Linköping University Post Print

Color Semantics for Image Indexing

Martin Solli and Reiner Lenz

N.B.: When citing this work, cite the original article.

Original Publication:

Martin Solli and Reiner Lenz, Color Semantics for Image Indexing, 2010, CGIV 2010: 5th

European Conference on Colour in Graphics, Imaging, and Vision, 353-358.

Copyright: Society for Imaging Science and Technology (IS&T)

http://www.imaging.org/

Postprint available at: Linköping University Electronic Press

http://urn.kb.se/resolve?urn=urn:nbn:se:liu:diva-57481

Color Semantics for Image IndexingMartin Solli and Reiner LenzDept. of Science and Technology (ITN), Linkoping University, Sweden

AbstractWe propose a color-based image descriptor that can be used

for image indexing based on high-level semantic concepts. Thedescriptor is based on Kobayashi’s Color Image Scale, whichis a system that includes 130 basic colors combined in 1170three-color combinations. Each combination is labeled with oneof 180 high-level semantic concepts, like ”elegant”, ”roman-tic”, ”provocative”, etc. Moreover, words are located in a two-dimensional semantic space, and arranged into groups based onperceived similarity. From a modified approach for statisticalanalysis of images, involving transformations of ordinary RGB-histograms, a semantic image descriptor is derived, containingsemantic information about both color combinations and singlecolors in the image. We show how the descriptor can be trans-lated into different levels of semantic information, and used inindexing of multi-colored images. Intended applications are, forinstance, image labeling and retrieval.

IntroductionMost available systems for image labeling or Content Based

Image Retrieval use objects as their prime descriptor of imagecontent. Typical tasks are finding (or labeling) images contain-ing a certain vehicle, an animal, a mountain, a face, etc. In re-cent years, however, the interest in image labeling and retrievalapproaches based on high-level semantic concepts, such as emo-tions and aesthetics, has increased. In a recent survey by Dattaet al. [7], the subject is listed as one of the upcoming topics inContent Based Image Retrieval. In this work we use ShigenobuKobayashi’s system [13] for color semantics of single colors andthree-color combinations, called the Color Image Scale. WithinGraphic Arts, the Color Image Scale is a well known and es-tablished tool used in, for instance, the selection of colors andcolor combinations. It is, of course, questionable if the samesystem can be applied in indexing of multi-colored images. Wewill, however, illustrate that the proposed method results in pre-dictions that can be useful in large scale image indexing. Asan illustration, the method is applied on a small database of 5000images. Some of the results are presented in this paper. But moreimportant, the findings are implemented in a publicly availabledemo search engine1, where readers can interact with the searchengine while reading this paper. Before continuing, one shouldemphasize that we cannot expect the proposed method to delivera useful semantic indexing for every possible image. The con-cept of color semantics is influenced by many factors, such ascognitive, perceptual, cultural, etc. Moreover, high-level imagesemantics can be obtained from various kinds of image content,not only color semantics. Possible relationships between colorsemantics and other types of high-level semantics are beyond thescope of this paper. The method presented is a first step towardsa broader use of high-level color semantics in image labeling andContent Based Image Retrieval.

1http://diameter.itn.liu.se/colse/

Related workEven if the interest for color semantics has increased, rela-

tively few papers are addressing the problem of including high-level semantics in image labeling and retrieval. Early experi-ments in this research field can be found in papers by Beretti etal. [1], Corridoni et al. [5], and Corridoni et al. [4]. Clusteringin the CIELUV color space, together with a modified k-meansalgorithm, is used for segmenting images into homogenous colorregions. Then fuzzy sets are used to convert intra-region prop-erties (warmth, hue, luminance, etc.) and inter-region properties(hue, saturation, etc.) to a color description language based oncolor semantics. A similar approach, using clustering and seg-mentation, is described by Wang and Yu [16]. Images are seg-mented by clustering in the CIELAB color space. Then imagesare converted to the CIELCH color space (the cylindrical versionof CIELUV), and segmented regions are converted to semanticterms through a fuzzy clustering algorithm. Both regional andglobal semantic descriptors are extracted. The user is able toquery the image database with emotional semantic words, like”sad” and ”warm”, and also with more complex sentences. In an-other paper, Wang et al. [17] present an annotating and retrievalmethod using a three-dimensional emotional space. From his-togram features, emotional factors are predicted using a SupportVector Machine. The method was developed and evaluated forpaintings.

Cho and Lee [2] propose an image retrieval system basedon human preference and emotion, using an interactive geneticalgorithm (IGA). Image features are extracted from average col-ors and wavelet coefficients. In a paper by Hong and Choi [11], asearch scheme called Fuzzy Membership Value (FMV) Indexingis presented. Using keywords such as ”cool”, ”soft”, ”roman-tic” etc., it allows the user to retrieve images based on high-levelsemantic concepts. Emotion concepts are derived from color val-ues in the HSI color space. Yoo [18] propose an image retrievalmethod using descriptors that are called query color code andquery gray code. The descriptors are based on human evaluationof color patterns on 13 emotion scales, most of them related tocolor. The image database is queried with one of the emotions,and a feedback method is utilized for dynamically updating thesearch result. In Lee et al. [14] they show how rough set theorycan be used to build an emotion-based color image retrieval sys-tem. Emotion data is extracted by letting people observe differ-ent random color patterns in category scaling experiments, usingthree different emotion scales. The primarily field of applicationis different color patterns, like wall papers etc.

Related to emotions is the concept of harmony (see an ex-ample by Cohen-Or et al. [3], where harmony is used in computa-tional imaging to beautify images), and the concept of aesthetics.In [6][8] Datta et al. study aesthetics in images from an onlinephoto sharing website, where images were peer-rated in the cat-egories aesthetics and originality. Support Vector Machines andclassification trees are used for comparing observer ratings withimage features corresponding to visual or aesthetical attributes(like Exposure, Colorfulness, etc.). In [9] Datta et al. introduce

CGIV 2010 Final Program and Proceedings 353

the phrase ”aesthetic gap”, and report on their effort to build areal-world dataset for testing and comparison of algorithms.

Similar labeling and search strategies have not yet been usedin commercial search engines, with one exception, the Japaneseemotional visual search engine EVE2. This search engine is alsousing the same emotion scales (soft - hard, and warm - cool) asused in this study.

Kobayashi’s Color Image ScaleThe Color Image Scale [13] is a book, or collection, devel-

oped by Shigenobu Kobayashi and his team at the Nippon Color& Design Research Institute (Japan). Early thoughts about thesystem are presented in [12]. In their psychophysical investiga-tions they have matched 130 basic colors and 1170 three-colorcombinations to 180 keywords, or image words, belonging tohigh-level semantic concepts related to the ways in which peo-ple perceive colors. The 130 basic colors are defined in the Hueand Tone System, basically a 2-dimensional color space wherethe axes corresponds to hue and tone. Examples of image wordsare ”elegant”, ”romantic” and ”provocative”. For each of the 130basic colors, also known as theme colors, nine color combina-tions with other basic colors have been created. Each combi-nation is labeled with one of 180 image words, together with0-5 stars indicating the frequency with which the theme color isused in the current color combination (a kind of popularity mea-surement). The relationship between a color combination and asingle image word can be described as the lowest level of abstrac-tion. On the highest level of abstraction, all color combinations,or corresponding image words, are located in a two-dimensionalKeyWord Image Scale, where the axes correspond to the scaleshard-soft and cool-warm. Fig. 1 illustrates the concept show-ing a few examples of color combinations, together with theirimage words, plotted in the KeyWord Image Scale. In this two-dimensional space, Kobayashi also defines different regions orcategories, called Patterns. There are 15 Patterns in total, eachsurrounding a group of image words. In addition, each themecolor is included in one or several Lifestyles, closely connectedto peoples taste in colors. There are in total eight Lifestyles, eachof them described with five or six image words.

Hue and Tone RepresentationThe standard color space for ordinary digital images is the

RGB color space. We must therefore first translate RGB valuesto the Hue and Tone System used by Kobayashi. RGB histogramsare very common in Content Based Image Retrieval, and there-fore we start with ordinary RGB histograms of images, denotedby hR, then transform them to Hue and Tone signatures denotedby hK . A typical RGB histogram consists of 512 entries witheight quantization levels (bins) per color channel. These 512 binswill be converted to a 130 components Hue and Tone signature,corresponding to the 130 basic colors in the Hue and Tone Sys-tem. We use a bin transformation described by a matrix T , ofsize 512× 130 (rows× columns), constructed as follows. Thebasic colors, and the mean RGB vector for each bin in the RGBhistogram, are converted to CIELAB coordinates (for a detaileddescription of CIE color spaces see [10]). For component num-ber n in the the Hue and Tone signature, that will be generatedfrom column n in T , we search for the closest neighboring binsfrom the RGB histogram. The distance metric used is the Eu-clidean distance in the CIELAB color space, known as ∆E. Binnumbers are retrieved for all bins less or equal to 10 ∆E away,

2http://amanaimages.com/eve/, with Japanese documentation anduser interface

lively

sweetpretty

calm

heavy and deep

quiet and sophisticatedwild

rustic

sporty

citrus

freshnatural

soft

hard

cool

war

mFigure 1. A few examples of three-color combinations, together with their

image words, positioned in the KeyWord Image Scale. (Color image in the

online version)

and their distances relative to the maximum distance 10 ∆E areused as weight factor in row r in T . For bin number r in theRGB histogram, with distance ∆Er to the closest Hue and Tonecomponent, the weight is calculated as

trn =

{1− ∆Er

10 +0.1 if ∆Er ≤ 100 else

for n = 1...130 (all columns in T ). We add the constant 0.1 toavoid a weight factor close to zero (which otherwise might be aproblem if all found bins have a ∆E distance close to 10, result-ing in an almost empty Hue and Tone component). A problemthat arises is that for some RGB bins, the closest component inthe Hue and Tone representation is more than 10 ∆E away. Wecannot ignore those bins since there might be images contain-ing those colors only. Instead, if no bin can be found within 10∆E, the closest bin is detected and the weight factor is set to 0.1.Since the intended usage is the use in general search engines, typ-ically available on the Internet, where we have no control overthe user environment, we make two assumptions: When trans-forming RGB values to CIELAB values, we assume images aresaved in the sRGB color space, and we use the standard illumi-nation D50. After multiplying the histogram vector hR with T ,we obtain the vector hK

hK = hR ·T (1)

a 130 components Hue and Tone signature describing the distri-bution of color values in the image. In this feasibility study wedo not take into account spatial relationships between colors.

A Semantic Image DescriptorThe next steps in the process are the conversions from a Hue

and Tone signature to image words, Patterns, and a position onthe scales hard-soft and cool-warm.

Image Words from Color CombinationsIn Kobayashi’s collection, the 130 components in hK are

combined to form 1170 three-color (or three-components) com-

354 ©2010 Society for Imaging Science and Technology

binations. For each color combination n, consisting of the threecomponents {b1

n,b2n,b

3n}, we define a signature as

σn = min(hK(b1n),hK(b2

n),hK(b3n)) (2)

The signatures for all color combinations are collected in the vec-tor hT = (σ1...σ1170). Thus hT is a 1170 components signaturedescribing the distribution of three-color combinations found inthe image.

Since each color combination is labeled with one of 180image words, we can create a matrix that will transform colorcombination signatures to image words distributions. For eachimage word n we create an index vector wn of length 1170, withwn(l) = 1 if word n is associated to color combination l. The vec-tors wn are collected in the 1170×180 matrix W , that can be usedalone for deriving image words signatures. We can also includethe ”star”-rating by using a 1170×1170 diagonal weight matrixM, where the value in the diagonal entry mll corresponds to the”star”-rating (+1, to avoid weighting with zero) for color combi-nation l. After multiplying the color combination signature, hT ,with W , and the weight matrix M

hC = hT ·M ·W (3)

we obtain the vector hC, a 180 components signature describingthe distribution of image words belonging to three-color combi-nations in the image.

Image Words from Single ColorsFor some images the method using three-color combina-

tions will produce empty image words histograms. The mostobvious example is an image containing only one or two col-ors. To make the proposed descriptor more robust we add the useof single colors to predict image words. We define the matrix Lof size 130×8 by the entries

lmk =

{1 if color m is related to lifestyle k0 else

where m and k represent rows and columns in L. In the same waywe define the matrix Q, size 8×180, by

qkn =

{1 if image word n is included in lifestyle k0 else

Combining the matrices with the Hue and Tone signature, hK ,gives the 180 components signature

hS = hK ·L ·Q (4)

describing the distribution of image words belonging to singlecolors in the image.

From Words to Patterns and ScalesThe signatures hC and hS, derived from three-color combi-

nations and single colors, are combined to a single image wordssignature. The signature is primarily based on hC, but if the num-ber of color combinations found approaches zero, the weight onhS will increase. The weighting is defined as

hW =hC

mC+ e−∑hC

hS

mS(5)

where mC and mS are the mean of the norm’s for all signaturesin our test database (this scaling is important since the range of

soft

hard

coolwarm

Figure 2. 100 images plotted according to their derived mean scores in

the KeyWord Image Scale. (Color images in the online version)

entries in hC and hS are in different intervals). The signature isnormalized, hW = hW /∑hW , and the result is an image wordshistogram containing the probability distribution of different im-age words.

Image words have been positioned in the two-dimensionalKeyWord Image Scale. Words close to the origin are describedby Kobayashi as ”neutral in value”. Following a dimension out-wards, for instance towards warm, the scale passes the coordi-nates ”fairly warm” and ”very warm”, and end in ”extremelywarm”. We compute the location in the KeyWord Image Scaleas the mean value given by the word histogram, hW . The posi-tion for each image word n is obtained and saved in column n,denoted en, in the matrix E. Multiplying hW with E, we obtain

s = hW ·ET (6)

s is thus an expectation vector containing the mean score for eachof the scale factors: hard-soft and cool-warm. In Fig. 2, 100images are plotted according to their derived mean scores in theKeyWord Image Scale.

To make the keyword scale easier to understand, Kobayashihas clustered image words that convey a similar image into broadcategories, called Patterns. Each image word is included in atleast one of totally 15 Patterns. To derive what Patterns an imagebelongs to we create the matrix U , of size 180×15, defined as

unk =

{1 if word n is included in pattern k0 else

The vector p, given by

p = hW ·U (7)

is thus an expectation vector for different Patterns. The positionwith the largest value corresponds to the Pattern with highestprobability. As we will illustrate in the next section, the descrip-tors obtained can be used in image indexing, and applied in bothimage labeling and retrieval.

CGIV 2010 Final Program and Proceedings 355

Image Words Pattern Image Scale

boldfestiveforcefulhotdynamic

DYNAMIC Fairly warmNeutral in soft/hard

agreeable to the touchlightgentleamiabledry

ROMANTIC Neutral in cool/warmVery soft

elegantrefinedsubtlesimplesedate

ELEGANT Fairly coolFairly soft

distinguishedpreciseauthoritativeurbanechic

MODERN Fairly coolNeutral in soft/hard

lightsoftinnocentwholesomeemotional

CLEAR Fairly coolVery soft

polisheddelicatesubtlesolemnurbane

ELEGANT Fairly coolNeutral in soft/hard

crystallinepurepolishedcleanrefined

CLEAR Fairly coolFairly soft

agreeable to the touchmildintimategentleamiable

NATURAL Neutral in cool/warmVery soft

quietplaciddignifiedsubstantialsharp

MODERN Fairly coolFairly hard

placidquiettraditionalsubstantialdignified

CLASSIC Neutral in cool/warmFairly hard

delicatesubtleculturedpolishedemotional

ELEGANT Fairly coolFairly soft

Figure 3. Images are labeled with five image words, one Pattern, and

the classification on the KeyWord Image Scale. (Color images in the online

version)

Query Five best matches

forcefulintensesharp

dreamylightmild

calmgracefulpeaceful

citrusfreshnatural

Figure 4. Image retrieval based on image words histograms, either defined

by a selection of image words, or derived from a query image. (Color images

in the online version)

356 ©2010 Society for Imaging Science and Technology

IllustrationsSome results are illustrated with a test database containing

5000 images, both photos and graphics. The database is a ran-dom subset of a much larger database used in previous and ongo-ing research. Examples of labeling are shown in Fig. 3. Imagesare labeled in three levels of semantic abstraction: 1) The fiveimage words that obtained the highest probability 2) The mostprobable Pattern 3) A semantic classification corresponding tothe position on the KeyWord Image Scale.

Since each image is positioned in the KeyWord Image Scale,we can also define a query by selecting a point, sq, in the space,and then search for images that have coordinates close to thispoint. Fig. 2 illustrates what images one will receive for differ-ent query points. One can also retrieve images based on a se-lection of either Patterns or image words. For the later, a queryhistogram, hWq, is created based on the selected image words,and the distance between hWq and image words histograms forall images in the database is derived as a measure of seman-tic similarity. Fig. 4 illustrates a few retrieval results based onthe selection of image words. The same figure holds exampleswhere a query image is used. Here derived image words his-tograms, hW , are used for finding images with a similar semanticcontent. An alternative is to perform the search in one of theother two abstraction levels. Notice the difference to ordinarycolor-based retrieval, with for instance RGB-histograms. Whenassigning colors to semantic concepts, it is rather common thatdifferent colors are associated with the same semantic term. Sev-eral illustrative examples can be seen in Fig. 4, for instance in row3, where completely different colors are associated with the theimage words ”forceful”, ”intense” and ”sharp”. In all examplesabove, the L2-norm is used as distance metric between coordi-nates or histograms. The findings of this study are implementedin our publicly available demo search engine.3 We encouragereaders to interact with the system while reading this paper.

EvaluationAn objective measure of performance (for instance by mea-

suring Precision and Recall, plotting ROC curves, etc.) is diffi-cult to design since color-based high-level semantic concepts arehard to define. One solution would be to evaluate the method inpsychophysical experiments. However, creating psychophysicalexperiments including all possible image words, Patterns, andpositions on the KeyWord Image Scale, would be very time con-suming, and beyond the scope of this initial study. Instead, weencourage readers to make a subjective judgment by interactingwith the publicly available demo search engine.

Image database statisticsTo show that the proposed method results in interesting and

promising image indexing, worthy of further investigations, weshow some statistics obtained from our test database containing5000 images. From hW belonging to each image, we save the fiveimage words that obtained the highest probability, and use themfor deriving image word probabilities for the entire database. Theresult can be seen in Fig. 5. The position of each circle corre-sponds to the position of the image word in the two-dimensionalKeyWord Image Scale, and the relative size of the circle repre-sents the probability. From the figure we notice that our imagecollection is biased towards cool, and slightly towards soft. Inearlier research ([15]), psychophysical evaluations of color im-ages showed similar results for another image database of muchlarger size. For our test database, the five most frequent image

3http://diameter.itn.liu.se/colse/

−3 −2 −1 0 1 2 3−3

−2

−1

0

1

2

3

< warm − cool >

< ha

rd −

sof

t >Figure 5. Image word probabilities for the entire test database. The po-

sition of each circle corresponds to the position of the image word in the

two-dimensional KeyWord Image Scale, and the relative size of the circle

represents the probability.

words are: urbane, quiet, delicate, precise, and chic. The mini-mum number of words that was assigned to an image is 9. Thethree most frequent Patterns are: ELEGANT, MODERN, andNATURAL. All these Patterns are located in regions with largecircles in the KeyWord Image Scale (see Fig. 5).

Next we investigate the distribution of image words for in-dividual images. If images are described by image words that arelocated rather close to each other in the KeyWord Image Scale, itwill indicate that the proposed extension of Kobayashi’s frame-work shows consistency when applied to images, reinforcing thatthe proposed method can be utilized as a tool for image index-ing. For every image in the test database, we obtain image wordpositions for the five words in hW with the highest probability.The standard deviation for all positions, belonging to all images,along the cool-warm and soft-hard dimensions respectively, are1.12 and 1.55. If we instead derive the standard deviation foreach image separately (still for the five image words with thehighest probability), the mean standard deviation over all imagesare down to 0.76 and 0.95 respectively. The distribution of stan-dard deviation scores for images in our test database are shownin Fig. 6. Values obtained show that for most images the spa-tial distribution of derived image words is much more compactthan, for instance, a random selection of image words. Finally,we show in Fig. 7 the five images with the lowest and the high-est standard deviation respectively, derived by the Euclidean dis-tance to the deviation [0 0]. Images with a low deviation containcolor combinations that can be found in a limited area of thetwo-dimensional KeyWord Image Scale, often corresponding toimages with a rather ”pure”, or well defined, content (in terms ofcolor semantics). Images with a high deviation usually containa lot of different color combinations, or color combinations thatare not defined by Kobayashi, which often results in image wordsthat are spread over the entire KeyWord Image Scale.

Summary and ConclusionsThe findings of this paper show that using Kobayashi’s

Color Image Scale on multi-colored images results in new and

CGIV 2010 Final Program and Proceedings 357

standard deviation

nr o

f im

ages

standard deviation

nr o

f im

ages

0 0.5 1 1.5 2 2.5 30

200

400

600

800

1000

1200

1400

1600

0 0.5 1 1.5 2 2.5 30

200

400

600

800

1000

1200

1400

(a) (b)

Figure 6. The distribution of standard deviation scores for image words po-

sitions (the five with the highest probability) for images in our test database.

The dimension cool-warm in (a), and soft-hard in (b).

(a)

(b)

Figure 7. Images in our test database that obtained the lowest (a), and

highest (b), standard deviation of image words positions. (Color images in

the online version)

interesting methods for image indexing based on high-level colorsemantics. The Color Image Scale was originally designed forsingle colors and three-color combinations. We show, however,that a modified approach for statistical analysis of images, in-volving transformations of ordinary RGB-histograms, results ina semantic image descriptor that can be used as a tool in bothimage labeling and retrieval. The descriptor can be translated,and used, in different levels of semantic information, spanningfrom image words to Patterns and positions on the scales hard-soft and cool-warm. It was shown that most images results in arather compact distribution of image words in the Color ImageScale. Labeling and retrieval results presented in this paper areonly illustrations. An objective measure of performance is dif-ficult to design since color-based high-level semantic conceptsare hard to define. Upcoming psychophysical evaluations willfurther established the usefulness of the proposed method. Aninteresting idea for future research is to investigate statistical re-lationships between the results obtained from the proposed in-dexing method, and image properties and keywords from moregeneral databases, for instance databases containing images andkeywords downloaded from the Internet.

AcknowledgmentsThe presented research is included in the project Visuella

Varldar, financed by the Knowledge Foundation, Sweden.

References[1] S. Berretti, A. Del Bimbo, and P. Pala. Sensations and psychological

effects in color image database. Proceedings of the 1997 Interna-tional Conference on Image Processing, volume 1, pages 560–563,Santa Barbara, CA, USA, 1997.

[2] S.-B. Cho and J.-Y. Lee. A human-oriented image retrieval systemusing interactive genetic algorithm. IEEE Trans Syst Man Cybern PtA Syst Humans, 32(3):452–458, 2002.

[3] D. Cohen-Or, O. Sorkine, R. Gal, T. Leyvand, and Y.-Q. Xu. Colorharmonization. In ACM SIGGRAPH 2006, volume 25, pages 624–630, Boston, MA, 2006.

[4] J.M. Corridoni, A. Del Bimbo, and P. Pala. Image retrieval by colorsemantics. Multimedia Syst, 7(3):175–183, 1999.

[5] J.M. Corridoni, A. Del Bimbo, and E. Vicario. Image retrieval bycolor semantics with incomplete knowledge. Journal of the Ameri-can Society for Information Science, 49(3):267–282, 1998.

[6] R. Datta, D. Joshi, J. Li, and J.Z. Wang. Studying aesthetics in pho-tographic images using a computational approach. In 9th EuropeanConference on Computer Vision, ECCV 2006, volume 3953 LNCS,pages 288–301, Graz, 2006.

[7] R. Datta, D. Joshi, J. Li, and J.Z. Wang. Image retrieval: Ideas,influences, and trends of the new age. ACM Comput Surv, 40(2),2008.

[8] R. Datta, J. Li, and J.Z. Wang. Learning the consensus on visualquality for next-generation image management. In 15th ACM Inter-national Conference on Multimedia, MM’07, pages 533–536, Augs-burg, Bavaria, 2007.

[9] R. Datta, J. Li, and J.Z. Wang. Algorithmic inferencing of aestheticsand emotion in natural images: An exposition. 15th IEEE Inter-national Conference on Image Processing, 2008. ICIP 2008., pages105–108, 2008.

[10] M.D. Fairchild. Color Appearance Models. Wiley-IS&T, 2005.[11] SY. Hong and HY. Choi. Color image semantic information re-

trieval system using human sensation and emotion. In ProceedingsIACIS, VII, pages 140–145, 2006.

[12] S. Kobayashi. Aim and method of the color image scale. Color ResAppl, 6(2):93–107, 1981.

[13] S. Kobayashi. Color Image Scale. Kodansha Intern., 1991.[14] J. Lee, Y.-M. Cheon, S.-Y. Kim, and E.-J. Park. Emotional eval-

uation of color patterns based on rough sets. In 3rd InternationalConference on Natural Computation, ICNC 2007, volume 1, pages140–144, Haikou, Hainan, 2007.

[15] M. Solli and R. Lenz. Color Emotions for Multi-Colored Images.Accepted for publication in Color Res Appl, 2010.

[16] W.-N. Wang and Y.-L. Yu. Image emotional semantic query basedon color semantic description. In Int Conf on Machine Learning andCybernetics, ICMLC 2005, pages 4571–4576, Guangzhou, 2005.

[17] W.-N. Wang, Y.-L. Yu, and S.-M. Jiang. Image retrieval by emo-tional semantics: A study of emotional space and feature extraction.In 2006 IEEE International Conference on Systems, Man and Cy-bernetics, volume 4, pages 3534–3539, Taipei, 2007.

[18] H.-W. Yoo. Visual-based emotional descriptor and feedback mech-anism for image retrieval. J. Inf. Sci. Eng., 22(5):1205–1227, 2006.

Author BiographyMartin Solli received a Master of Science in Media Technology and

Engineering from Linkoping University, Sweden, where he currently ispursuing a Ph.D. degree at the Department of Science and Technology(ITN). His research is focused on specialized topics within Content BasedImage Retrieval and image indexing, such as color emotions and high-level semantics, color harmony, and font recognition.

Reiner Lenz received the Diploma in Mathematics from the GeorgAugust University, Gottingen, Germany, and the Ph.D. degree from theElectrical Engineering Department, Linkoping University, Sweden. He isan Associate Professor at Linkoping University, Sweden. He has workedon the visualization of 3-D images and the application of group theoret-ical methods in image and signal processing and in computational colorimage processing.

358 ©2010 Society for Imaging Science and Technology