Semantic-Driven Selection of Printer Color Rendering Intents

Transcription

1 Semantic-Driven Selection o Printer Color Rendering Intents Kristyn Falkenstern 1,2, Albrecht Lindner 3, Sabine Süsstrunk 3, and Nicolas Bonnier 1. 1 Océ Print Logic Technologies S.A. (France). 2 Institut TELECOM, TELECOM ParisTech, LTCI CNRS (France). 3 School o Computer and Communication Sciences, EPFL, Switzerland. Abstract In this paper we introduce a uniied ramework that automatically selects the optimal color rendering intent or a given print job. We irst present how we extract inormation rom both the image eatures and the semantic inormation contained in keywords attached to this image. Then we show how our ramework uniies the two inputs to select the optimal ICC rendering intent. The ramework is evaluated with a psychophysical experiment on an image data set printed with the ICC media-relative colorimetric and perceptual intents using an Océ large ormat printer. We ind that our method is correctly able to predict the observers preerences in 81% o the images tested when the keyword is included compared to 58% when the keyword is not included. Introduction The aim o our research is to automatically generate optimal print reproductions o images using an inkjet printing system. The International Color Consortium (ICC) provides a consistent worklow to manage the color gamut changes between an original image and its reproduction via a given technology (ink or toner based print, silver halide photograph, electronic display). The ICC has deined our dierent intents to address the dierent reproduction objectives a user may have [1]. Each one represents a dierent color reproduction compromise. In this work we ocus on two o these intents, the perceptual and media-relative colorimetric intents. The media-relative colorimetric intent aims or a colorimetric match while the perceptual intent aims or a pleasing reproduction[1, 2]. The visual impact o the media-relative colorimetric intent is likely to cause the reproductions to appear more colorul and with more contrast and the perceptual intent will oten prioritize the details over other qualities. From this observation it appears that the selection, by a user looking or an optimal worklow, o one rendering intent versus the others may also depend on the document content. Perhaps or a very colorul image, the user pays more attention to the color reproduction over the details and chooses the rendering intent accordingly. For images with details, the user may choose another rendering intent that produces a better rendering o details. Our aim is to help the user by building a tool to automatically select the optimal rendering intent or a given print job. Much eort has been made towards creating an adaptive processing worklow where the inal processing is driven by the input document s content [3 5]. These adaptive worklows oten use a training set o documents which require two inputs: 1. eatures and 2. perormance input. The eatures used in these worklows help to summarize dierences between documents or to group documents into categories. The terms statistics, properties, actors, image characteristics and descriptors have also been used to describe a document s eatures. The perormance input is oten the result o a psychophysical evaluation. Our worklow requires the ability to easily change which ICC proile and rendering intent combination (color worklow) to apply. This requirement excludes the use o time-consuming psychophysical data as the perormance input. Instead, we use a set o perormance results derived rom metric tests, where each metric compares the color worklow perormance o a speciic perceptible quality attribute [6]. The psychophysical validation results used to test the quality o our irst implementation showed that the observers preerence between color worklows was more signiicant between rendering intents than between ICC proiles. For most test images, we were able to adaptively select the same rendering intent as the observers when the observers preerence between rendering intents was signiicant [7]. Our goal is to improve the automated selection o which rendering intent is optimal or documents that embed several conlicting image characteristics by using semantic inormation. The inclusion o the semantic inormation will add an understanding o the scene on a higher semantic level which will improve our prioritization o the conlicting characteristics. The aim o this paper is to introduce a uniied ramework that automatically selects the optimal rendering intent or a given print job. We irst present how we extract inormation rom both the image and the semantic keywords. Then we show how our ramework uniies the inormation to select the rendering intent. We then show some early results obtained with our ramework to demonstrate its advantages and compare these ramework results with the results o a psychophysical evaluation. A Uniied Framework or Image Features and Semantic Inormation This section explains in two separate subsections, how to extract cues rom two very dierent sources: 1) image pixels eatures and 2) semantic context. Ater this we complete the ramework by uniting the two methods into a single estimation that can be used or the selection o the best rendering intent. Cues rom Numeric Pixel Values We use eight quality attributes, which are Colorimetric Accuracy (CA), Colorulness (CO), Gamut Boundary (GB), Smoothness (SM), Details (DE), Shadows (SH), Highlights (HL), and Neutrals (NT), respectively [7]. We denote the set o all quality attributes Q. Each quality attribute is represented by 1 expert-selected example images. For a new input image we assess its relatedness to a quality attribute by measuring similarity to its set o example images. This is a typical classiication task and we implement a standard method with multivariate Gaussians. 2th Color and Imaging Conerence Final Program and Proceedings 323

2 Mathematical Background Given an example set o images, or each quality attribute, we pre-compute a eature vector or all o the images. The eature vector can contain any type o pixel-based image descriptors, such as lightness, color, or texture eatures. As a irst step we whiten the data by subtracting the mean and dividing through the variance in each dimension separately. The mean and variance are computed globally over all quality attributes q and images. For the rest o this paper the whitened eatures are used without being reerred to as whitened. We estimate or each quality attribute the mean (µ q ) and covariance (Σ q ) matrix o its associated point cloud o images in the eature space. Each quality attribute can then be represented as a multivariate Gaussian density distribution in eature space: g q ()= 1 (2π) N/2 exp Σ q 1/2 [ 1 ( ) T ( ) ] 1 µ q Σq µ q (1) 2 where is a point in the whitened eature space, while and ( ) 1 are the determinant and inverse o a matrix, respectively. For a new input image I we compute its eature vector I and use the outcome o the Gaussian density estimators g q ( I ) to quantiy its relatedness to each quality attribute. Because it is based on numeric pixel values we call this numeric relatedness and denote it as q (I). Since an image relates to the distinct quality attributes in dierent proportions, we always scale the relatedness values so that they sum to 1%, i.e. q (I)=1%. Practical Implementation The eatures or a classiication task have to be chosen with respect to the intended application. In our printing worklow, the key image characteristics are texture and color. Texture is important because it relates to image sharpness that is limited by the print resolution. Color is crucial in two ways: gray levels have to be careully reproduced since they are sensitive to color casts and very saturated colors might cause diiculties due to a printer s limited color gamut. Consequently we use eatures that describe the indicated characteristics. We convert an input image I to CIELCH color space and compute the mean values o the lightness L and chroma values C and the entropy o the lightness channel E, respectively. Entropy is a statistical measure o randomness and can be used to characterize image textures. The complete eature vector is I = [ L, C, E ] T. Figure 1 shows two example images, N7 (top) and N1 (bottom) [8, 9], and their computed relatedness values. It is visible how image characteristics are relected in the values. The top image N7 contains many details and saturated colors and thus the Details (DE) and Colorulness (CO) quality attributes have the highest scores. The bottom image N1 has many neutral colors and details on the bride s dress and bouquet as expressed by high scores or Neutrals (NT) and Details (DE). Cues rom Semantic Image Context The semantic context o an image can be given by associated keywords in the EXIF ile header. Example annotations o the images in Figure 1 are colors and blackwhite, respectively. Adding this inormation to the decision process can add an understanding o the scene on a higher semantic level. For the N7 image, a Figure Two input images, with associated numeric relatedness values or the eight quality attributes (horizontal axis). Top: the content o the image N7 relates to Details (DE) and Colors (CO). Bottom: the image N1 contains details (DE) and many neutral colors (NT). keyword colors indicates that the colors are important to the user and thus the relatedness to Colorulness (CO) has to be higher at the expense o Details (DE). The goal is to estimate or a given keyword its relatedness to the eight quality attributes, as we did or a given image. The challenge is that keywords come rom an uncontrolled vocabulary and we do not want to limit it, because this has considerable drawbacks. One drawback is the diiculty to choose a subset that is small enough to handle, but large enough to cover the semantic space. Our statistical approach handles vocabulary sizes in the order o thousands so that we can cover a signiicant part o the semantic space. We use a statistical signiicance test to estimate the relations between the three image eatures and an arbitrary keyword [1]. The test is based on the MIR Flickr database, which contains one million high quality images rom Flickr along with their annotations given by the internet community [11]. Theoretical Background We start by deining the set o all eature values (or each eature) rom images with a speciic keyword w in the annotation string and designate a second set or the remaining images, the disjoint set. In order to quantiy how a keyword w inluences the given image eature, the eature distributions o the two disjoint sets are compared. This is accomplished with the Mann-Whitney- Wilcoxon ranksum test that assess whether the medians o two populations dier [12, 13]. We have a means o measuring the signiicance a given keyword has on the eature vectors by determining the signiicance value (z) o the normalized ranksum statistic (T), with its expected mean (µ T ) and variance (σt 2 ), [1]. The used MIR Flickr database contains one million annotated images [11], which makes it possible to compute signiicance values or a large number o keywords. There are 2858 keywords in the database that occur in at least 5 image annotations. Based on our experience, this is a scale where statistical estimations are robust enough to be exploited in this ramework Society or Imaging Science and Technology

3 The signiicance value z w indicates whether the underlying eature has the tendency to be larger (z w > ), smaller (z w < ) or unaected (z w ) or images annotated with keyword w. To give a better intuition about the z values we show some examples that appear in the database. For the lightness mean we obtain or example z L shadows = 21 and z L design = 25, or the chroma mean z C blackwhite = 86 and z C colors = 63, and or the entropy z E red = 28 and z E trees = 27. Keywords with z values close to zero are z L water =.41, z C place =.1, and z E libraryocongress =.4, respectively. The example z values clearly show that the signiicance measure is a solid estimate or what a keyword implies or an image eature or what it does not imply. Practical Implementation We propose a simple yet eective way to compute the relatedness o a keyword w to the eight quality attributes using signiicance values. The example image set o the quality attribute q is centered at µ q = [ µ L q, µ C q, µ q E ] T in the training eature space. We then deine or all eatures F = { L, C,E} the minimum and maximum values: m M = min µ q = max µ q F (2) I a z w value or keyword w and eature is positive, the quality attribute q that has the maximum mean value M q is most related to the keyword and the quality attribute with the minimum mean value m q is least related. For a negative z w value, it is the opposite. We thus estimate a keyword w s semantic relatedness q (w) to a quality attribute q as: q (w) = q(w), (3) µ q m q z q(w) Mq m w zw q = q Q, F (4) µ q M q z Mq m q w zw < Similar to the values we scale also the values so that q (w)=1%. they sum to Figure 2 shows the semantic relatedness values or our keywords colors, nature,, and blackwhite, respectively. The bar plots show that colors and nature are both related to Colorimetric Accuracy (CA), Colors (CO), Gamut Boundary (GB) and Smoothness (SM) whereas and blackwhite relate stronger to Details (DE), Neutrals (NT), Highlights (HL) and Shadows (SH). The keyword blackwhite is only very weakly related to Colors (CO) and not at all to Gamut Boundary (GB). The plot also shows that two dierent words can stand or similar image characteristics in terms o ICC rendering intent. The keywords colors and nature both imply colorul images. As the statistical ramework detects these relationships automatically on a large scale, it is not necessary to deine them as synonyms. Uniication o Numeric and Semantic Cues In the two previous subsections we have presented methods to estimate the relatednesses o quality attributes to numeric im nature Figure colors blackwhite Semantic relatedness values or our keywords as indicated in the titles or the eight quality attributes. The keywords colors and nature both imply the importance o Colorulness (CO) and Gamut Boundary (GB), whereas and blackwhite relate stronger to Neutrals (NT) and Details (DE). Note that colors and nature are similar in terms o their meanings or ICC rendering intents. age pixels based eatures and to semantic image context, respectively. The numeric relatedness q (I) takes as input an image I, and the semantic relatedness q (w) a keyword w. In order to estimate how an image in a speciic context relates to the quality attributes, the values have to be united in a single relatedness measure r q (I,w). Because the and values express probabilities, it is reasonable to use a multiplication: r q (I,w)= q (I) q (w), (5) In this early work, multiplication has been the only way used to combine the relatedness measures, other methods should be investigated or any uture work. Figure 3 reproduces the two example images rom Figure 1, but with an associated semantic context given by a keyword. The bar graphs show the numeric q (I) and semantic relatedness q (w) as well as the united relatedness r q (I,w). Note how the relatedness values adapt to both the image and its context, enabling an automatic semantic-driven selection o the printer ICC rendering intent. From Relatedness Measures to Rendering Intents We are interested in understanding how the keywords impact the choice o rendering intent, which includes the numerical relatedness () and the uniied relatedness (r), and no longer the semantic relatedness () results. We combine the metric ranked results (rri) rom Table 1 irst using the numeric relatedness () results and then the uniied relatedness (r) with the ollowing: χ k, j = (k rri) q, j, k {,r} (6) where rri is the ranked metric score o the two rendering intents, q is the quality attribute index, j is the rendering intent index 1:m, and m = 2 the number o rendering intents in this experiment. The inal selection (S k ) is determined by: S k = argmin(χ k, j ). Our j method o combining the metric results and the relatedness scores 2th Color and Imaging Conerence Final Program and Proceedings 325

4 (a) semantic context: colors r (b) relatedness values r Table 1: The evaluation results o each quality attribute. Each quality attribute is listed with the metric used to assess it, the type o document used in the assessment, the results and the ranking o the Perceptual (Per) and Media-Relative colorimetric (MR) rendering intent, [7]. Quality Attribute Metric Per(rRI) MR(rRI) Color Accuracy CIE E* 94 (in gamut target) 3.79(2) 2.63(1) Colorulness Cui [14] (targets) 1.81(2).24(1) Gamut Boundary CIE LCHab [6] (out o gamut gradients).93(2).95(1) Smoothness 2 nd Derivative[15] (gradients) 2.73(1) 3.(2) Details Visual Inormation Fidelity [16] (images).21(1).26(2) Shadow Details CIE L* STDV [7] ( L*3).15(1).48(2) Highlight Details CIE L* STDV [7] ( L*75).2(1).3(2) Neutrals CIE C H ab [6] ( L* target).71(2).66(1) (c) semantic context: blackwhite Figure 3. (d) relatedness values Same images as in Figure 1, but with a semantic context indicated by the keyword in the caption. The bar graphs show the numeric q (I), semantic q (w) and united relatednesses r q (I,w), respectively. Top: the semantic context colors reduces the relatedness to Details (DE) and increases it or Color (CO) and Gamut Boundary (GB). Bottom: the semantic context blackwhite makes Details (DE), Highlights (HL) and Neutrals (NT) the dominant quality attributes. has not yet been proven as the most optimal. Other methods should be considered and compared or uture work. Since and r have been scaled so that they sum to 1% then the ollowing holds true, m j=1 χ k, j = m (m+1) 2 and the scores rom 6 can be scaled so the sum o χ k is equal to 1%. Let the index j= 1 correspond to the perceptual intent and j=2 correspond to the media-relative colorimetric intent, so that i m = 2 then i χ k,1 < 5% the perceptual intent is selected and i χ k,1 > 5% the media-relative colorimetric intent is chosen. Additionally, when m = 2 then χ k,1 can be inerred rom χ k,2 and vice versa or the same k. For the rest o this paper we reer to only the perceptual intent χ k,1 when discussing the rendering intent scores. To simpliy we set χ = χ,1 and χ r = χ r,1. We express the keyword s impact on our selection as, d = χ r χ. When d > it is illustrated with, indicating a move towards the media-relative colorimetric intent and when d < a is used to illustrate a move towards the perceptual intent. Experimental Validation For our irst implementation we chose to compare the perceptual and media-relative colorimetric rendering intents or a custom ICC v2 output printer proile. The ranked metric results used in Equation 6 are summarized in Table 1, along with a description o the metrics used and their results. As is expected, the Perceptual (Per) intent perorms better with the more details related attributes, while the Media-Relative intent perorms better with Color Accuracy and Colorulness. Demonstration o the Framework We irst consider the images in Figure 3 to compare the numeric ramework that uses only the image eatures to the uniied ramework that uses the semantic inormation along with the image inormation. As shown in Table 2, the addition o colors changes the selection or image N7 rom the perceptual intent (< 5%) to the media-relative intent (> 5%). However, when the keyword is shop the selection does not change, it moves urther towards the perceptual intent. For the N1 image, i the ramework uses blackwhite the selection does not change, the inclusion o causes a change in selection to the perceptual intent. Table 2: A comparison o the scaled results obtained with our ramework, irst the selection with the numeric solution and then the uniied ramework. The Perceptual (Per) intent is chosen or s below 5% and the Media-Relative colorimetric (MR) intent when above. image numeric keyword uniied 39% Per 54% MR colors shop blackwhite 52% MR 22% Per 58% MR 49% Per Image & Keyword Selection The results rom our past psychophysical experiments have shown that or many images the observers and the automatic selection resulted in the same rendering intent [7]. However, images which embed several conlicting image characteristics resulted in the observers having a less signiicant choice o which rendering intent to choose and reduced the correlation between the automatic selection and observer results. For this experiment, we intentionally chose these more challenging images, which exhibited conlicting characteristics, since these are the images which are most likely to be impacted by the addition o the keywords. O the 26 test images, 16 o them were used in our past experiments. 1 additional images were included, which also have conlicting characteristics. From the 2858 keywords which appeared in the MIR Flickr database at least 5 times, we chose 26 o them or the demonstration o the ramework. The number o times the words appeared within the set ranged rom 551 (shapes) to (sky). We chose 14 keywords which changed the automatic selection, 7 changed the selection rom media-relative to perceptual and 7 rom perceptual to media-relative. The remaining 12 keywords either increased the certainty o the rendering intent choice or moved the choice closer to the opposite selection without causing a change o selection. Additionally, words were subjectively Society or Imaging Science and Technology

5 paired with each test image, which were considered to be relevant to the content o the given image. Keywords 1% Observer Mean Preerence Media-Relative 5% Perceptual % Images Figure 4. details trees ceiling shadows vehicle preerence without keyword change o preerence when keyword added depth roses ur blue color skysot women text design preerence with keyword (towards media-relative) preerence with keyword (towards perceptual) advertising contrast edge shapes red bubbles metal kids vivid old Plotted are the average observer preerences or each image both without and with the keyword. An arrow is plotted when the average preerence changed with the inclusion o the keyword. I the arrow head points down, the preerence moved towards the perceptual intent. I the tip o the arrow head is below the line the observers avored the perceptual intent and i its above the line the observers preer the media-relative intent. Thumbnail examples o the images and the keywords given with each sample can also be ound on the plot. Psychophysical Experiment To validate our automatic smart selection o rendering intent, we asked a set o observers to give their preerence o over image quality between an image printed with the perceptual intent and the media-relative colorimetric intent. The reproductions or the test were printed on the Océ ColorWave 6 wide ormat printer with the LFM9 uncoated Océ Top Color Paper. It was o interest to know i the preerence changed when the image pair was given with the selected keyword versus without. To accomplish this, we divided the observer pool o 2 Océ employees into two sets o 1. A survey was conducted to help divide the pool o observers into two roughly equivalent sets, based on: age, occupation, nationality, country o origin, native language, luency in any additional languages, experience assessing image quality, interest and experience in imaging and photography. The experiment was conducted entirely in English and all keywords were English. The observers were presented with 54 pairs o documents and asked to choose which one o the pair they most preerred or overall image quality. Image quality was deined or them as: the impression o the overall merit or excellence o an image, as perceived by you [17]. The observers judged each pair o images twice, once without a keyword and once with a keyword. When the keyword was included, the observers were told that The word was given to the sample by the creator. Please irst read the word and then assess the pair o samples and choose which one o the pair you most preer or overall image quality. The pairs o samples were viewed in the Judge II viewing booth at a color temperature o 515 Kelvins and illumination level o 115 ± 75 lux and a distance o approximately 6 cm. The observers were allowed to move the samples around reely. The evaluation was conducted in a room without windows and the lights were turned o, the room is designed and operated or visual testing with neutral gray surround, walls and counter-tops. Psychophysical Results The results o the psychophysical evaluation are summarized in Figure 4. The perceptual intent (< 5%) was avored or 8 o the 26 test images when a keyword was not given and 1 when the keyword was given. The average preerence changed rendering intents or 4 images, 3 to perceptual and 1 to media-relative colorimetric. The test was a orced choice, so we are unable to determine i the points close to the rendering intent threshold (5%) are there because the observers had conlicting preerences or their preerences were not signiicant. This question will need to be investigated in the uture work. Keywords 1% no kw with keyword Observer Mean Preerence Media-Relative 5% Perceptual % Images Figure / 14 Images Automated Selection = Observer Mean Preerence roses old sot ur blue color sky text advertising red vehicle women metal 1% 5% In this igure we ocus on the 14 images where the observers or the automated selection changed which rendering intent to select with the inclusion o the keyword, a subset o the 26 images used. The observer choices are in blue and to the let o the automated selection which is in orange. The irst 4 images (beore the dotted line) are the instances where the rendering intent selection changed with both the observers and the automatic selection. For all 4 images, the automated selection chose the same as the observers. The next 8 images (beore the solid line) are the instances where the automated selection changed to the same rendering option as the observers when the keyword was added. The urthest 2 images to the right are the only instances where the addition o the keyword negatively impacted the rendering intent selection. Assessing the Framework Results In total, our ramework predicted the same rendering intent as the observers or 81% o the images when the keyword was included, an improvement compared to 58% without. In Figure 5 we ocus on the 14 images where either the observers or the automated selection changed with the inclusion o the keywords, 12 o which were in agreement with the observers. There are 4 instances where the observers mean choice changed rendering intents, or these images the ramework selected the same rendering intent both with and without the keyword. There were only two instances where the inclusion o the keyword had a negative impact on the automated selection. In general, the keywords had a greater impact on the automated selection than with the observers preerences. Future work may include investigating which key- % Automated Selection 2th Color and Imaging Conerence Final Program and Proceedings 327

6 Observer Mean Preerence 1% 5% Figure 6. no keyword Pearson s correlation =.19 5% 1% Automated Selection Observer Mean Preerence 1% 5% keyword Pearson s correlation =.57 5% 1% Automated Selection Plotted are the results o the automated selection versus the psychophysical (observer mean preerence) results, both without the keywords and also with the keywords. The keywords, or most images, improved the results by moving the points closer to the diagonal line, which is the ideal. word and image combinations impact the automated tool versus combinations that impact the observers preerences. Pearson s correlation coeicient was used to describe the relationship between the ramework results and the observer results. The coeicient indicates the linear relationship between two variables on a scale o +/- 1, the closer the values are to +1 the better the correlation is. The correlation without the keywords was.19. When the keywords were added the correlation increased to.57 and the directional correlation was.46. Again, images which were expected to be diicult or both the observers and our ramework were intentionally chosen. It is expected that images without the conlicting characteristics would result in higher correlations. We have illustrated the improvements in Figure 6, which shows the improved correlation o the automated selection when the keywords were included. Conclusion In this paper we introduced a uniied ramework that automatically selects the optimal color management settings or a given print job. The ramework takes cues rom both the image pixels and the semantic context. The cues rom the image pixels are represented in the orm o a eature vector and we use a multivariate Gaussian classiication engine to compute the numeric relatedness to eight quality attributes. The semantic cues are given by a keyword associated with an image. We use a statistical ramework to relate the keyword with the same eight quality attributes. Finally, we uniy the relatedness values rom the numeric and the semantic cues. The uniied relatedness values represent the image together with its semantic context. Our uniied ramework has improved the results o our automated selection. We were able to correctly predict the observers preerence or 81% o the images tested when the keyword was included, compared to 58% when the keywords were not included. The correlation o the automated selection to the observer choice improved to.57 rom.19. Future work includes investigating other ways o uniying the two methods. We will also consider other types o image eatures. Another area or uture work is to investigate what types o keywords have the most impact on either the observer choice or the automated selection. Author Biography Kristyn Falkenstern is a PhD student under the direction o Hans Brettel at Télécom ParisTech, sponsored by Océ Printing. In 29, she completed her MSc at the London College o Communication, where she studied various methods o inding spectral relectance estimates. She worked rom 22 to 28 in an Image Science team at the Vancouver acility o Hewlett-Packard. She also has a BS in Imaging rom Rochester Institute o Technology. Reerences [1] International Color Consortium. White paper 9: Common color management worklows and rendering intent usage [2] International Color Consortium. White paper 2: Perceptual rendering intent use case issues [3] P. Sun and Z. Zheng. Selecting appropriate gamut mapping algorithms based on a combination o image statistics. In SPIE/IS&T Electronic Imaging, volume 5667, 25. [4] T. Newman, T. Kohler, and J. Haikin. Dynamic gamut mapping selection, 25. US Patent 6,947,589. [5] A. Golan and H. Hel-or. Novel worklow or image-guided gamut mapping. Journal o Electronic Imaging, 17: , 28. [6] K. Falkenstern, N. Bonnier, H. Brettel, M. Pederen, and F. Viénot. Using metrics to assess the ICC perceptual rendering intent. In SPIE/IS&T Electronic Imaging, volume 7867, 211. [7] K. Falkenstern, N. Bonnier, H. Brettel, M. Felhi, and F. Viénot. Adaptively selecting a printer color worklow. In IS&T CIC, 211. [8] ISO Graphic technology Prepress digital data exchange Part 2: XYZ/sRGB encoded standard colour image data (XYZ/SCID). ISO, 24. [9] ISO Graphic technology Prepress digital data exchange Part 3: CIELAB standard colour image data (CIELAB/SCID). ISO, 27. [1] A. Lindner, N. Bonnier, and S. Süsstrunk. What is the color o chocolate? - extracting color values o semantic expressions. In IS&T CGIV, 212. [11] M. Huiskes and M. Lew. The MIR lickr retrieval evaluation. In ACM MM, 28. [12] F. Wilcoxon. Individual comparisons by ranking methods. Biometrics Bulletin, 1(6):8 83, [13] A. Lindner, A. Shaji, N. Bonnier, and S. Süsstrunk. Joint statistical analysis o images and keywords with applications in semantic image enhancement. In ACM MM, 212. [14] C. Cui and S. Weed. Colorulness? in search o a better metric or gamut limit colors. In IS&T PICS, 2. [15] P. Green. A smoothness metric or colour transorms. In SPIE/IS&T Electronic Imaging, volume 687, 28. [16] H. Sheikh and A. Bovik. Image inormation and visual quality. IEEE Transactions on Image Processing, 15:43 444, 26. [17] ISO Photography - psychophysical experimental methods to estimate image quality - part 1: Overview o psychophysical elements. ISO, Society or Imaging Science and Technology