Showing papers presented at "Color Imaging Conference in 2012"

The Role of Bright Pixels in Illumination Estimation

Abstract: The White-Patch method, one of the very first colour constancy methods, estimates the illuminant colour from the ma ximum response of three colour channels. However, this simpl e method has been superseded by advanced physical, statistic al and learning based colour constancy methods. Recently, a few re search works have suggested that the simple idea of using max imum pixel values is not as limited an idea as it seems on first glance. These works show that in several situations some man ipulations can indeed made it perform very well. Here, we exten d the White-Patch assumption to include any of: white patch, h ighlights or light source; let us refer to these pixels in an imag e as the “bright” pixels areas. We propose that bright pixels are surprisingly helpful in the illumination estimation process. In this paper, we investigate the effects of bright pixels on several current colour constancy algorithms. Moreover, we describe a simple framework for an illumination estimation me thod based on bright pixels and compare its accuracy to well-know n colour constancy algorithms applied to four standard datas e s. We also investigate failure and success cases, using bright pixels, and propose desiderata on input images with regard to th e proposed method. Introduction Illumination estimation, which is the main step in colour constancy processing, is an important prerequisite for man y computer vision applications. One of the first colour constancy methods, the so-called White-Patch or Max-RGB method estimates the illuminant colour from the maximum response of three col our channels [24]. With the advent of newer and more precise colo ur constancy methods such as Grey-World [3], Gamut Mapping [12 ], Grey-Edge [31] and many other complex methods (refer to [21] for an overview), few researchers or commercial cameras use the White-Patch method. On the other hand, recent research such as that on perceptual color contrast enhancement by Choudhu ry and Medioni [4] or on the “rehabilitation” of MaxRGB by Funt and Shi [15] propose that a local mean calculation such as local blurring as a preprocessing step can significantly impro ve the performance of this simple method, consisting of simply find ing the maximum in each colour channel. Simply put, these works propose it is advantageous to calculate the max of a local mea n image. Recently, Drew et al. [6] found analytically that the geomet ric mean of bright (generally, specular) pixels is the optim al estimate for the illuminant, based on a standard dichromatic m odel for image formation (which accounts for the matte and highli ght appearance of objects). This work proposes that in the prese nc of specular highlights the “mean of the max” is the best illuminant estimate, in contradistinction to previous works [1 5, 4] which say it is the “max of the mean.” The analytical approach [6] claims performance comparable with very complex colour con stancy methods despite its simplicity. The bright areas of images can be white surfaces or light sources as well as highlights and specularity, and all are he lpful in the illumination estimation process. Highlights and whi te surfaces both tend to have the colour of light in ideal condition s for dielectric materials such as plastic. In this paper, we investigate the effects of bright pixels on different colour constancy algorithms. We describe a simpl e framework for an illumination estimation method based on br ight pixels and compare its accuracy to well-known colour consta ncy algorithms applied to four standard datasets. We also inves tigate failure and success cases, using bright pixels, and draw con clusions on input images with regard to the proposed method. Illumination Estimation by Specular reflection In specular reflection, light from a single incoming directi on is reflected into a small cone of outgoing directions. This co ntrasts with diffuse reflection, where light is partially absorbed a nd partially scattered within the surface material. Areas of imag es that are specular tend to be bright. Moreover, the spectral power distribution (SPD) of specular reflections is the same as the ill um nation’s SPD, within a Neutral Interface Reflection (NIR) [25] condition, which mostly obtains for the surfaces of optically i nhomogeneous objects (such as ceramics, plastics, paints, etc.) ; however it does not always hold for the surfaces of optically homogen eous objects (such as gold, bronze, copper, etc.) [20]. These pro perties make specular reflection, which is usually in bright are as of image, an appropriate tool for estimating illumination. Ma ny illumination estimation methods derive from the dichromatic mo del for specular reflection proposed by Shafer [28]. Klinker et al. [23] showed that when the diffuse colour is constant over a surface, the colour histogram of its image fo rms a skewed-T shaped distribution, with the diffuse and specula r pixels forming linear clusters. They used this information to esti mate a single diffuse colour. Therefore in order to use this princ iple, their approach needed to segment an image into several regio ns of homogeneous diffuse colour. Lee [26] proposed a method which uses specularity to compute illumination by using the fact that in the CIE chromatic ity diagram [33] the coordinates of the colours from different p oints from the same surface will fall on a straight line connected t o the specular point. This is the case when the light reflected from a uniform surface is an additive mixture of the specular compo nent and the diffuse component. This seminal work initiated a sub stantial body of work on identifying specular pixels and using th ese to attempt to discover the illuminant [27, 30]. Another appr oach extending these algorithms is to define a constraint on the po ssible colours of illumination, making estimation more robust [8, 9]. Extending the White Patch Hypothesis The White-Patch hypothesis is essentially that there is always a white surface in the image. Let us extend this assumpti on to include any of: white patch, specularities, or light sour ce (or an effective white, e.g. a bright yellow and red pixel which c ombined have the same max R, G and B as a white patch). We also use the termgamutof bright pixels, in contradistinction to maximum channel response of the White-Patch method, which typi cally deals only with the brightest pixel in the image. Obvio usly, using a single pixel or very small area is noisy and not robust . Since are we dealing with bright pixels we need to be very careful about clipped pixels, i.e. pixels where the light re flection exceeds the dynamic range of the camera. Here for each colour channel we remove pixels which exceed 90% of the dynamic range. Then we simply define bright pixels as the top T% of luminance given byR+G+B. To investigate the utility of this assumption, we carry out a simple experiment. We check whether or not the actual illuminant colour falls inside the 2D gamut of bright pixels. We find that the actual illuminant colour falls in the 2D gamut of the top 5% brightness pixels of each image for the SFU Laboratory Dataset [2] for 88 .16% of images, in 74 . 7% of images for the ColorChecker dataset [29], and in 66 .02% of images for the GreyBall Dataset [5]. Fig. 1 shows the 2D gamut in chromatici ty space{r,g} = {R,G}/(R+G+B), with the top-5% brightness pixels in green. The actual measured illuminant is shown as a red star. Clearly, as Fig. 1(c) shows, with no supporting evi d nce it may happen that the illuminant does not fall within the bri ght region. 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8

Aluminum Induced Immunoexcitotoxicity in Neurodevelopmental and Neurodegenerative Disorders

Abstract: A great deal has been learned about the neurotoxicity of aluminum over the past two decades in terms of its ability to disrupt cellular function. Newer evidence suggests that a more central pathophysiological mechanism may be re- sponsible for much of the toxicity of aluminum and aluminofluoride compounds on the brain. This mechanism involves activation of the brain's innate immune system, primarily the microglia, with a release of neurotoxic concentrations of ex- citotoxins and pro-inflammatory cytokines, chemokines and immune mediators. A large number of studies suggest that excitotoxicity plays a significant role in the neurotoxic action of a number of metals, including aluminum. Recently, re- searchers have found that while most of the chronic neurodegenerative effects of these metals are secondary to prolonged inflammation, it is the enhancement of excitotoxicity by the immune mediators that is responsible for most of the metal's toxicity. This enhancement occurs via a crosstalk between cytokine receptors and glutamate receptors. The author coined the name immunoexcitotoxicity to describe this process. This paper reviews the evidence linking immunoexcitotoxicity to aluminum's neurotoxic effects.

Three-Dimensional Porous LiFePO4: Design, Architectures and High Performance for Lithium Ion Batteries

Abstract: The olivine-structured lithium ion phosphate (LiFePO4) is one of the most competitive candidates of cathode materials for the sustainable lithium ion battery (LIB) systems. However, the major drawback of olivine-structured LiFePO4 is the poor intrinsic electronic and lithium ion conductivities arising from the lack of mixed valency and the one- dimensional lithium ion diffusion, which influence its high electrochemical performance, especially high rate capability. Nano-structured LiFePO4 materials offer a potential solution to enhance surface-to-volume ratio and reduce transport length for mobile charges, but they have high interfacial energy, aggregate easily and need more agglutinant in electrode, which seriously impact the electrochemical performance and practical applications of LiFePO4. Furthermore they con- tinue to experience limitations as energy and power requirements escalate with the evolution of technology. Recently, three-dimensional (3D) porous LiFePO4 architectures have been widely designed and studied. This has led to increased in- terest in the development of cathode materials and processing capabilities necessary to enable high-performance, next- generation LIB system that can deliver large amounts of energy at high rates. In this review, we focus on 3D porous LiFePO4 architectures for high power LIBs, summarize and discuss its structure, synthesis, electrochemical behaviors, mechanism, and the problems encountered in its application. The major goal is to highlight the recent progress of 3D po- rous LiFePO4 architectures with high rate capability, high energy density and application.

Realtime Estimation of Illumination Direction for Augmented Reality on Mobile Devices.

Abstract: Augmented reality simulations aims to provide realistic blending between real world and virtual objects. One of the important factors for realistic augmented reality is correct illumination simulation. Mobile augmented reality systems is one of the best options for introducing augmented reality to the mass market due to its low production cost and ubiquitousness. In mobile augmented reality systems, the ability to correctly simulate in realtime the illumination direction that matches the illumination direction of the real world is limited. Developing a mobile augmented reality systems with the ability to estimate illumination direction presents a challenge due to low computation power and dynamically changing environment. In this paper, we described a new method that we have developed for realtime illumination direction estimation for mobile augmented reality systems, using analysis of shadow produced by a reference object that doubles as a 3D augmented reality marker. The implementation of the method could estimate the direction of a single strong light source in a controlled environment with a very good degree of accuracy, with angular error averaging lower than 0.038 radians. The current implementation achieved 2.1 FPS performance in a low-end Android mobile device, produced proper estimation within 15 seconds using a uniform surface, and demonstrated scalability potential. Background An augmented reality system aims to add simulated virtual objects, for example computer-generated 3D objects, to a real world scene, and presents the combined scene to the user. This combination of the real world and virtual objects augments the user’s perception of the real world, and offer a new method of interacting with his or her environment. Augmented reality is used in multiple fields, including medical, manufacturing, entertainment and military [1]. Taken in a larger context, Milgram et al. [2] positioned augmented reality as an intermediate between a real world environment and a virtual reality environment. They also define real world environment and virtual reality environment as polar opposites in what they called a reality-virtuality continuum. Mobile Augmented Reality One important aspect of an augmented reality system is portability [1]. An augmented reality system with high portability would give the user greater degree of liberty during his or her interaction with the augmented environment. Based on the positioning of its display, two major categories of augmented reality system that have relatively high degrees of portability are head-worn augmented reality and hand-held augmented reality [3]. From these two categories, the hand-held augmented reality systems are currently considered to be the best for introducing augmented reality to the mass market due to its low production cost and ease of use [3]. The ubiquitous proliferation of smart mobile devices that could be used as hand-held augmented reality systems also helps the ease of adoption by potential users. Registration and Tracking To properly align the simulated virtual object with the real world scene and produce a consistent simulation, an augmented reality system needs to be able to detect and track its position and orientation with respect to the real world. This is typically performed using markers with known shapes or textures [4]. Augmented reality markers are typically two-dimensional markers, but some libraries provide the ability to use three-dimensional markers. Recent development using feature extraction and recognition provides a new markerless approach for augmented reality tracking, although this approach is still under development [5]. Reproduction Fidelity Tracking and alignment is required to provide overlay information. However, if the objective is to make the virtual object appear integrated into the real world scene, we must embed the virtual objects in the real world [2]. To improve the user experience the system must provide more realistic blending between the virtual objects and the real world. The degree of this blending could be represented as reproduction fidelity [2]. Low reproduction fidelity is exemplified by having a simple wireframe simulation of the virtual objects, while higher reproduction fidelity takes into account more complex factors for the simulation such as shading, texture, transparency, and illuminations, with the ultimate target of producing simulations with photorealistic fidelity. Current state of the reproduction fidelity for mobile augmented reality systems is limited to texture, shading, and transparency simulation of the rendered virtual objects. The ability to properly simulate the illumination condition of the virtual objects to match the illumination condition of the real world is still limited, mainly due to the inability to estimate the illumination conditions of the real world environment. Without this estimate, simulating the illumination condition in an augmented reality systems using abstract parameters would risk producing less realistic simulations. In order to address this issue, several methods are proposed and developed to estimate the illumination conditions and incorporate the estimate in an augmented reality system [6, 7, 8, 9, 10, 11, 12, 13, 14]. However, none of the explored methods is applied to mobile augmented reality systems, and most of them requires off-line processing that precludes real time application 20th Color and Imaging Conference Final Program and Proceedings 111 of the illumination estimation on mobile devices. In this paper, we developed a illumination estimation method that is suitable for simulating real time illumination on mobile augmented reality systems. Illumination in Augmented Reality The first issue in understanding illumination in augmented reality is understanding the types and properties of illumination sources that could be simulated in an augmented reality system. Since the simulated part of an augmented reality system is essentially a virtual reality simulation, types of illumination sources that could be simulated in an augmented reality follow the types normally found in a virtual reality. Strauss [15] defines three basic types of illumination sources in a virtual reality simulation: 1. Point illumination source. The illumination originates from a single point in space (e.g. a light bulb). 2. Directional illumination source. The illumination rays run parallel from a certain direction (e.g. the sun). Could be considered similar to a point illumination source, but positioned at a large or infinite distance. 3. Ambient illumination source. The illumination is spread diffusely and evenly (e.g. illumination from a lamp equipped with lampshades). Strauss [15] states that direction and color of the illumination source are two important properties necessary to simulate proper illumination effect. Moreover, experiment by Slater, Usoh, and Chrysanthou in [16] shows that the existence of shadows could assist the spatial perception of the user. We therefore focused on detecting directional and point illumination sources in order to be able to simulate realistic shadows for a mobile augmented reality system. To achieve this, we require real-time estimation of the illumination condition of the environment surrounding the mobile device. Methods for Estimating Real World Illumination Conditions We explored three different methods that could be used to estimate the real world illumination condition. The first method is using a light probe. A light probe is a spherical object coated with reflective materials that could be used to capture the surrounding illumination of an environment [17, 18]. Since the intensity of illumination sources in an environment normally spans a large dynamic range compared to the environment itself, it is common to use HDR imaging to capture the light probe images. The light probe method has been applied to augmented reality simulations to produce photorealistic virtual objects [6, 7, 8, 9, 10]. In general, the hardware setup of augmented reality simulations developed using this method is using two separate cameras [6, 8], one camera dedicated for viewing the light probe and another camera dedicated to viewing and tracking the AR marker. A one-camera setup is possible [9, 6], but necessitate the light probe to be constantly viewable in the same scene with the AR marker. An additional problem exists since the position of the light probe is dynamically changing with respect to the camera, thus introducing the need to detect and track the position of the light probe. Mounting the probe in a fixed platform is one possible solution for this problem [9]. The second method is using a fish-eye lens to directly detect and estimate the illumination sources surrounding the environment [19]. HDR imaging could be done by varying the shutter speed and aperture size of the imaging device. However, since capturing the environment in several different dynamic range requires the time of several seconds, this means that this method is not suitable for real time applications. The fish-eye lens method has been applied to augmented reality simulations [6, 20]. The third method is using analysis of shadows generated by reference objects with known geometry to estimate the direction of the illumination sources [21, 22] and, in an outdoor environment, the strength of sky irradiance [11]. Panagopoulos et al. use von Mises-Fisher distribution [21] and Markov Random Field [22] to model the direction of the illumination sources from the shadows detected in the input image. Their method and the resulting illumination models was implemented to simulate shadows of virtual objects for augmented reality, but they noted that their algorithm requires three to five minutes of processing time per image, making it unsuitable for real-time augmented reality. Madsen and Nielsen [11] use contrast obtained by comparing the brightness of the shadowed and unshadowed region of an outdo

Salen Mn Complexes are Superoxide Dismutase/Catalase Mimetics that Protect the Mitochondria

Abstract: Salen Mn complexes, including EUK-134, EUK-189 and a cyclized analog EUK-207, are synthetic superoxide dismutase (SOD) and catalase mimetics that are beneficial in many models of oxidative stress. Though not designed to target the mitochondria, salen Mn complexes show "mito-protective" activity, that is, an ability to attenuate mitochondrial injury, in various experimental systems. Treatment with EUK-134 prevents respiratory chain abnormalities induced by ionizing radiation in rat astrocyte cultures. Treatment with salen Mn complexes also prolongs survival, protects mitochon- drial enzymes and prevents oxidative pathologies in Sod2-/- mice, which lack the mitochondrial form of superoxide dis- mutase. Recently, EUK-207 was shown to attenuate ischemia reperfusion injury, including mitochondrial dysfunction, in hearts from ABC-me-/+ mice, which are deficient in a mitochondrial transporter and more vulnerable to oxidative stress. Since mitochondrial dysfunction has been implicated in many forms of injury and degeneration, this "mito-protective" property may explain some of the cytoprotective effects of salen Mn complexes in vivo, and may also enhance their poten- tial therapeutic value.

Toward a Perceptually Based Metric for BRDF Modeling.

Abstract: Measured materials are used in computer graphics to enhance the realism of synthetic images. They are often approximated with analytical models to improve storage efficiency and allow for importance sampling. However, the error metrics used in the optimization procedure do not have a perceptual basis and the obtained results do not always correspond to the best visual match. In this paper we present a first steps towards creating a perceptually-based metric for BRDF modeling. First, a set of measured materials were approximated with different error metrics and analytical BRDF models. Next, a psychophysical study was performed to compare the visual fidelity obtained using different error metrics and models. The results of this study show that the cube root metric leads to a better perceptual approximation than other RMS based metrics, independently of the analytical BRDF model used. More benefit of using the cube root metric compared to the RMS based metrics is obtained for sharp specular lobes, and as the specular lobe broadens the benefit of using the cube root metric decreases. The use of the cube root error metric will improve the visual fidelity of renderings made using BRDF approximations and expand the usage of measured materials in computer graphics.

The Pharmacokinetics and Toxicology of Aluminum in the Brain

Abstract: The chemical forms (species) of aluminum in blood plasma and brain extracellular fluid are considered, as they are the candidates for brain aluminum uptake and efflux. The blood-brain barrier is the primary site of brain aluminum up- take. The mechanism of brain uptake of aluminum transferrin, long thought to be mediated by transferrin-receptor medi- ated endocytosis, requires further investigation. Brain Al citrate uptake has been attributed to the sodium-independent L- glutamate/L-cystine exchanger system, system Xc-. Reports have suggested aluminum can compromise blood-brain barrier in- tegrity, however the studies were conducted with aluminum concentrations greatly exceeding those seen in human blood plasma. Aluminum appeared in cerebrospinal fluid suggesting it can cross the choroid plexus and in brain after intranasal appli- cation suggesting it can be taken up by cranial nerves, but neither of these routes has been definitively demonstrated. Brain aluminum efflux appears to be carrier-mediated, however the mechanism has not been identified. A small increase in brain aluminum seems sufficient to produce neurotoxicity. Once aluminum enters the brain it persists there for a very long time; es- timates of the half-life range from 20% of the lifespan to greater than the lifespan. Al persistence in bone, which maintains the majority of the body burden, may influence brain Al, due to equilibrium among the body's organs. Chelation therapy with des- ferrioxamine has been shown to reduce some manifestations of aluminum toxicity although it may increase redistribution of aluminum to the brain to increase aluminum-induced neurotoxicity. An orally-effective aluminum chelator that is an improve- ment over desferrioxamine has not yet been demonstrated. Although a non-essential metal, there are mechanisms enabling aluminum to get into the brain, accumulating over the lifespan, and creating the potential to contribute to many neurodegenera- tive disorders.

Efficient spectral imaging based on imaging systems with scene adaptation using tunable color pixels

Abstract: A method for adaptive spectral image capture that may be performed via an image capture device is disclosed. A default capture parameter is applied to an imaging assembly and a sample image of a scene is captured by an image capture device. The sample image is analyzed to identify transition zones between multiple different regions. An initial guess as to which spectral regions a filter mode might work best is obtained based on dominant transition region spectrum and a first iterated step in which numerical values for the filter mode are calculated. A second iterated step in which each such filter mode is evaluated for effectiveness against other filter modes. The regions in which a specific filter mode works best becomes associated with the filer mode and these regions become the guess for the next iteration.

Automated Pre-processing Method for Dermoscopic Images and its Application to Pigmented Skin Lesion Segmentation.

Abstract: In this paper, we put forward a new pre–processing scheme for automatic analysis of dermoscopic images. Our contributions are two-fold. First, we present a procedure, an extension of previous approaches, which succeeds in removing confounding factors from dermoscopic images: these include shading induced by imaging non-flat skin surfaces and the effect of light-intensity falloff toward the edges of the dermoscopic image. This procedure is shown to facilitate the detection and removal of artifacts such as hairs as well. Second, we present a novel simple yet effective greyscale conversion approach that is based on physics and biology of human skin. Our proposed greyscale image provides high separability between a pigmented lesion and normal skin surrounding it. Finally, using our pre–processing scheme, we perform segmentation based on simple grey-level thresholding, with results outperforming the state of the art.

Quantum-mechanical studies of reactions performed by cytochrome P450 enzymes

Abstract: We review density functional theory studies of various types of reactions performed by the cytochrome P450 family of enzymes. We describe the various reactions on equal footing with an emphasisis on models to predict sites of metabolism for an arbitrary molecule. The activation barriers range between 0 and 109 kJ/mol, depending more on the atoms surrounding the reactive site than on the type of reaction. Therefore, the intrinsic reactivity can rather well be predicted by simple chemical rules. However, for a full predictive model, the steric effects of the enzyme surrounding the heme group need also to be modeled, which often is harder.

Revisiting Spectral Printing: A Data Driven Approach.

Abstract: Spectral printing is a well–established part of imaging that can boast of a rich body of literature. Nonetheless there has been limited commercial uptake of this approach to visual content reproduction, in spite of its clear benefits. The aim of the present paper is therefore to explore what may lie behind this apparent mismatch by looking at how colorimetric (metameric) and spectral reproduction compare on an 11–ink printing system. To aid the above exploration, the paper proposes a new metric for evaluating spectral reproduction in a visually meaningful way and presents an analysis of the spectral properties of colorimetric and spectral reproductions of a variety of original content including spot colors and fine art. Introduction A key choice when making a print is to decide how it is to relate to original content. This can range from the print becoming the first „original‟ (e.g., fine art created digitally, where its viewing on a display is only an intermediate step of the creative process), via its aim being to please (e.g., holiday snaps) to it being as close to a facsimile as possible (e.g., fine art reproduction, proofing). In the last case, the question arises of how broadly the match needs to hold: only under specific viewing conditions or under any (or a broad range of) lighting and viewing. Here the former is a colorimetric (metameric) reproduction while the latter is a spectral one, which has the benefits of mimicking an original more closely so that looking at it gives the same visual experience as looking at the original would, regardless of where they are viewed and who does the viewing. Conversely, the colorimetric case is set up for specific lighting (typically D50 or D65) and with a specific viewer in mind (usually the 2° CIE Standard Observer) and tends to break down under other conditions (hence its „metameric‟ label). As the case looks very strong for spectral reproduction, it is worth putting two caveats on the table: first, how accurately a spectral match can be achieved and second, how much closer it is to an original than the spectral match obtained when colorimetric matching is set up. The two questions are related in that both spectral and colorimetric reproduction have potentially the same spectral variety at their disposal (being a consequence of the inks, substrate and their interactions) where the difference between an explicit spectral match and the spectral fit of the colorimetrically– selected match may be significantly smaller than the mismatch of either of these to the original reflectance spectrum. In other words, a key question is the spectral „compatibility‟ of the original content and printing system‟s potential. Spectral printing is a topic that can boast of a rich body of literature exploring its various aspects, developing its component building blocks (e.g., spectral capture, printer models (e.g., Taplin 1996), gamut mapping, error metrics for minimization) and applying it in various ways (e.g., fine art reproduction, proofing – including of textiles). Given such a well–established field, it is maybe surprising that it has not found more commercial application and the aim of the present paper is also to look for possible reasons for this fact. Two test cases will therefore be considered: fine art reproduction and spot color proofing, both of which are, in principle, a very good match to the benefits of spectral reproduction. The spectral properties of originals and the way they relate to the spectral variety accessible using two printing setups will then be evaluated. Finally the closest achievable matches will be quantified using an evolution of existing multi–illuminant ∆E metrics that aims to be more representative of an original–reproduction pair being viewed under a broad variety of viewing conditions. Before proceeding with an overview of the state of the art, it may be worth underlining why the above two aspects of spectral characteristics (a physical, „device dependent‟ feature) and perceived difference under a variety of conditions (a psychophysical, „device independent‟ aspect) are considered side by side. The reason for this is that image reproduction is concerned precisely with the interplay between reproduction capabilities and their effects on a viewer – i.e., the device dependent seen in a device independent way. State of the art of spectral match metrics Before turning to the analysis outlined above, two areas of the literature will be reviewed: dimensionality reduction (allowing for an analysis of spectral „compatibility‟) and metrics for evaluating spectral matches. In terms of dimensionality reduction, the basic idea is that the underlying variance in spectral data is often of lower dimensionality than that of the measured reflectance space (i.e., typically having 31D for a 400–700nm range sampled at 10nm steps) and that it can therefore be expressed as a weighted combination of a smaller number of bases. In other words: R=B*w, where R is a 31x1 reflectance vector, B is a 31xn matrix containing n bases and w is an nx1 set of weights for combining the bases linearly. Then there are numerous choices of how to obtain the bases, how many of them to use and what space to use this representation in. Here Ramanath et al. (2004) present a survey that compares Principal Component Analysis (PCA), Independent Component Analysis (ICA) and Neural Networks (NN) and also covers methods for obtaining sets of all–positive bases (e.g., Non– negative Matrix Factorization) and their results show similar performance for these approaches when using three bases, with PCA performing best for their data. Tzeng (1999) introduces an important consideration to dimensionality reduction – that the choice of space in which bases are computed and where their combinations are made plays an important role. He then goes on to show that the dimensionality reduction of spectra measured from an IT8.7/2 chart (a three–dye photographic print) suggests that six bases are needed, while it is known that there are only three independent components at play. Tzeng shows how a conversion into the Kubelka Munk K/S absorption space before PCA results in the same level of variance >99.9% being spanned by only three bases. Finally, an important question is, how much variance coverage is enough? One approach is to state that 99.9% ought to be plenty and then select the number of bases that give the necessary coverage. Another is to look for meeting a 1 ∆E threshold under a reference illuminant and choose the number of dimensions to achieve it. Finally, a very well reasoned approach is to use psychophysics to find how many bases it takes to match hyperspectrally–captured scenes. Here Nascimento et al. (2005) report that 8 bases were needed for a 55% discrimination threshold (corresponding to a mean ∆E*ab of 0.7–0.8, which corresponds to a ∆E00 of around 0.4 (Sun and Morovic, 2002)) even though 5 bases would have been sufficient to get to the unit ∆E threshold. Turning to the evaluation of spectral match metrics, Imai et al. (2000) and Viggiano (2004) presented two excellent surveys, comparing metrics that range from spectral–only methods like RMS (the root mean square difference between two reflectance spectra) and GFC (Hernandez–Andres et al.‟s (2001) goodness of fit coefficient), via various weighted version of RMS, e.g., using the diagonal of Fairman‟s (1987) matrix R derived from tristimulus weights for a given illuminant and observer, to metamerism indices (which report the color difference under a test illuminant – e.g., A – for a match under a reference illuminant – e.g., D65) and even a combined spectral and colorimetric metric: CSCM (Lopez–Alvarez et al., 2005). The conclusions of both these surveys are that none of these metrics can be universally recommended over the others and that their choice is a matter of what application it is being used for. The basic challenge here is that while RMS expresses the physical difference between a pair of spectra, it is not visually meaningful. The fact that the starting point is often a mismatch already under a reference illuminant rather than a strict match is a complication, which means that metamerism indices are often applied not directly to an original–reproduction pair, but to one that has been „corrected‟ (e.g., using Fairman‟s (1997) method) to force a match so that the metameric difference under a test illuminant can be expressed. A different approach is then taken by Alsam and Hardeberg (2004) and Bastani et al. (2007) who consider ∆E statistics under multiple illuminants: 6 in the former and 11 in the latter case. Given the above approaches to dimensionality reduction and spectral match metrics, the following sections will first introduce a new alternative to the reflectance or absorption based PCA approaches, proceed to make a more explicit comparison between original and reproducible spectra, propose a new spectral match metric that extends the multi–illuminant methods mentioned previously and finally apply them to the example original and print conditions.

A Large-Scale Multi-Lingual Color Thesaurus

Abstract: We present a color thesaurus with over 9000 color names in ten different languages. Instead of using conventional psychophysical experiments, we use a statistical framework that is based on search results from Google Image Search. For each color name we compute a significance distribution in CIELAB space whose maximum indicates the location of the color name in CIELAB. A first analysis discusses the quality of the estimations in the context of human language. Further, we conduct an advanced analysis supporting our choice to use a statistical method. Finally, we demonstrate that a color name mainly depends on the chromatic values and varies more along the lightness axis.

Removing Outliers in Illumination Estimation.

Abstract: A method of outlier detection is proposed as a way of improving illumination-estimation performance in general, and for scenes with multiple sources of illumination in particular. Based on random sample consensus (RANSAC), the proposed method (i) makes estimates of the illumination chromaticity from multiple, randomly sampled sub-images of the input image; (ii) fits a model to the estimates; (iii) makes further estimates, which are classified as useful or not on the basis of the initial model; (iv) and produces a final estimate based on the ones classified as being useful. Tests on the Gehler colorchecker set of 568 images demonstrate that the proposed method works well, improves upon the performance of the base algorithm it uses for obtaining the sub-image estimates, and can roughly identify the image areas corresponding to different scene illuminants. Introduction There are two types of outliers that create problems for illumination-estimation algorithms. In the illumination-estimation context, an outlier is an observation that does not fit the illumination model well. One type of outlier arises from noise in the image data created, for example, by a speck of dust on the imaging sensor, or by clipping of high digital counts. A second, more interesting, type of outlier arises from scenes that do not fit the expected model of the scene illumination—for example, a predominantly indoor scene with some light also coming through a window. In this paper, we apply the RANSAC (random sample consensus) [5] technique to randomly sampled subwindows of the input image as a means of handling outliers of both these types. The proposed algorithm was evaluated on Shi’s reprocessed version of Gehler’s original ‘colorchecker’ set of 568 images [7, 14] and found to reduce the high-percentile angular errors by roughly 30%. Illumination estimation is the crucial step in standard automatic white balancing or ‘color constancy’. The goal in illumination estimation is to determine the chromaticity of the overall scene illumination. The accuracy of the estimate is often measured in terms of the angular difference in degrees between the estimated chromaticity and the actual chromaticity when treated as vectors in 3-space. Many illumination-estimation algorithms have been proposed and are surveyed by Barnard et al. [1], Hordley et al. [10] and Gijsenij et al. [8]. Noting that outliers of the two types mentioned above will mislead most illumination-estimation algorithms, we propose a novel technique to improve upon any given algorithm or set of algorithms by explicitly taking outliers into account. The proposed method divides the input image into smaller sub-images, runs the algorithm(s) on each of the parts independently and then combines the resulting estimates. Outliers are identified and eliminated as part of the process of combining the estimates. The rational behind the proposed method is that many illumination-estimation algorithms rely on information that can be significantly influenced by a small part of an image. For example, MaxRGB [7] and retinex [11] both can be influenced by a single pixel having a spuriously high R, G or B value. Gamut mapping algorithms such as Forsyth’s [6] can be influenced by a single erroneous pixel that happens to stretch the convex hull of the gamut significantly in the wrong direction. For a single-illuminant scene, the estimates from multiple sub-images should be consistent. Those that are inconsistent can be identified as outliers and eliminated. Sub-images, of course, do not carry as much information as the whole image. Therefore, the performance of an algorithm on each of the parts is likely to be worse than its performance on the full image; that is, assuming the full image contains no outliers. However, it is often the case that even a fairly small sub-image contains enough information for the underlying illuminationestimation algorithm to work reasonably well. As an example, consider the images in Figure 1. It is likely that quite a few of the vertical or horizontal slices will cover a sufficient proportion of the complete set of image colors to make gamut mapping algorithms work. Similarly, there is a good chance that they contain the necessary high R, G or B digital counts that MaxRGB requires, or are sufficiently textured for Edge-based Color Constancy [15] to succeed. Illumination-estimation algorithms generally assume there is a single illuminant lighting the imaged scene, or at least that even if there is more than one illuminant then there is only one dominant illuminant. It is expected that white balancing the image relative to the dominant illuminant will suffice. In terms of the second type of outlier—those related to multiple illuminants—the information from a sub-image is likely to be more reliable, not less, than that from the image as a whole, because by being smaller the sub-image is more likely to involve only a single illuminant. Proposed Algorithm The proposed algorithm combines illumination estimates obtained from sub-images using RANSAC as a method of eliminating outliers. The core idea of RANSAC is to determine which observations are inliers and which are outliers and to base the final result only on the inliers. The data is assumed to fit some underlying model defined by some parameters (e.g., a model could be a straight line with the slope and intercept being the parameters). The process works by: (1) randomly selecting some observations; (2) determining the model parameters that best fit those observations; (3) testing all the remaining observations and classifying them as inliers or outliers based on how well they conform to the model; (4) checking to see that a sufficient number of inliers remain, and if not, discarding the model; (5) recomputing the model parameters based on the complete set of inliers. The algorithm repeats the steps (1) to (5) to generate many 20th Color and Imaging Conference Final Program and Proceedings 105 possible models. The model that fits the observations the best is returned as the result. RANSAC is used to deal with the two types of outliers mentioned above. The underlying model of the image data that RANSAC fits is different for each of them. To handle multiple sources of illumination, the model is that the scene contains 3 distinct illuminants. Any additional illumination chromaticities that are observed (i.e., calculated from a sub-image by an illumination-estimation algorithm) will be classed as outliers. Applying the RANSAC steps in this case means obtaining illumination estimates from 3 sub-images of random size and location (the ‘observations’), sorting the remaining estimates as either inlier or outlier, checking that a sufficient number of inliers has been found, and re-computing the estimates for each of the 3 clusters obtained this way based on the final set of inliers. For this last step, the inlier estimates within each cluster are simply averaged. In terms of the outliers of the pixel-noise variety, the scene model is that there is only a single illuminant, and hence all the sub-image estimates should conform to this. Those that do not will be considered to be outliers. Note that this definition of outlier will also deal with secondary illuminants to a certain extent. If a secondary illuminant only lights a small portion of the scene then the estimates from the corresponding sub-windows will be excluded as outliers and therefore not influence the estimate of the dominant illuminant. Related Work Combining estimates from multiple illumination-estimation algorithms applied to whole images has been explored before, but generally not to sub-regions of images. One simple method of combining estimates is to use the arithmetic mean [3, 2]. Other strategies are to use a weighted average with weights determined by a training stage using least squares, or to use a trained neural network or support vector regression [12] to combine the estimates [3]. Shades of Grey [4] and Edge-based Color Constancy [15] combine MaxRGB-type and Greyworld-type clues. In both cases, the results are proportional to the Minkowski or p-norm of the form

Image Difference Measure Optimized Gamut Mapping

Abstract: Even though there is still room for improvement, recent perceptual image-difference measures show a prediction performance that makes them interesting to be used as objective functions for optimizing image processing algorithms. In this paper, we use a color enhanced modification of the Structural Similarity (SSIM) index for optimizing gamut mapping. An iterative algorithm is proposed that minimizes this measure for a given reference image subject to in-gamut images. Since distortions within remote image regions contribute independently to the measure a descent direction can be specified locally. The step-length is chosen to be a fraction of the just-noticeable-distance ensuring a decrease of the measure. Results show that the proposed approach preserves contrast and structural information of reference images. Some artifacts suggest modifications of the employed image-difference measure.

CIC@20: Multispectral Imaging.

Abstract: A variety of multi-spectral imaging methods are discussed for acquiring spectral information from a scene. We first review conventional multispectral imaging approach. The conventional imaging systems are mostly constructed by multi-band imaging devices with different filtration mechanism at the sensor side under passive illumination. We show some imaging devices, estimation algorithms, and applications. Recently, active spectral imaging attracts much attention as promising technology. The active spectral imaging method has the possibility of recovering spectral reflectance information and estimating tristimulus values of object surface in high speed. We introduce a spectral imaging technology by synchronizing a programmable light source and a high-speed monochrome camera. Two effective applications to spectral reflectance recovery and tristimulus imager are described. Introduction Multispectral imaging technology is a useful technology that is now widespread in all fields related with visual information. So far a variety of multispectral imaging systems and methods have been proposed for acquiring spectral information from a scene. Figure 1 shows the number of papers related with multispectral imaging, presented in the CICs of the past 19 years as a function of CIC number. The first paper in CIC was entitled "Analysis multispectral image capture" by Peter D. Burns and Roy S. Berns, 1996 [1]. In the same year, "Multichannel vision system for estimating surface and illuminant functions" by the author was published in JOSA, 1996 [2]. Therefore we can consider that the year of 1996 was the starting point of multispectral imaging. The largest number of papers at 14 were presented in CIC19 of the last year, when MCS (Multispectral Color Science) joined to CIC. The session titles related with multispectral imaging in the past CICs are listed as follows: CIC6 : Input CIC7 : Image Capture CIC9 : Spectral Image Analysis CIC12 : Multi-spectral / Multi-primary Systems CIC13 : Spectral Imaging CIC14 : Multi-spectral imaging CIC15 : Spectral Color CIC17 : Spectral Color CIC19 : Multispectral Color Science (joint with 13 MCS). The research contents cover a broad range of areas, including spectral image capture, spectral reflectance estimation, illuminant estimation, spectral image compression, color reproduction system for spectral image, and computer graphics based on spectral images. The application fields of multispectral imaging look medicine, human skin, art (mainly art paintings), and wide gamut technology in the past. However, it is quite certain that application in the field is expanding. Figure 1. Number of papers related with multispectral imaging, presented in

Deriving Appearance Scales.

Abstract: The concept of color space has come to be an unquestioned threedimensional representation of color stimuli, or color appearance, intended to simplify the relationships among physically measurable attributes of light, mathematical formulae, and human sensations and perceptions. The notion of three-dimensional mathematical spaces as adjuncts for color is often helpful, but perhaps also misleading at times. Color appearance models requiring five or six dimensions to represent color appearance illustrate some of the limitations of historic spaces. This paper poses the question of whether color appearance would be better represented by independent appearance scales with no requirement that they be related as a higher-dimensional space. In other words, is color better represented by six one-dimensional color scales than one or two three-dimensional color spaces. A framework for implementing such appearance scales is described and one implementation is presented along with discussion of the ramifications for color difference metrics. Introduction Color scientists and engineers have become accustomed to the fundamental concept of color space to the point that the concept itself goes unquestioned. Much like most accept the fact that the earth is nearly spherical, those in the color-related fields proceed merrily along without a doubt that color space is three dimensional. Further, some continue to seek the holy grail of a three-dimensional color space in which perceived color differences can be expressed as uniform Euclidean distances despite an apparent lack of psychophysical evidence that such a space might exist. Perhaps it is time to, once again, step back and ask the question of whether the concept of a Euclidean distance metric in three dimensions really makes sense for describing color, even approximately. Perhaps some insight into appropriate descriptions of color appearance can be gained from a cursory examination of the other human senses.[1,2] Our perception of taste has at least five distinct dimensions, sweetness, bitterness, sourness, saltiness, and umami, and seldom does anyone speak of changes in taste perceptions as a Euclidean difference space. Similarly our sense of smell is served by something on the order of 1000 different receptor types. Some have tried to reduce the dimensionality to approximately six including flowery, foul, fruity, spicy, burnt, and resinous. Our sense of hearing is actually spectral (plus intensity) in terms familiar to color scientists as humans are able to detect frequencies within sounds (no aural metamerism) and the relative intensities of each frequency. Finally our sense of touch might well be too complex to even attempt to summarize in a sentence or two. None of the perceptions arising from any of these senses are commonly expressed in terms of multi-dimensional spaces with Euclidean (or similar) difference metrics. Given these similarities in our other senses, why should we think color is different? Is it the relatively low dimensionality? Is it the seemingly simple perceptual relationships such as color opponency? Is it the nature of additive color mixing? Additive color mixture under photopic conditions provides ample evidence for trichromacy, the three-dimensional nature of color matching/mixture. Adding Grassmann’s laws allows expression of color matches in various sets of primaries via simple 3x3 linear transformations analogous to a change of basis in a three-dimensional, linear space (where Euclidean distances mean something mathematically). Perhaps it is this property of color matching, which is not a direct representation of perception or appearance, that leads to an almost irresistible next step to start expressing color matches in three-dimensional Euclidean spaces. And then, apparently without clear justification, the concept is carried forward in attempts to express appearances and differences in similar three-dimensional Euclidean spaces such as the CIELAB color space. Perhaps those attempts were always as doomed as any explorers who might have set out to “circumnavigate” a flat earth. Color science is not devoid of examples typically described as color spaces that are actually descriptions of color perception one dimension at a time.[3] For example, the Munsell system, despite its common embodiments, was derived as a system of three independent perceptual dimensions, hue, value, and chroma. Similarly, Guth’s ATD model of visual perception was typically described in terms of independent dimensions, although the temptation to plot some of them together for some examples proved irresistible. Likewise, color appearance models such as CIECAM02 were developed with independent predictors of the six perceptual dimensions of brightness, lightness, colorfulness, saturation, chroma, and hue. This was somewhat compromised by requests for rectangular color space dimensions which appeared as CIECAM97s evolved to CIECAM02. However it should be noted that cylindrical representations of the appearance spaces were common even before the requests for rectangular coordinates. Lastly, the NCS system provides a useful example of hue being treated separately from whiteness-blackness and chromaticness. And while NCS whiteness-blackness and chromaticness are plotted in two-dimensional trilinear form, the dimensions are largely independent since the anchor of maximal chromaticness appropriately varies from hue to hue. All of this insight leads to the hypothesis that perhaps color space is actually a one-dimensional space, rather than a threedimensional space, and that Euclidean distance metrics might indeed be successful in such a space. Of course, color appearance cannot be properly described in a single one-dimensional space. Instead six of them are required. There are three fundamental appearance attributes for related colors, lightness, saturation, and hue. Combined with information about absolute luminance, colorfulness and brightness can be derived from these and are important and useful appearance attributes. Lastly, chroma can be derived from lightness and saturation if desired as an alternative relative colorfulness metric. Thus, color is rightfully and fully described with six one-dimensional appearance spaces (or scales), four of which are fundamental for related colors and two of which are derived from the fundamental scales. This paper provides some detail of the conceptual framework of a color model made up of one-dimensional spaces and an implementation of that framework for future application and investigation. Note: One-dimensional “spaces” are more commonly referred to as “scales” in color science, thus the term “scale” is used preferentially for the remainder of the paper. Conceptual Framework A set of color appearance scales (or dimensions, or spaces) following these principles has been derived and an implementation is presented in the next section. This section provides the general framework that could be easily adapted to different specific implementations of the concept. The first step is to apply a chromatic adaptation model to compute corresponding colors for reference viewing conditions (e.g. D65, 315 cd/m2). Then the IPT model, derived specifically for accurate hue representations, is used to compute a hue angle (h) and then a hue composition (H) can be computed based on NCS unique hues. For the defined hue, saturation (S) is computed using the classical formula for excitation purity applied along lines of constant h in the u’v’ chromaticity diagram. For that chromaticity, the luminance for zero gray content, G0, is defined as the reference for lightness (L) computations that follow a power function with offset model found to perform well in recent research for high-dynamic-range lightness-brightness scaling. The remaining dimensions are then derived from L and S along with luminance information. Brightness (B) is lightness (L) scaled by a factor derived from the classic work of Stevens and Stevens that illustrated terminal brightness as a function of adapting luminance. The derived scales are colorfulness (C), which is simply saturation (S) scaled by brightness (B), and chroma (Ch) which is saturation (S) times lightness (L). This type of formulation allows accurate description of color appearance for lights and objects across a variety of adaptation conditions and for lowor high-dynamic-range scenes. To the degree that each perceptual scale is accurate, differences on each of the dimensions should be easily calculated and, as long as the temptation to combine those differences into a single Euclidean distance metric is resisted, quite effective results can be obtained. The next section steps through a proposed implementation in detail. Implementation Fairchild,[4] at ISCC/IS&T/SID meeting on color spaces, outlined a methodology for computing the set of three fundamental appearance attributes of hue, saturation, and lightness for related colors from which the attributes of brightness and colorfulness can be derived as a function of the absolute luminance along with chroma. As the hue-linearized space IPT, based in opponent color theory, is considered exceptionally uniform in hue, the hue scale (h) is computed as a simple hue angle using the IPT model.[5] The required inputs for the IPT hue angle computation are the CIE tristimulus values in XYZ for the corresponding colors in CIE Illuminant D65. A chromatic adaptation transform is required to obtain corresponding colors for Illuminant D65 if the stimuli of interest are viewed under a different state of adaptation. The CAT02 transformation imbedded in the CIECAM02 color appearance model is recommended with a simple von Kries transformation on cone fundamentals a second choice. If luminance information is available and impacted by the selected chromatic adaptation transformation, then transformation to a white-point luminance of cd/m2 is recommended. Hue composition (H) can be obtained by recognizing th