Visual Image Quality Assessment - Term Paper Example

COVER PAGE DETAILS HERE… Contents COVER PAGE DETAILS HERE… 1 Image Quality 3 1. Introduction 3 1.1 Challenges of Image Quality Assessment 5 1.2 Motivation of Image quality Assessment 7 1.3 Objectives of Image Quality Assessment 8 2. Image Quality 9 2.1 Sharpness 9 2.2 Noise 10 2.3 Dynamic range 10 2.4 Contrast 10 3. Facial image 11 3.1 Pose 13 3.3 Contrast 16 3.3.1 Weber Contrast 17 3.3.2 Michelson contrast 17 3.3.3 Root Mean Square Contrast 17 3.4.1 Sharpness method 1(Variance approach) 18 3.4.2 Sharpness Method 2 (frequency approach) 19 4.Summary 20 References 22 Image Quality 1. Introduction The field of visual image quality measurement and testing is involved in establishing the fidelity and acuity of images. This is important because images are subject to distortion during their capture, compression, storage and subsequent uses. Distortion suffered by images could make the difference between proper image matching for biometric comparisons or other image applications. This could mean the difference between letting in unauthorized persons into a building that uses facial recognition for access control or a security system letting a profiled terrorist go undetected under the full glare of security cameras. The field of Visual Image Quality assessment seeks to find metrics and algorithms that can be used to reliably detect and asses the quality of images. This could be done dynamically during a video stream for example to allow a system to allocate resources to a stream for optimal quality. A reliable metric can also be used in the development of algorithms for pre-filtering and bit assignment algorithms at the encoder and also used for optimal image reconstruction, post filtering and error concealment algorithms at the decoder end of a system [1]. A reliable image quality measure can also be used as a bench mark for other image processing algorithms to measure distortion from the actual and the resultant efficiency of the algorithm. Currently the most commonly used measure of image quality is the Mean Squared Error which is found by calculating the average of the squared differences of distorted and reference image pixels [1]. Peak Signal to Noise Ratio PSNR is also used as a common metric. These two methods are desirable because they are easy to compute and have clear mathematical and physical implications and meanings [2]. There are three most common methods used to appraise quality of images [1]. The main one is full reference methods commonly abbreviated as FR. In this method the image to be assessed is compared to its equivalent that is considered perfect. The comparison is then easy to measure and assign a scalar value for the image quality. A second method of quality assessment is where there is no reference image to be measured against. This is termed No Reference or “Blind” referencing method, NR for short. This method usually demands that more resources and more elaborate algorithms are used to establish image quality. The final method is in between the first two. It is called the reduced reference method where only a partial reference image is made available and quality assessment has to be made based on the partial ‘perfect ‘ image. Quality Assessment methods try to get consistent quality prediction built by modeling psychological and psycho-visual features on the human visual system abbreviated to HVS or by signal fidelity measures. Sheikh and Bovik [2] in their paper choose to look at image quality assessment problems as being information fidelity problems and seek to quantify the relation between image information and visual quality but mainly attribute the loss of image quality to the distortion process. 1.1 Challenges of Image Quality Assessment Most frame works for image quality assessment are based on a common error sensitivity philosophy [5]. This is as a direct motivation or result of research in human psychological visual science research. Further research into the area shows that the human visual error perceptions vary depending on the on spatial and temporal frequency and directional channels. Based on this premise, Wang and Bovik propose a different approach that models image degradation as distortions as opposed to treating them as errors. Wang and Bovik, summarize the quality measurement models as: two images; a reference image and a test image are subjected to processing followed by channel decomposition that yields two sets of transformed signals. Further scoring based on human sensitivity measures are carried out on the individual channels. The channels are scored for error metrics usually by a Contrast Sensitivity Function (CSF). Visual masking is applied to the new model which serves to hide the errors in the background signal. This is then summed up using an error pooling algorithm to yield a single value. Admittedly the applications and implementations of various IQA methods vary in complexity and intensity but the underlying approach by and large remains as that described here [1]. As Stated before, there are considerable geo-spatial or temporal factors that come into play when the HVS system is used as a model to rank images. Human image perception is affected by different inputs and writing an algorithm that then mimics a faithful image formed in human vision is at the very list a tedious process. To try and even out the differences that the environment has on facial images the ISO/IEC standard 19794-5 has detailed instruction relating to focus, lighting, pose and so on. This aims to form more uniform images in products such as E-passports and so on. If this standard is adhered to it can then be assumed that the difficulties associated with the environmental interferences on image acquisition can be at least evened out but not entirely done away with. Wang and Bovik also explain further that the methods conventionally used in image quality assessment as shown above work on nine assumptions. These assumptions are: 1. It assumes that reference images are flawless in quality 2. The HVS has visual channel responses that can be adequately simulated by a select set of channel transformations and manipulations 3. Contrast Sensitivity Function variance and intra-channel masking effects are the main factors that influence the HVS’s perception on each channel that has been transformed 4. CFS weighting and masking sufficiently portray the magnitude of error and the distortion observed by the Human Visual System, in turn the HSV can be modeled as a non linear function 5. The influence of CSF weighting and masking on different coefficients is negligible 6. The effect of interaction between channels is small enough to be ignored 7. Perceived distortions can be summed up in a monotonically increasing fashion from individual channels and coefficients. 8. Early Vision system holds most responsibility for the for perceived image quality. Higher level processes like pattern matching, image completion and cognitive understanding of that are done by the human brain are less effective 9. Human active visual processes such as attention being paid to an object being viewed, fixation and adaptive adjustment of spatial resolution are not very effective Image quality assessment as a field of research is the meeting point of several research specializations and sciences. These include but not limited to: signal and image processing computer vision visual psychophysics neural physiology information theory machine learning design of image acquisition, communication, and display systems 1.2 Motivation of Image quality Assessment The general quality of an image is chiefly determined by the method used in imaging, the nature of the equipment, and the imaging parameters therein set by the equipment operator [6]. This could be images for a passport, access control, medical imaging or some other bio-metric application. Parameters of Image quality assessment such as contrast on a grey scale MRI or X-ray image could mean the difference between a malignancy being identified by a medical practitioner, or the patient going on to suffer more intrusive methods to find the said malignancy. While this paper is biased more towards human facial images and their quality measurement the same principles could still be used in other imaging applications. As mentioned earlier Image Quality Assessment has three main applications by the work of Wang, Bovik et Al [3-5]. The three applications are [7]: To benchmark video and image processing algorithms, devices and systems in aspects such as de-noising, de-blurring or compression for storage. If systems and algorithms meet a particular threshold set by agreed upon IQM metrics they can be used and set as conventions for general application and rejected if they do not meet set benchmarks. IQM measures can be used to monitor video streams or other dynamic image feeds. The quality of these feeds can be checked against set metrics to evaluate their quality. This allows the system to allocate resources such as band width and processing power evenly and optimally to such streams and feeds. IQM metrics can be embedded into a system at both the encoder and decoder end. This will set benchmarks and thresh holds for compression and transmission algorithms at the encoder end. At the decoder the metrics could be used to find an acceptable de-noising, decompression and de-blurring standards. This would inevitably obtain optimal peak signal to noise ratios PSNR. 1.3 Objectives of Image Quality Assessment The main objective of image quality assessment is to find a metric that can be used to measure the quality of an image in a manner that is reliable and consistent. This yields further applications and objectives such as being able to predict the desirability of an image to human eyes which are most times the ultimate recipients of images in most image processing systems. This is also the reason of using the HSV as the reference for image quality measurement. Since most image quality assessment methods are based on the HVS [8] this paper shall seek out to explain the most commonly used methods and their challenges. This paper shall also analyze newer methods of IQA that treat image degradations as structural differences as opposed to signal errors. This approach is proposed by Wang and Bovik [1], [5-8]. Other applications of the techniques discussed in this paper go beyond the realm of image quality for the human eye but also spill over into image recognition. It is possible to reject poor images at the enrollment stage or even filter out distortions to an extent where the image can then be reliably used in image recognition if proper metrics for image quality are used [9]. This paper will also focus on ISO standards for facial image acquisition (ISO/IEC standard 19794-5). This standard aims to create uniformity and standards for images used in electronic or other systems but use the human facial images as one of the inputs. This paper shall touch on all objective image quality assessment methods that is Full Reference (FR), Reduced Reference (RR) and No Reference (NR). 2. Image Quality An image capture device works by receiving and recording light on a light sensitive plane. This could be the typical Silver Nitrate film, a digital light sensitive surface such as in a digital camera or other films such as an X-ray film. In actual sense all the methods of light capture digital or otherwise are still just approximations of the total amount of light coming off the object being photographed. All just advancements based on the pinhole camera or the human eye. Most image capture images further undergo post processing. This could be as basic as the colour and resolution differences that are formed by two different developers from the same film in traditional photography or image enhancement methods applied to images on a digital camera. In the case of the human eye, factors such as auto-completion and cognitive processes serve as post processing. Pre-processing factor such as lens focus, aperture, and curvature serve to distort an image before it has been captured on the image plane, in this case a film or eye’s retina. By defining quantifiable image quality measures or standards it allows for an image to be measured for its deviation from an ideal image. It also allows for benchmarks and thresholds to be set for image quality. Image quality metrics can be absolute or subjective, a good example of a subjective measure is if a human observer was to judge image reduction of noise quality from an image the observer’s opinion would be formed mainly based on how much the noise obscures the information that the viewer seeks rather than the overall strength of noise reduction [13], [6]. Image quality factors for the purposes of IQA that is based on HVS systems include but are not limited to: 2.1 Sharpness Sharpness can be termed as the total amount of detail an image conveys. Sharpness is directly determined by the system’s lenses and light sensor or receptor. These two can be broken down to focal length, distance form image, aperture and axis of the lens. The light sensor will directly affect resolutions most commonly pixel count and anti-aliasing filter. During image capture, sharpness can be affected by shaking, focus accuracy and thermal effects. There are several image sharpening algorithms and techniques available to increase the sharpness of an image. Over sharpening however causes contrast edges to be too “hard” and causes the appearance of halos [13]. 2.2 Noise Noise is defined as random variations in image density. This becomes apparent on the image as grainy presentations. It is caused by the inevitable effects of the image formation process usually the packet (photon) nature of light and heat energy which light sensors sometimes pick up. This becomes manifest as noise in images. Noise reduction algorithms and software can be used to reduce noise but this is often a trade off to sharpness since the smoothing that achieves noise reduction can obscure fine and low contrast parts of an image. 2.3 Dynamic range Dynamic range is the variety of luminance levels that a sensor can capture often measured in exposure value (EV) or zones. This is usually directly related to noise as low dynamic ranges usually yield high noise. 2.4 Contrast Contrast also known as gamma, is a non-linear operation used to encode and decode luminance. High contrast often sacrifices dynamic range and the reverse is also true. 3. Facial image While facial images are also susceptible to the degradations discussed in the section 2 of this paper there are also other factors that are unique to facial images. ISO/IEC standard 19794-5 seeks to outline conventions and standards for taking consistent facial images. According to Griffin [18] there are 4 basic types of facial images. Basic images, frontal images, full frontal images and token images. Basic images are images that are not necessarily taken to fit any set standards for lighting, scene, photography and image type. Frontal images are images that adhere to set standards for facial recognition system and enrolment of facial images. This group further yields the other two types of images. Full frontal images are images taken to set standards, they include full view of the subjects hair, neck and shoulders. Token images are images that are taken to set ISO standards. The standard specifies the precise positioning of the eyes, the head size, pose and scene setting standards. Token images run from the crown of the head to the chin, neck and shoulders are not visible. This image is deal for storages since it takes up less space. As is apparent token and full frontal images are then the most used for image recognition and match up to image quality assessment standards. The ISO/IEC standard 19794-5 has guidelines related to pose, exposure, background, lighting and illumination. It is the opinion of the author that some of the standard guidelines put forward in ISO/IEC standard 19794-5 or open to subjective interpretation. For example focus guideline is that focus needs to be sharp. An image may be deemed to have good sharpness by one human observer, fair or poor by two other observers. Other parameters such as expressions are also quite subjective and may vary with one person due to individual temperament and mood. Some of the guidelines for the ISO/IEC standard 19794-5 are discussed under the following sub-topics. 3.1 Pose Pose should ideally have the individual looking squarely at the camera. Angles of pose are described by pitch, roll and yaw. Pitch describes a lateral movement, which is a movement away from center line. Roll describes front to back movements and yaw describes the left to right movement maintaining the vertical axis of the heard. The ideal situation is 0 degrees for all the parameters though the acceptable range is +30 or -30 degrees of roll. Eyes should be maintained open, held in Frankfort, with no hair covering the eyes. Red eye effect should be edited out. Glasses may be worn if they are for corrective reasons and the individual cannot do without them. However frames of the glasses should not cover the subject’s eyes. Glasses should also not be reflective and should allow the viewer to see the individual’s eyes. If the individual wears an eye patch for medical reasons or because the person has different eye colour in either eye an eye patch can be allowed but it must obscure as little of the face as possible. Head coverings are not allowed with the exception of those worn for religious reasons. In that case however the face should be visible from the bottom tip of the chin to the top of the forehead. Either side of the subject’s face should also be clearly visible. The individual’s face must be the only object in the picture and must hold a neutral expression. Frowning, smiling, grimacing or other facial expressions are not allowed. The mouth should also remain closed to the best of the subject’s ability. Fake beards are not allowed in pictures. The skin must show the individual’s natural dermis and no noticeable makeup should be worn. Facial jewelry and other jewelry such as earrings should also be taken off. The subject’s face should take 70% to 80% of the picture. The width of the head must be more than five sevenths (5/7) of the width of the photograph. The height of the head should not exceed 80% of the photograph’s height. The background should be light coloured and not reflective. Shoulders should also be visible in the photograph. Inter eye distance must not be less than 120 pixels in digital photographs for the ANSI standard. The ISO standard however specifies that the distances between the subjects pupils be no less than one quarter the width of the resultant image. Inter eye distance can be measured manually and adjusted by changing the distance to the camera. Closer to the subject increases inter-eye distance and further decreases inter eye distance. Inter pupil distance can be digitally adjusted by enlarging the image using a bi-cubic interpolation technique. Other variables such as pitch, roll and yaw can be adjusted using different algorithms and techniques. Roll alignment for example is tied to the angle between pupils. A straight line is plotted between the two eyes. The skew of the line is determined using trigonometric techniques and appropriate rotation of the whole image is then determined from the skew of the angle between the eyes. Only one person should be visible in the photograph. 3.2 Lighting The standards for lighting by ISO seek to provide a uniformly lit image, devoid of degrading shadows and presenting of the subject’s true skin tone. Since the background affects or is affected by lighting of the image, it is addressed in the ISO/IEC standard 19794-5 standard. The standard states that a lightly coloured non-reflective background should be used. The phrase “lightly coloured” is however open to individual interpretation and bias. India’s standard for image acquisition for e-government applications [19] and guides seek to do away with this ambiguity by expressly stating that a white background should be used. Where there is not enough distinction between the individual’s face or hair with the image’s background a light grey background is also permissible. In this case the grey background can be up to 18% grey levels. Natural lighting shall uniformly illuminate the face of the subject. The whole of the subject’s face from the crown of their head to the tip of their chin shall be visible and have no shadows. More than one light source may be used for this [6]. Care should be taken to avoid hot spots on the subjects face. Hot spots are areas of the face that appear to be shinning or are more brightly lit than the rest of the face. The iris and pupils of the eyes should be clearly visible. There should be no shadows on the sockets of the eyes caused by the ridges and brow of the face. If eye glasses obstruct the view of the pupils or iris they should be taken off. 3.3 Contrast Contrast has various definitions. There are different ways of understanding the concept that is contrast. Some factor in colour and others do not. This part of this paper shall endeavour to sample the various approaches and definitions for contrast. In its simplest form contrast can be simply termed as the variance between minimum and maximum pixel intensity in a given image. It is the ability of a visual system to differentiate between bright and dim areas of a static image. It is what makes a particular image or part of an image visible and distinguishable from its surrounding. An alternate definition is: Contrast also known as gamma, is a non-linear operation used to encode and decode luminance. By the HVS, the human eye decodes contrast in typical band pass filter methods. This is to mean that contrast has to surpass a certain threshold for the human eye to perceive it [22]. The high drop off rate is an explicit illustration of the limitations of the HVS. The human eye is therefore unable to distinguish detail that is apparent to many digital devices. The ability to distinguish contrast and by effect pick up detail is called contrast sensitivity. Contrast sensitivity in actual effect is the ability to distinguish areas of different luminosity. This alternate definition then brings out the relation between luminosity and contrast. If contrast is exaggerated the image starts to have a sort of strobe lighting effect on it. Since there are varying definitions of contrast; some including colour and others not, it creates an inconvenience in understanding and researching the topic [22]. The problem is further compounded in application of solutions and comparisons of the works of different researchers. One of the ways to understand the interpretation and weakness of the HVS is using the Weber-Fechner law. This is a general law of psycho-physics which seeks to quantify the human response to given stimuli. The general law is that: the “Just noticeable difference” in a given stimuli is directly proportional to the magnitude of the stimuli [22]. 3.3.1 Weber Contrast Based on the Weber-Fechner law one of the methods of finding contrast is the Weber contrast. In Weber’s contrast, contrast is calculated by the difference of the luminosity of an image and the luminosity of the image’s background divided by the luminosity of the background. This allows one to calculate how different an object glows compared to its own background there by bringing out our initial definition of contrast. It’s most common application is on small features on a big and uniform background. 3.3.2 Michelson contrast Michelson is also known as a visibility measure. It is most commonly used in pattern images where light and dark patterns take up similar portions of space on the image; a good example is sine wave gratings. This is calculated by dividing the difference of the highest luminosity and lowest luminosity divided the double the mean of the two. This formula efficiently calculates a quantifiable value for periodic functions and is also referred to as the modular function of a periodic signal. The modular function mf refers to the contrast of a modular wave as compared to its average value. 3.3.3 Root Mean Square Contrast Root mean square contrast or RMS in short can be termed as the standard deviation of the intensity of pixels in an image. This method is not dependent on the angular frequency of content or the distribution of contrast across the image. This algorithm is most useful in calculating luminosity or adjusting luminosity of an image digitally manipulated. 3.4 Sharpness Sharpness just like contrast increases the detail with which an image is viewed. Unlike contrast that is more concerned with luminosity sharpness is more concerned with edge contrast. In photography this is often subjectively (and sometimes incorrectly) [22] termed as acutance. Acutance is correctly termed as a subjective perception of sharpness. This can be altered by adjusting edge contrast which makes an image look sharper. This however has its draw backs as it causes halos if over done. In actual sense perceived sharpness is a combination of acutance which can be altered digitally and the resolution of an image which cannot be adjusted digitally. One of the most common methods of adjusting for sharpness in digital photography is un-sharp masking. It is called that because it uses a blurred negative image to create a mask for the actual image which creates perceived sharpness. This is done by controlling three main variables of the negative blurred image: radius, amount and threshold. Radius is the width of the edges to be enhanced. Greater radii make for wider edge rims. Amount can be thought of as the intensity of the contrast of the edges. This affects how much darker or lighter the edge rims become. It however leaves the radii unchanged. This value is measured in percentages. The threshold value affects the minimum amount of brightness to be adjusted and how far apart tonal values have to be for the filter takes effect on them. Un-sharp masking can sometimes be used with large radii and small percentage amounts. This yields an image with increased local contrast. This method is called local contrast enhancement. 3.4.1 Sharpness method 1(Variance approach) One method for calculating sharpness of an image is suggested by Ratha et al [25] who takes a variance based approach to the challenge. In this method an image is subdivided into regions based on its height and width. The regions’ variance is then individually computed and if the variance passes a predefined bench mark the region’s sharpness is good enough and is considered a ‘good region’. The total variance is then scored as a ration of the total number of ‘good regions’ to the total number of regions. This is can be written as the below equation. 3.4.2 Sharpness Method 2 (frequency approach) The second method of calculating sharpness suggested by Sang et al [24] utilizes IDCT and DCT. This method is informed by the knowledge that low frequency information in an image directly corresponds to the general shape of components while the high frequency information matches the details of the skin. Sang et al show clearly that a blurred image looses high frequency information. Sang et al first use a high DCT operations to get the frequency domain of a given image (I) . The coefficients which occupy a certain pre-determined ratio of total energy is conserved after which it is further inverse transformed back into the image space using IDCT (inverse discrete transform). This yields a recovered image R. The sharpness is then calculated by finding the difference of the input image and the recovered image. Sharper images will yield greater differences. This can be expressed in the below equation: In experiments by Sang et al [24] it becomes apparent that the frequencies based approach yields far more precise results for measuring sharpness. It is also easier to model mathematically. It is however an approach that does not stem from HVS and thus would be difficult to model standards off of. 4.Summary In this paper we analyze the trends and standards in image recognition, the imperativeness of facial image quality to appropriate biometric matching and enrolment processes. We analyze the ISO standards for image quality and look at a few other standards. It also comes across that many of the standards set by ISO and ANSI are open to subjective interpretation. This may be due to a person’s individual cognisance of the standard specification or due to errors in HVS image perception. We further draw out the weaknesses of the HVS model as a reference point or basis for design of biometric system. This paper does not however completely cast aside the usefulness of the HVS as a reference for understanding the dimensions of image quality. It is clear that most images are taken t be observed by human beings and as such the HVS has been given prominence by many researchers and experts in the field of image quality measurement. It is also important since it allows researchers and engineers to understand how images are perceived and thus how quality is perceived. The weakness in HVS referencing and bench marking is made apparent in the review of two methods of measuring sharpness. While the variance approach by Hatha et al is useful and has been applied often the method of frequencies proposed by Sang et al proves more accurate. This is a particular example of where the HVS is useful for understanding the perception of the image but when it comes to developing algorithms the method proves insufficient due the innately subjective nature of human vision. HVS methods are still important because more often than not the enrolment and capture of images is done manually by human operators. For the operators to understand and decipher images and their apparent quality standards and definitions have to be based almost exclusively on the HVS system. References [1] Z. Wang et Al , “Image Quality Assessment: From Error Visibility to Structural Similarity,” IEEE Trans. Image Process, vol. 13, no. 4, pp.1-12, Apr. 2004. [2] R. H. Sheikh and A. C. Bovik, “Image Information and Visual Quality,” IEEE Trans. Image Process, vol. 15, no. 2, pp.1-3, Feb. 2006. [3 ] R. H. Sheikh, M. F. Sabir and A. C. Bovik, “A Statistical Evaluation of Recent Full Reference Image Quality Assessment Algorithms,” IEEE Trans. Image Process, vol. 13, no. 4, pp. 1-12, Mar. 2006. [4] X. Gao et Al, “Standardization of Face Image Sample Quality,” National Lab of Pattern Recog., Beijin, Rep. 230026, 2007. [5] Z. Wang and A. C. Bovik, “Why Is Image Quality Assessment So Difficult?,” Lab for Image and Video Pattern Recog., Austin, Tx. [6] P. Sprawls. (2004). Image Characteristics and Quality [Online]. Available: http://www.sprawls.org/ppmi2/IMGCHAR/ [7] K. Thung and P. Raveendran, “A Survey of Image Quality Measurement,” Dept. of Elec. Eng. University of Malaya, Kuala Lumpur, Malaysia. [8] D. M. Chandler. (2013). Seven Challenges in Image Quality Assessment: Past, Present and Future [Online]. Available: http://www. hindawi.com/journals/isrn/2013/905685/ [9] Z. Wang, “Applications of Image Quality Assessment Methods,” IEEE Signal Processing Mag., Vol.28, pp. 137-143, Nov. 2011. [10] Z. Wang and A. C. Bovik, “Mean squared error: Love it or leave it? A new look at signal fidelity measures,” IEEE Signal Processing Mag., vol. 26, pp. 98–117, Jan. 2009 [11] G. Piella, “New quality measures for image fusion," in Proc. Int. Conf. on Information Fusion, 2004, pp. 542-546. [12] VQEG. (2000, Mar.) Final report from the video quality experts group on the validation of objective models of video quality assessment. [Online]. Available: http://www.vqeg.org [13] N. S. Kopeika, A system engineering approach to imaging, WA: SPIE, 1998. [14] R. V. Hsu et al., “Quality Assessment of Facial Images,” in Biometrics Symposium, [2006] [15] H. R. Sheikh, Z. Wang, L. Cormack, and A. C. Bovik, “LIVE image quality assessment database, Release 2,” 2005, available at http://live.ece.utexas.edu/research/quality. [16] A. M. Eskicioglu and P. S. Fisher, “Image quality measures and their performance,” IEEE Trans. Communications, vol. 43, no. 12, pp. 2959–2965, Dec. 1995. [17] O. Yuridia, G. Castillo, “Survey About Facial Image Quality,” Fraunhofer Institute for Computer Graphics Research, 2004, Darmstadt, Germany. [18] P. Griffin. (2005). Understanding The Face Image Format Standards. [Online]. Available: http://biometrics.nist.gov/cs_links/standard/archived/workshops/workshop1/presentations/Griffin-Face-Std-M1.pdf [19] Face Image Data Standard for e-Governance Applications in India, eGov.BIDS:01, 2010. [20] ISO 19794-5: 2005(E) Information Technology – Biometric data interchange formats – Part 5 Face image data, ISO, 2005 [21] W. F. Campbell, (1968). "Application of Fourier analysis to the visibility of gratings". Journal of Physiology [22] N. Lavie and S. Cox, “On The Efficiency of Visual Selective Attention: Efficient Visual Search leads to inefficient Distracter Rejection,” APA., London, 1997 [23] A. Michelson,. (1927). “Studies in Optics.” Uni. of Chicago Press. [24] J. Sang, Z. Lei, S. Z. Lei, “Face Image Quality Evaluation for ISO/IEC Standards 19794-5 and 29794-5,” Center for Biometrics and Security Research, no. 4, pp.3-10, Apr. 2009. [25] Ratha, N.K., Chen, S.Y., Jain, A.K. “Adaptive Flow Orientation-based Feature Ex-traction in Fingerprint Images. Pattern Recognition .” pp 2-14, Mi, Mi State Uni, 1995. [26] ISO/IEC 19785 (all parts), Information technology — Common biometric exchange formats framework [27] ISO/IEC 19794-1, Information technology — Biometric data interchange formats — Part 1: Framework [28] ISO/IEC-15444-1 Information technology – JPEG2000 image coding system : Core coding System [29] H. R. Sheikh, Z. Wang, A. C. Bovik, and L. K. Cormack, “Image and video quality assessment research at LIVE,” http://live.ece.utexas.edu/research/quality/. [30] Z. Wang and A. C. Bovik, “Embedded foveation image coding,” IEEE Trans. Image Processing, vol. 10, pp. 1397–1410, Oct. 2001. [31] C. M. Privitera and L. W. Stark, “Algorithms for defining visual regions-of-interest: Comparison with eye fixations,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 22,pp. 970–982, Sept. 2000. [32] A. B. Watson, “Visual detection of spatial contrast patterns: Evaluation of five simple models,” Optics Express, vol. 6, pp. 12–33, Jan. 2000. [32] Z.Wang and A. C. Bovik, “A universal image quality index,” IEEE Signal Processing Letters, 2002. [33] E. P. Simoncelli, W. T. Freeman, E. H. Adelson, and D. J. Heeger, “Shiftable multi-scale transforms,” IEEE Trans. InformationTheory, vol. 38, pp. 587–607, 1992. [34] B. Girod, “What’s wrong with mean-squared error,” in DigitalImages and Human Vision (A. B. Watson, ed.), pp. 207– 220, the MIT press, 1993. [35] A. B. Watson, “Visual detection of spatial contrast patterns: Evaluation of five Simple models,” Optics Express, vol. 6, pp. 12–33, Jan. 2000. [36] T. N. Pappas and R. J. Safranek, “Perceptual criteria for image quality evaluation,” in Handbook of Image and VideoProcessing (A. Bovik, ed.), Academic Press, May 2000. [37] V. Strela, J. Portilla, and E. Simoncelli, “Image denoising using a local Gaussian scale mixture model in the wavelet domain,” Proc. SPIE, vol.4119, pp. 363–371, 2000. [38] K. Somarsundaram and N. Palaniappan, “Personal ID Image Normalization Using ISO 19794-5 Standard For Facial Recognition Improvement,” Ghadiradhgram Rural Institute, Rep. 230026, 2007. [39] R. V. Hsu, J. Shah and B. Martin, “Quality Assessment of Facial Images,” Identic Corporate Reseach, NJ, 2006. [40] A. P. Bradley, “A wavelet visible difference predictor,” IEEE Trans.Image Process., vol. 5, no. 5, pp. 717–730, May 1999. [41] A. P. Bradley, “Gabor Feature Based Classification Using the Enhanced Fisher Linear Discriminant Model for Face Recognition,” IEEE Trans.Image Process., vol. 11, no. 4, pp. 468–478, Apr 2002. Read More