International Journal of Scientific & Engineering Research, Volume 4, Issue 11, November-2013 1386
ISSN 2229-5518
Image Quality using Attributes
Yusra Al-Najjar, Dr. Chen Soong Der
—————————— ——————————
On the other hand, image quality has many overlapping
attributes that could be classified differently into measureable
MAGE quality is not an attribute that could be measured for an image, quality takes into consideration many parameters. Human eye is the best judge for image quality,
but what is meant by quality?
Quality means how natural, colorful, bright, clear,
distortion free the image is, beside other attributes in which
normal user sees it difficult to be found directly.
Image enhancement is the simplest and most common field
in digital image processing techniques. The main idea of image enhancement is to increase the contrast of the image to get specific details.
RGB colour space, which is used by most of the conventional methods, could not be adjusted easily to Human Visual System (HVS) property because it does not express the colors as a combination of lightness and chromaticity. On the other hand, the CIELUV colour space is capable of performing colour quantization. The HVS is very sensitive to the colour that is considerably different from the surrounding colors, even if it occupies relatively smaller region of the image. Yet, this property may be ignored by the quantization method using Mean Squared Error (MSE).
Naturalness is defined as the degree of consistency between visual representation of the image and the knowledge of reality as stored in memory [1]. It could also be considered by addressing e.g. how manipulated changes in attributes such as Sharpness or Chroma affect the perceived naturalness of an image [2].
Naturalness attributes also said to include some artifactual
attributes (due to e.g. low-level quality), preferential
attributes (e.g. colour balance, preferred colors), aesthetic
attributes (e.g. lighting, perspective, cropping), and personal
attributes (e.g. personal connection, familiarity) [2].
attributes (e.g. contrast, colorfulness, noise. brightness) and hard to be measures attributes - since it depends on observers’ point of view - (e.g. usefulness, balance, pleasantness, completeness), see Fig. 2.
Following is a discussion for some of these attributes:
CIE defined by the as the attribute of a visual sensation according to light percentage emitted from the area. It measures the subjective sensation produced by a particular luminance, i.e., the brightness is the perceived luminance [3]. Stevens [4] proposed an expression for the apparent brightness that gives a suitable relationship between luminance and brightness for simple objects. In [5] Krawczyk et al. proposed an operator in order to get an accurate estimation of lightness in real world scenes by a theory called anchoring theory of lightness perception. This method is based on automatic decomposition of HDR (High Dynamic Range) image into framework (consistent areas), where lightness is then estimated by anchoring to the luminance level that is perceived as white, and then the global lightness is calculated.
Image contrast could have more than one definition, but it is usually related to variations in image luminance, see Fig. 1.
Fig. 1: The simultaneous contrast effect: despite the two inner rectangles are of the same shade of gray, but the one to the left appears lighter than the one to the right.
Michelson’s definition [6] for contrast is defined by Eq. (1):
————————————————
𝐶 = 𝐿𝑚𝑎𝑥−𝐿𝑚𝑖𝑛
𝐿𝑚𝑎𝑥+𝐿𝑚𝑖𝑛
(1)
• Yusra Al-Najjar Author is a PhD student in Universiti Tenaga
Nasional, Malaysia (e-mail: yusra.najjar.2011@gmail.com).
• Chen Soong Der is with the College of IT, Universiti Tenaga Nasional, Putrajaya, Malaysia (e-mail: chensoong@uniten.edu.my).
where 𝐿𝑚𝑎𝑥 and 𝐿𝑚𝑖𝑛 are the maximum and minimum
luminance values. Michelson’s definition is sometimes used
when measuring the global contrast of the image.
IJSER © 2013 http://www.ijser.org
International Journal of Scientific & Engineering Research, Volume 4, Issue 11, November-2013 1387
ISSN 2229-5518
Using this method for global contrast is considered 
misleading, because it got affected by random, irrelevant pixels with extreme values. Peli in [6] proposed another definition of local contrast that is suitable for complex images; this method assigns a constant value to every point in the image as a function of the special frequency band where the constant is defined as the ratio of bandpass-filtered image at that frequency to low-pass image filtered on an octave below the same frequency. This method produces a multi-scale representation of the effective image contrast.
The sensation of colour is considered to be an important aspect of the human visual system, and the correct reproduction of colors may help to increase the realistic appearance of the output image. One important feature of the human visual system is the ability to see the level of colors in a bright environment. This capability, measured as colour sensitivity, is reduced in dark environments, as the light sensitive rods take over for the colour-sensitive cone system.
As the luminance level is raised, the cone system becomes
active and colors begin to be seen starting with the long
wavelength reds and moving ahead toward the middle
wavelength greens. Only at relatively high luminance, where short wavelength blue targets start to appear [7]. In colour reproduction, naturalness can be assessed by the mental recollection of the colors of familiar objects, i.e. memory colors. Colorfulness is often considered by the observers when evaluating overall image quality [2]. Correct reproduction of colors can make the output image more real. The success of a colour reproduction lies in how close the reproduced scene is to the original scene.
Usefulness is the accuracy of the visual representation of the image. It could also be defined by the presence of enough information required to identify a certain object in the image. It is how much we can benefit from the image regardless of how natural or good it is. In short words, image should serve a purpose [8].
This attribute is hard to be computed since it balances between different other attributes of the image depending on observers’ point of view. A well balanced image should be bright, colored, clear and no distortion, and the main thing is to be acceptable by human eye.
In this paper we are mainly concerned with naturalness, colorfulness, and contrast attributes in order to find out an image quality index that is close to human perceptual judgment.
Image Quality
Noise Contrast Colourfulness Naturalness of colours
Fig. 2: The attributes that shares most in defining image quality
Following scheme clarifies the relationship between image attributes and their effect over image quality. As seen Image Quality is affected by many attributes either positively or negatively, i.e. it strongly depends on the perceived brightness where bright scenes appear bright and dark scenes appear dim. Contrast should also be reproduced in order to make the resultant image natural. Details and visibility reproduction have their effect as well in image natural appearance.
Fig. 3: The relations between image attributes [7], dashed boxes are the attribute of our concern
The word ‘Colour’ maybe understood in different ways such as a certain kind of light, the effect of light on human eye, or the effect of this light on human mind. Colour is perceptual result of visible light that lies in the range of 380nm, and
750nm of spectrum wavelength, incident upon the retina [9] .
Colour space is a method by which colors are specified,
created, and visualized [9]. They are usually specified by
using three attributes or coordinate which represent its position within a specific colour space. These coordinates do not tell us what the colour looks like, only where it is located within particular colour space. Colour models are 3D coordinate systems, and a subspace within that system, where each colour is represented by a single point.
This colour space is represented by the colors Red (R), Green
IJSER © 2013 http://www.ijser.org
International Journal of Scientific & Engineering Research, Volume 4, Issue 11, November-2013 1388
ISSN 2229-5518
(G), and Blue (B) mainly and is an additive system. The additive system is specified by the chromaticity of its primaries and white point.
RGB values can be transformed into a CIE XYZ colour space by three-by-three matrix transform. The transform is done using the following matrix:
𝑋 0.412453 0.35758 0.180423 𝑅
�𝑌� = �0.212671 0.71516 0.072169� �𝐺 � (2)
Unlike the RGB colour space, each axis of the CIELuv
colour space has different expression ranges. The range 𝐿𝑟𝑎𝑛𝑔𝑒 of the expression for 𝐿∗ component has the 101 steps ranging from 0 to 100, while the 𝑢∗ component has 355 steps ranging from 134 to 220, and the 𝑣 ∗ has 263 steps ranging from -140 to
122.
The colour difference between any two colors in the
CIELuv colour space is calculated using the Euclidian
𝑍 0.019334 0.119193 0.950227 𝐵
distance as follows:
∗
∗ 2 ∗ 2
∗ 2 1⁄2
However, RGB is not very efficient when dealing with
∆𝐸𝑢𝑣 = [(∆𝐿 )
+ (∆𝑢 )
+ (∆𝑣 ) ]
, (7)
real-world images. To generate any colour within the RGB colour cube, all three RGB components need to be of equal pixel depth and display resolution. Also, any modification of the image requires modification of all three planes [10].
Since white is normalized to unity, the middle row sums to unity. To recover the white point, the RGB is transformed into CIE XYZ, then x, y are computed.
There are three types of cone receptors in the retina, and they are acceptable to describe the colour. The set of perceivable colors is only three-dimensional space. The convenient set used for colour measurement, is the CIE 1931 (X, Y, Z) – system adopted by the Commission International de l’Eclairage (CIE) [11]. Each point in the XYZ colour space represents a unique colour perception.
In XYZ colour space, colour components such as Hue and
Chroma, can be represented with x and y chromaticity
coordinates defined as
where ∆𝐿∗, ∆𝑢∗, and ∆𝑣 ∗ represent the difference between two colors in the 𝐿∗ − 𝑎𝑎𝑎𝑎, 𝑢∗ − 𝑎𝑎𝑎𝑎, 𝑎𝑎𝑎 𝑣 ∗ − 𝑎𝑎𝑎𝑎 [4].
In our experiment we intend to measure the naturalness, colorfulness, and contrast of colored images in order to verify the quality of the image. Naturalness is the degree of conformity between human perception and reality world, and will be presented as Naturalness Index CNI; colorfulness on the other hand represents colour vividness degree, and will be presented by the Colorfulness Index CCI; while the Contrast Index is CoI.
Yendrikhovskij introduced a model for best colour
regeneration of natural images which are based on the colour
quality of natural images based on perceived naturalness and
colorfulness of these images [12]. Hasler gives another precise metric for colorfulness, which will be considered in the algorithm we use [13].
𝑎 = 𝑋
𝑋+𝑌+𝑍
, 𝑦 = 𝑌
𝑋+𝑌+𝑍
(3)
In this paper, the CNI will be computed as follows:
1. Converting the RGB image into CIELuv colour space where it is easier to compute the factors needed for
Perceptual uniformity means that the distance between two points in the colour space makes the equal perceived colour difference. The XYZ, RGB, or even HSV colour spaces do not exhibit perceptual uniformity.
There are two other perceptually uniform colour spaces,
the CIElab and the CIELuv. Considering the CIELuv, we find
that the three variables 𝐿∗, 𝑢∗ , and 𝑣 ∗ are defined as follows:
1
116 � 𝑌 �3 − 16 𝑎𝑓 𝑌 > 0.008856
calculating image naturalness such as Chrome, Hue, Saturation, and Luminance. Conversion process passes through XYZ colour space which acts as an intermediate step [14] .
2. Computing the Luminance (L), Hue (H), and
Saturation (S) correspondingly. L* should scale from
0 to 100 for luminance (𝑌/𝑌𝑎) scaling 0 to 1. There
are three meaningful polar parameters that closely
match human visual experience: Chroma, C*, Hue,
L ∗ = �
𝑌𝑛
903.3 𝑌
𝑌𝑛
𝑌𝑛
𝑜𝑡ℎ𝑒𝑟𝑤𝑎𝑎𝑒
′
(4)
ℎ𝑢𝑣 , and saturation 𝑆𝑢𝑣 , as follows:
𝐶 ∗ = (𝑢∗2 + 𝑣∗2 )0.5 (8)
ℎ = arctan �𝑣 ∗ � (9)
𝑢
𝑢∗=13𝐿∗�𝑢′− 𝑢𝑛�
′
𝑛
𝑆𝑢𝑣 =
𝐶 ∗
𝐿∗
(10)
where 𝑢′, 𝑣 ′ are calculated from,
3. Thresholding L and S components so that L values
𝑢′= 4𝑋
′ 9𝑋
that are between 20 and 80 are kept and S values
𝑋+15𝑌+3𝑍 , 𝑣 = 𝑋+15𝑌+3𝑍
(6)
over 0.1 are kept.
′ 4𝑋𝑛
′ = 4𝑌𝑛
𝑢𝑛= 𝑋𝑛+15𝑌𝑛+3𝑍𝑛 , 𝑣𝑛
𝑋𝑛+15𝑌𝑛+3𝑍𝑛
4. Defining three kinds of pixels according to hue
𝑋𝑛, 𝑌𝑛 , and 𝑍𝑛 values are the values of the illuminant, with
𝑌𝑛 equals to 1.
values: 25-70 is called “skin” pixels (pixels that
represent skin colour), 95-135 is called “grass” pixels
(pixels that represent green colour), and 185-260 is
IJSER © 2013 http://www.ijser.org
International Journal of Scientific & Engineering Research, Volume 4, Issue 11, November-2013 1389
ISSN 2229-5518
called “sky” pixels (pixels that represent the blue colour)
5. Calculating averaged saturation values for the “skin” as Saverage_skin, for the “grass” Saverage_grass, and for the “sky” Saverage_sky, and number of “skin” pixels as nskin, “grass” pixels as ngrass, and “sky” pixels as nsky.
6. Calculating local naturalness index CNI values for
the “skin”, the “grass”, and the “sky” pixels as follows:
difference between two images, the original with low contrast image, and the enhanced image in pairs displayed sequentially on a wide screen. Each pair is displayed at a time to be quality judged.
𝑆𝑎𝑣𝑒𝑟𝑎𝑔−𝑠𝑘𝑖𝑛−0.76 2
𝑁𝑠𝑘𝑖𝑛 = exp �−0.5 ∗ �
� �, (11)
0.52
𝑆𝑎𝑣𝑒𝑟𝑎𝑔−𝑔𝑟𝑎𝑠𝑠 −0.81 2
𝑁𝑔𝑟𝑎𝑠𝑠 = exp �−0.5 ∗ �
� �, (12)
0.53
𝑆𝑎𝑣𝑒𝑟𝑎𝑔−𝑠𝑘𝑦 −0.43 2
𝑁𝑠𝑘𝑦 = exp �−0.5 ∗ �
� �, (13)
0.22
7. Calculating global colorfulness index CCI values:
𝑛𝑠𝑘𝑖𝑛∗𝑁𝑠𝑘𝑖𝑛 +𝑛𝑔𝑟𝑎𝑠𝑠 ∗ 𝑁𝑔𝑟𝑎𝑠𝑠 + 𝑛𝑠𝑘𝑦∗ 𝑁𝑠𝑘𝑦
𝑁𝑖𝑚𝑎𝑔𝑒 =
(𝑛
𝑠𝑘𝑖𝑛
+𝑛
𝑔𝑟𝑎𝑠𝑠
+𝑛
𝑠𝑘𝑦
(14)
)
𝑁𝑖𝑚𝑎𝑔𝑒 varies from 0 (the most unnatural image) to 1 (the
most natural image)
8. Computing the colorfulness index 𝐶𝐶𝐶 as in the (15):
𝐶𝐶𝐶 = ((0.2 ∗ 𝑆𝑘 ) + 𝜎𝑘 ), (15)
where Sk is the average saturation of image k, 𝜎𝑘 is
standard deviation, and the value 0.2 is determined
experimentally. 𝐶𝐶𝐶 varies from 0 (achromatic image)
to 𝐶𝐶𝐶max (most colorful image).
9. Compute the contrast of the image according to Peli
[6] as the standard deviation of pixels intensities
Fig. 4: Sample of images used in the experiment
TABLE 1
CATEGORY INTERVALS FOR IMAGE QUALITY
𝐶𝑜𝐶 = 1
∑𝑁−1 ∑𝑀−1(𝐶
− 𝐶)2, (16)
𝑀𝑁
𝑖=0
𝑗=0
𝑖𝑗
Where intensities 𝐶𝑖𝑗 are the 𝑎 − 𝑡ℎ, 𝑗 − 𝑡ℎ elements of a two dimensional image of size M by N, and 𝐶 is the average
intensity of all pixel values in the image.
Computing image quality IQ as a combination of
naturalness, colorfulness, in addition to contrast with
different ratios as in (17):
𝐶𝐶𝑄 = (1 − 𝑤) ∗ 𝐶𝑁𝐶 + 𝑤 ∗ 𝐶𝐶𝐼
max(𝐶𝐶𝐼)
+ 𝐶𝑜𝐶, (17)
Where 𝑤 equals 0.75, experimentally determined.
Fig. 4 shows some of the images used in our experiment.
A group of 30 subjective have been selected to evaluate 42 natural scene images’ quality. A simple training has been done. The experiment took place in a dark room, where each pair of the original and test images have been displayed in front of the subjective.
Each from the 42 contrast-reduced images was processed
using SGHESE [15] with different parameters’ values to
generate one stimulus for each of the following categories.
Forty two pairs of images, thirty subjective made 1260
judgments.
The categories generates five-point intervals ranging from
1 to 100, see Table 1. The observers rated image quality
Table 2 displays the results of applying several methods including the proposed one.
IJSER © 2013 http://www.ijser.org
International Journal of Scientific & Engineering Research, Volume 4, Issue 11, November-2013 1390
ISSN 2229-5518
TABLE 2
The performance of the proposed method CIQ has been assessed and compared with other methods for the forty two colored images used in the experiment. For the comparative study, we use Pearson, Spearman, and RMSE indices as illustrated in.
There are so many image quality evaluating metrics used for gray images, but still we lack image quality metrics IQM to be implemented over colored images in order to get the quality of the image. In our experiment we gathered three attributes of the image: naturalness, colorfulness, and contrast to compute image quality index. We believe that image depends on more than one factor in order to measure its quality.
[1] | T. a. B. F. Janssen, "Predicting the usefulness and naturalness of color reproductions," J. Imaging Sci. Technol., vol. 44, pp. 93-104, 2000. |
[2] | S. W. P. O. Raisa Halonen, "Naturalness and interestingness of test images for visual quality evaluation," Proc. SPIE7867, Image Quality and System Performance VIII, 24 January 2011. |
[3] | E. Adelson, "Lightness Perception and Lightness Illusion," The New Cognitive Neuroscience, vol. 2, pp. 339-351, 2000. |
[4] | M. D. Faichild, Color Appearance Models, vol. 2nd Ed., Chichester: Wiley-IS&T, 2005. |
[5] | M. K. a. S. H. P. Krawczyk G., "Computational model of lightness perception in high dynamic range imaging," Human Vision and Electronics Imaging XI, 2006. |
[6] | E. Peli., "Contrast in Complex Images," Journal of the Optical Society of America, pp. 2032 - 2040, 10 1990. |
IJSER © 2013 http://www.ijser.org
International Journal of Scientific & Engineering Research, Volume 4, Issue 11, November-2013 lSSN 2229-5518
1391
Yusra A. Y. Al-Najjar was born on January 12, 1965 in Zarqa, Jordan. She received her B.E. degree in Computer Science from Kuwait University at 1986, M. Sc. in Computer Science from Al-Balqa'a Applied University at
2008. She is a LECTURER in KSA King Faisal University. Here research interest includes image enhancement, noise reduction, and Networks.
Soong Der Chen (M'2002) was born in October 6, 1973 in Kuala Lumpur, Malaysia. He received his B.E. in Electronics/Computer (1997), M.Sc (2000) and Ph.D (2008) from the Dept. of Comp. and Comm. Engineering in Universiti Putra Malaysia (UPM), Malaysia. Currently he is a senior lecturer in the Dept. of Graphics and Multimedia, College of IT, Universiti Tenaga Nasional (UNITEN). His research interest includes image quality assessment, image enhancement, computer vision and image compression.
IJSER © 2013