard L. Is there a best color space for color image characteri- tral Asia Internet2005; 5. zation or representation based on multivariate Gaussian mixture model? [62] Liu Z, Li Y, He X, Wang W, Zhou J, Li K. Extraction of face regions in Comp Vis Image Understanding. 2009;113:867-877. color image. IEEE Intl Conf Mach Learning Cybern. 2002;3:1340-1343. [91] Cai J, Goshtasby A. Detecting human faces in color images. Image Vis [63] Noda H, Niimi M, Korekuni J. Simple and efficient colorization in YCbCr Comput. 1999;18:63-75. color space. Paper presented at: 18th Intl. Conf Pattern Rec2006; [92] Wang SJ, Jeng DL, Tsai MT. Early fire detection method in video for ves- 685–688. sels. J Syst Softw. 2009;82:656-667. [64] Noda H, Takao N, Niimi M. Colorization in YCbCr space and its applica- [93] Smolka B, Chydzinski A. Fast detection and impulsive noise removal in tion to improve quality of JPEG color images. IEEE Intl Conf Image Pro- color images. Real-Time Imaging. 2005;11:389-402. cess. 2007;4:385-388. [94] Mlsna P, Zhang Q, Rodriguez J. 3-D histogram modification of color [65] Travis D. Effective Color Displays: Theory and Practice. Vol 1991. images. IEEE Intl Conf Image Process. 1996;3:1015-1018. London: Academic press; 1991. [95] Chen Y, Hao P. Optimal transform in perceptually uniform color space [66] Vezhnevets V, Sazonov V, Andreeva A. A survey on pixel-based skin and its application in image retrieval. Paper presented at: 7th Intl. Conf color detection techniques. Proc Graphicon. 2003;3:85-92. Signal Process2004; 1107–1110. [67] De Dios J, García N. Face detection based on a new color space YCgCr. [96] Shen S. Color difference formula and uniform color space modeling and IEEE Intl Conf Image Process. 2003;2:909-912. evaluation, RIT Thesis2009. [68] Zhengzhen Z, Yuexiang S. Skin color detecting unite YCgCb color space [97] Ebner F, Fairchild M. Development and testing of a color space (IPT) with with YCgCr color space. Intl Conf Image Analysis Sig Process. 2009; improved hue uniformity. Paper presented at: 6th Color Imaging 221-225. Conf1998; 8–13. [69] Asif M, Qi C, Fareed MMS. Multiple license plate detection for Chinese [98] Ebner F. Derivation and Modeling of Hue Uniformity and Development of vehicles in dense traffic scenarios. IET Intell Transp Syst. 2016;10: IPT. Rochester, NY: Rochester Institute of Technology; 1998. 535-544. [99] Xue Y. Uniform Color Spaces Based on CIECAM02 and IPT Color Dif- [70] Saber E, Tekalp A. Frontal-view face detection and facial feature extrac- ference Equations. Rochester, NY: Rochester Institute of Technology; tion using color, shape and symmetry based cost functions. Pattern 2009. Recogn Lett. 1998;19:669-680. [100] Derhak M. Spectrally Based Material Color Equivalency: Modeling and [71] Wesolkowski S, Jernigan M, Dony R. Comparison of color image edge Manipulation. Rochester, NY: Rochester Institute of Technology; 2015. detectors in multiple color spaces. IEEE Intl Conf Image Process. 2000;2: [101] Derhak M, Berns R. Introducing Wpt (waypoint): a color equivalency rep- 796-799. resentation for defining a material adjustment transform. Color Res Appl. [72] Gomez G, Sanchez M, Sucar L. On selecting an appropriate colour space 2015;40:535-549. for skin detection. Mexican Intl Conf Artificial Intell. 2002;69-78. [102] Derhak M, Berns R. Introducing WLab—going from Wpt (waypoint) to a [73] Ohta Y, Kanade T, Sakai T. Color information for region segmentation. uniform material color equivalency space. Color Res Appl. 2015;40: Comput Graph Image Process. 1980;13:222-241. 550-563. [74] Lin C. Face detection in non-uniform illumination conditions by using [103] Oleari C. Corresponding color datasets and a chromatic adaptation model color and triangle-based approach. Joint Conf Info Sci. 2006. based on the OSA-UCS system. J Opt Soc Am A. 2014;31:1502-1514. [75] Wesolkowski S. 1999Color Image edge detection and segmentation: a [104] Mirzaei H, Funt B. Gaussian-metamer-based prediction of colour stimulus comparison of the vector angle and euclidean distance color similarity change under illuminant change. In Proc. AIC 2011 Midterm Meeting, measures. MS Thesis, University of Waterloo, UWSpace. http://hdl. Intl. Colour Assoc., 2011. handle.net/10012/937 [105] Finlayson G, Hordley S, Hubel P. Color by correlation: a simple, unifying [76] Jiang S, Huang Q, Ye Q, Gao W. An effective method to detect and cate- framework for color constancy. IEEE Trans Pattern Analysis Machine gorize digitized traditional Chinese paintings. Pattern Recogn Lett. 2006; Intel. 2001;23:1209-1221. 27:734-746. [106] Munsell AH. Atlas of the Munsell Color System. Wadsworth: Howland & [77] Nevatia R. Color edge detector and its use in scene segmentation. IEEE Company, Inc.; 1915. Trans Syst Man Cybern. 1977;7:820-826. [107] Ye Q, Gao W, Zeng W. Color image segmentation using density-based [78] Wesolkowski S, Jernigan M, Dony R. Comparison of color image edge clustering. IEEE Intl Conf Acoustics Speech Signal Process. 2003;3: detectors in multiple color spaces. IEEE Intl Conf Image Process. 2000;2: 345-348. 796-799. [108] Hård A, Sivik L, Tonnquist G. Ncs, natural color system—from con- [79] Braquelaire JP, Brun L. Comparison and optimization of methods of color cept to research and applications. Part I. Color Res Appl. 1996;21: image quantization. IEEE Trans Image Process. 1997;6:1048-1052. 180-205. [80] Kaiser PK, Boynton RM. Human color vision. Washington, DC: Optical [109] Balasubramanian R, Dalal E. Method for quantifying the color gamut of Soc Amer. 1996. an output device. Proceedings of Color Imaging: Device-Independent [81] Levine MD. Vision in Man and Machine. New York, NY: McGraw-Hill Color, Color Hard Copy, and Graphic Arts, Vol. 3018. 1997;3018: College; 1985. 110–116. [82] Encoding PC, Schemes C. Article No. 4–from the Fully Utilizing Photo [110] RAL Colour System. http://allral.com/ CD Images Series. Rochester, NY: Eastman Kodak Company; 1994. [111] MacAdam D. Uniform color scales. J Opt Soc Am. 1974;64:1691-1702. [83] Kekre H, Thepade S. Image Blending in Vista Creation using Kekre's [112] MacAdam D. Colorimetric data for samples of OSA uniform color scales. LUV Color Space. Paper presented at: IEEE Colloquium and Intl. Conf., J Opt Soc Am. 1978;68:121-130. Sardar Patel Institute of Technology2008; 4–5. [113] Strutz T. Multiplierless reversible color transforms and their automatic [84] Kekre H, Thepade S. Color traits transfer to grayscale images. Paper pre- selection for image data compression. IEEE Trans Circuits Syst Video sented at: 1st Intl. Conf Emerging Trends in Engg Technol2008; 82–85. Technol. 2013;23:1249-1259. [85] Topiwala P, Tu C, FastVDO L L C. New invertible integer color trans- [114] Kim S, Cho N. A lossless color image compression method based on a forms based on lifting steps and coding of 4: 4: 4 video. ISO/IEC new reversible color transform. IEEE Vis Commun Image Process (VCIP). JTC1/SC29/WG11 and ITU2003. 2012;1-4. KAHU ET AL. 33

[115] Hao P, Shi Q. Comparative study of color transforms for image coding and derivation of integer reversible color transform. 15th IEEE Intl Conf at R.T.M. Nagpur University, Nagpur, India. He is a member Pattern Recognition 2000; 224–227. of ISTE. His research interests include display and printing [116] Starosolski R. New simple and efficient color space transformations for applications of image processing. lossless image compression. J Vis Commun Image Represent. 2014;25: 1056-1063. KISHOR M. BHURCHANDI received a BEng and MEng [117] Sheikh H, Sabir M, Bovik A. A statistical evaluation of recent full refer- degrees in electronics engineering in 1990 and 1992. He ence image quality assessment algorithms. IEEE Trans Image Process. 2006;15:3440-3451. then received a Ph.D. degree from Visvesvaraya Regional [118] Schaefer G, Stich M. UCID: an uncompressed color image database. College of Engineering, Nagpur University, Nagpur, India in Proceedings of Storage and Retrieval Methods and Applications for 2002, where he is currently working as a Professor. He is Multimedia, Volume 5307. 2004; 472–480. principal investigator of two major funded research projects AUTHORS' BIOGRAPHIES in the field of signal processing and embedded systems. His research interests include color image processing and analy-

SAMRUDDHI Y. KAHU received BEng degree in Electronics sis, computer vision, digital signal processing and embedded and Communication Engineering from Nagpur University, systems. Nagpur, India in 2011 and MTech degree in Communication Systems Engineering from Visvsesvaraya National Institute of Technology, Nagpur, India in 2014, where she is cur- rently pursuing a doctorate. Her research interests include How to cite this article: Kahu SY, Raut RB, image and video compression and processing. Bhurchandi KM. Review and evaluation of color RAJESH BRAUT received BEng and MTech degrees in spaces for image/video compression. Color Res Appl. electronics in 1997 and 2006 respectively from Nagpur Uni- 2019;44:8–33. https://doi.org/10.1002/col.22291 versity, Nagpur, India. Currently, he is pursuing a doctorate