54

Color Perception of Clothes Images ― Observation Point and Perceived of Clothes ―

Saori KITAGUCHI*♯, Sae TANAKA*, Tetsuya SATO*,

Satoru HIROSAWA**, and Tadashi HAYAMI**

* Kyoto Institute of Technology, Kyoto, Japan ** Kyoto Municipal Institute of Industrial Technology and Culture, Kyoto, Japan

Abstract In our living environment, a single-colored object may not always have an even surface color. For example, displayed clothes have lighter and darker areas due to the shape of clothes and the environment. Despite this, we still understand them as a single-colored items. However, we cannot be sure whether or not we all interpret the color in the exact same way. Therefore, this study was carried out to understand how people perceive the color of an object by focusing on clothes. In this study digital images were used. The subjects’ observation point and perceived color were investigated similar to the environment of online shopping. Various styles of the clothes were used as samples but only single-colored clothes were selected as samples. The results of the study indicated the importance of the presentation methods. Presenting the clothes with no or little crease was appropriate for giving a common understanding of color. However, the subjects’ perception did not always agree with the representative color of the clothes. A large variation of color in clothing was found from the creased clothes or the clothes worn by a model. Furthermore, there was also a large variation of perceived color from the subjects. This large variation in perceived color indicates that one subject may understand a color differently from another subject even when looking at the same image. (Received September 25, 2017; Accepted February 28, 2018)

Key words: color perception, digital image, clothes, online shopping

(Journal of the Japan Research Association for Textile End-Uses, Vol.59, pp.534–541, 2018)

1. Introduction we can understand this as a single-colored object. Even if an object consists of a single color, However, if it is displayed differently, we may certain areas may appear different in terms of interpret it as a slightly different color. We color, influenced by our living environment. It is cannot be sure whether or not we all interpret due to a variety of display methods such as the the color in the exact same way. Therefore, the shape of an object and/or lighting and study was carried out in order to better background color which do not relate to an object understand our perception of the color of a but are factors of display environments. For single-colored object. instance, with displayed clothes in a shop there Visual assessments were carried out are lighter parts and darker parts even for focusing on observation point and perceived single-colored clothes. Despite these differences, color. Observation point is a location in clothes #Corresponding Author: E-mail: [email protected]

( 534 ) 0037⊖2072⊘2018⊘0700⊖0534 $ 01.00⊘0ⓒ2018 Jpn.Res.Assn.Text.End-Uses. 繊消誌 55 where people focus on /see carefully when they positioned at the center of the image sample. decide the color of clothes. Perceived color is the Since the shapes of the clothes were not all the color that a person interprets the clothes to be. same, the number of pixels belonging to each In this study, digital images of clothes were clothing image area was slightly different. used as samples. Real clothes could be used, but A subject was first asked to find an this would be more complicated than the use of observation point which was selected by clicking digital images for identifying the observation a point on the sample with a mouse. The location point and perceived color. In a shop, we can take of this point was recorded. The color at the an item of clothing in our hands and look at it observation point was then presented in a small whilst changing the shape or the lighting angle, box next to the sample as shown in Fig.1(b). The allowing us to observe in detail the item under subject was asked whether this color matched different environmental conditions. However, the perceived color; namely whether they during online shopping, we decide on the color of believed this was a true representation of the clothes from a few images on a computer screen color of the clothing sample. If the subject or smartphone display. Therefore, the agreed it was, then this color was recorded as differences in observation point and perceived the perceived color. If not, the subject was asked color between people could be larger. Therefore, to continue clicking until they found a color to visual assessments were carried out under best represent the clothing. conditions designed to represent online The total number of image samples were shopping. Trends in the characteristics of the 100. These images were obtained from online observation point and perceived color, and the stores (a list of the websites of the online stores factors which influenced them were discussed. can be found in the Appendix). All the samples used were single-colored clothes. Of all the 2. Experiments samples, 70 consisted of only the image of 2-1 Visual assessments clothes, of which 40 displayed clothes which Visual assessments were carried out under were not creased or only a little creased, normal room conditions with fluorescent categorized as the ‘non-creased’ sample, and 30 lighting (NEC: FLR40SEX-N/M/36-HG) as samples displayed creased or very creased shown in Fig.1(a). An image sample was clothes, categorized as the ‘creased’ sample. The presented to a subject on a standard monitor remaining 30 image samples consisted of clothes (ASUS Designo MX279H). A color whose worn by models. At the beginning of the (R), (G), (B) digital input values assessment, three random samples were (RGB values) were R=255, G=255, B=255, was presented for familiarity and the results of these displayed on the monitor and was measured samples were excluded from data analysis. The using a spectroradiometer (KonikaMinolta CS- 100 samples were then presented to subjects in 1000). The luminance of the white color was a random order. The participants comprised of 269 cd/m2 and the CIE coordinates 40 university students: 20 female and 20 male. [1] were x=0.31 and y=0.32. The interface for this visual assessment was created using C# (Visual Studio 2013). In Fig.1(b), an example of an assessment screen with a sketch of an image sample was given because of the copyright of the images. In an assessment, not sketches, but an image sample of a real clothes was presented to (a) (b) a subject on the monitor. All the image samples Fig.1 (a) Experimental settings (b) Example of had a size of 500×500 pixels including the white an assessment screen presented to a subject background. The clothes to be observed were with a sketch of an image sample

Vol.59 No.7(2018) ( 535 ) 56

2-2 Monitor characterization expressions can be written for the green and In order to obtain device independent color blue channels. X, Y and Z define the XYZ values, information of the samples [2], RGB values, the subscripts g and b stand for green and blue which were device-dependent values, were channels, max stands for the maximum output converted to a device independent color of CIE and Xk, Yk and Zk are the level (R=G=B=0) XYZ tristimulus values (XYZ values) [1] and output. then to CIELAB values [1] using a gain-offset- The model performance was evaluated with gamma (GOG) model [3]. A transfer model to 27 test whose RGB values were the obtain corresponding XYZ values of displayed combinations of the R, G or B digital input RGB values of the samples was derived from a values of 0, 127 or 255. The same as the training training data set. The training data set data set, the XYZ values of the test colors were * consisted of a set of RGB and XYZ values. The measured. The CIELABΔEab [3] RGB values were 18 neutral colors having the between these measured XYZ values and the same digital input values of R=G=B from 0 to XYZ values estimated from the RGB test color 255 with a 15 interval. Thus, it started from using the Eq.(1) and (2) were calculated. The R=G=B=0, and then R=G=B=15, until equations to calculate CIELAB values from XYZ * R=G=B=255. In addition, the primary colors values, and the color difference CIELABΔEab [1] which were red (R=255, G=B=0), green (R=0, from CIELAB values are given in Eq.(3) and (4), G=255, B=0) and blue (R=G=0, B=255), as well respectively. The average color difference CIELAB * as white (R=G=B=255) and black (R=G=B=0) ΔEab of these 27 colors was 1.65. Using these were used as the training data. These RGB equations, the RGB values of each pixel of the values were displayed one by one on the monitor, samples were converted to the XYZ values, then and the XYZ values were then obtained by the CIELAB values were obtained. measuring them using a spectroradiometer

(Konika Minolta CS-1000). The measurements L *116 f (Y /Yn ) 16

a *500[ f (X / X n ) f (Y /Yn )] were made under the same conditions as the Eq.(3) b * 200[ f (Y /Yn ) f (Z / Zn )] visual assessment. Then, the system gain, offset 2 2 C *ab  a * b * and gamma of the monitor was computationally 1 hab  tan (b * / a*) derived from the relationship of these RGB  ()1 / 3 if  (24 /116)3 f ()  3 values and the measured XYZ values (training (841/108)()16 /116 if  (24 /116) data set). The equations to obtain the XYZ values from the RGB values are given in Eq.(1) where X, Y, Z and Xn, Yn, Zn are XYZ values for and (2). a sample and for a reference white (white color

in this study), respectively.   r  R   R   g  o  if  g  o   0  r r  r r    2  2  2  255   255  Eq.(1) E *ab (L *1 L *2 ) (a *1 a *2 ) (b *1 b *2) Eq.(4) s r     R       0 if  g r or  0   255  where the subscripted number 1 and 2 indicates the two colors whose color difference is to be

X Xr,max Xk Xg,max Xk Xb,max Xk  sr  Xk  compared.         Y  Y Y Y Y Y Y s  Y Eq.(2)    r,max k g,max k b,max k   g   k             Z Zr,max Zk Zg,max Zk Zb,max Zk  sb  Zk  3. Results 3-1 Observation point where R is the digital input value of the red There were three types of sample: non- channel, and g r, o r andγr are the system gain, creased, creased and model-worn samples. The offset and gamma respectively, for the red observation points selected for most of the channel, r. Eq. (1) is for the red channel. Similar samples were around the body area of the

( 536 ) 繊消誌 57 clothes, not the sleeves, neck or edge of the the samples). Some of the samples showed clothes. However, the observation points were clearly lighter and darker areas. not concentrated in one area. They were rather Thus, the observation points were divided scattered over the body area as shown in among the subjects. Some of them focused on the Fig.2(a-c): (a) for the non-creased sample, (b) for lighter areas and others focused on the darker the creased sample and (c) for the model-worn areas as per the examples shown in Fig.2(d-f). It sample (note that these are only the sketches of is not easy to see from the sketch but the lighter and darker areas of the sample given in Fig.2(d) occurred because of the gathered fabric. Therefore, there was a repeat of the and dark areas at the front of the clothes. In the sample in Fig.2(e), the lighter and darker areas were made by a strong directional illumination onto a top left area of the clothing sample. Some

(a) (b) of the subjects finalized their color match from the lighter area on the left and others from the darker area on the right. In the case of Fig.2(f), the lighter and darker areas were created by the complicated influences of the illumination and the gathering of the fabric and curvy shapes due to the body shape and pose of the model. The results of these samples demonstrate that the

(c) (d) subjects were not always looking at the same color area. They see the different colors of areas depending on the styles of the clothes and the presentation methods.

3.2 Perceived color differences The differences of the perceived colors between the subjects were investigated. This (e) (f) was analyzed in terms of color difference CIELAB ΔEab*. The average ΔEab* of all the Fig.2 (a-c): Examples of the samples whose combination of the subjects’ perceived color was observation points were around the body calculated for each sample. In Fig.3, a histogram area, (a) the non-creased sample, (b) the of the ΔEab* of all the samples was given with the creased sample, and (c) the model-worn indication of the three sample groups. The non- sample. The black dots () indicate the creased samples showed relatively small color observation points of all the subjects. (d-f): Examples of the samples showed lighter and differences among the subjects compared with darker areas due to (d) the undulating of the other groups. It means that the subjects’ clothes, (e) a strong directional illumination perception of the color of these samples was onto a small area of the clothing area, (f) the similar. The creased and the model-worn curvy shape due to the body shape or pose of samples showed larger color differences. This the model. The open circles (○) indicate the could be because of the creases, body shapes and observation points located in lighter area and illuminations causing inconsistency between the cross (×) indicate the observation points observation points and hence the large scatter located in darker area. The lines in the clothes of data. The largest average color difference, indicate some of the crease lines ΔEab*, was 10.1. In apparel industry, ΔEab*

Vol.59 No.7(2018) ( 537 ) 58

representative color of the clothes. However, the histograms of the color of the clothes showed not only unimodal but also bimodal (double-peaked) and multimodal distributions. Of the 100 samples, 78 showed a unimodal distribution. Many of the non-creased samples showed a unimodal distribution. Most of the bimodal and multimodal distributions were found from the

Fig.3 The histogram of the ΔEab*of the creased or the model-worn samples. Similarly, samples with the indication of the three the histograms of the perceived colors also sample groups: the non-creased samples, showed not only unimodal but also bimodal or the creased samples and the model-worn multimodal distributions. samples In order to analyze the distributions of the histograms of the perceived color and color of the value of over 4 could fail the general color clothes, the variances of both the perceived color fastness quality assessment. Thus, this and color of the clothes, for the samples that only largeΔEab* value indicates a high level of showed a unimodal distribution were calculated. disagreement and a lack of common Then, the variances of the perceived color and understanding between subjects. This result color of the clothes of these samples were demonstrates a potential problem with online compared. It was found that there was high shopping and indicates the importance of the correlation (r = 0.73), indicating that a higher presentation of clothes. variation in the color of clothes results in a larger variation in the perceived color. This 3.3 Perceived color versus color of the clothes suggests that our perceived color tends to vary The perceived color and color of the clothes when clothes include a variety of L* values. For were analyzed. The perceived color is the color these samples, the L* values at the peaks of the that the subjects interpreted the clothes to be. distributions of the perceived color histogram, The color of the clothes means the colors (each and the color of the clothes histogram were pixel values) contained in the clothing area of compared. The L* values of some of these each sample. In order to analyze the color of the samples were very similar; The differences in L* clothes, the main area of the clothes was values (ΔL*) at the peaks were less than 4 as extracted; namely the clothing area excluding shown in Fig.4(a). But, in some samples, the L* buttons, tags, and the areas with different differences were more than 4 as shown in materials was analyzed. Histograms of the Fig.4(b). In most cases, the L* value at the peak CIELAB values of the main area of all the of the perceived color was larger than that of the samples were obtained. In comparison, color of the clothes; namely the perceived color histograms of the CIELAB values of the subjects’ at the peak of the distribution was a lighter color perceived colors of all the subjects were also than that of the color of the clothes. obtained. In this paper, only the results of the In other cases, the perceived color showed a CIELAB L* values were introduced as all other bimodal distribution, although the color of the CIELAB values were not showing noticeable clothes showed a unimodal distribution as an differences in characteristics. example is given in Fig.4(c). These samples were The samples used in this study were all either the creased or the model-worn samples. single-colored clothes. Thus, the histograms of In these samples, the small lighter color areas the sample colors were thought to show were observed in each clothing area. This could unimodal (single-peaked) and the peak of the be due to the influence of the illumination. distribution was thought to show a Because some of the subjects focused on that

( 538 ) 繊消誌 59 small lighter color area but the others focused As an example shown in Fig.4(d-e), the on the other darker areas, the distribution of the bimodal distribution was observed from the perceived color became bimodal. The color of the clothes. Most of these samples were distribution of the color of the clothes of these the creased or the model-worn samples. Two samples was also thought to be bimodal, different peaks were corresponding to the however it was often unimodal. The lighter color lighter and the darker areas on the clothing areas were relatively small, so that the number area. Another example given in Fig.4(d) is that of the pixels belonging to the lighter color area the lighter and the darker areas were found was not large enough to create any peak. at the upper and lower areas of the clothing,

(a) (b)

30 Color ofSample the clothes color Color of the clothes (%) 20 PerceivedPerceived color color Perceived color

10 Frequency 0 0 10 20 30 40 50 60 70 80 90 100 CIELAB L*

(c) (d)

15 30 Sample Color color of the clothes Sample color (%)

(%) Color of the clothes

Perceived color 10 Perceived color 20 PerceivedPerceived color color

5 10 Frequency Frequency 0 0 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100

(e) (f)

20 Sample color 15

(%) Color of the clothes Sample color

(%) Color of the clothes PerceivedPerceived color color 10 PerceivedPerceived color color 10 5 Frequency

0 Frequency 0 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 CIELAB L* CIELAB L* Fig.4 Examples of the CIELAB L* histogram of the perceived color and the color of the clothes with the sketch of the samples. (a) The perceived color and the color of the clothes both show a unimodal distribution. The L* values at the peaks of them are similar. (b) The perceived color and the color of the clothes both show a unimodal distribution. The L* values at the peaks are different. (c) The perceived color shows a bimodal distribution, and the color of the clothes shows a unimodal distribution. (d) The perceived color shows close to a unimodal distribution, and the color of the clothes shows a bimodal distribution. (e) The perceived color shows a random distribution, and the color of the clothes shows a bimodal distribution. (f) The perceived color and the color of the clothes both show a bimodal distribution

Vol.59 No.7(2018) ( 539 ) 60

respectively. These were created by the curvy for giving a common understanding of color, shape of the clothes due to the body shape of namely the perceived colors were similar among the model and perhaps by the direction of the subjects. For these presentations, the illumination. From the histogram, it can be seen subjects often observed color from the flat and that the perceived color was concentrated central area of the clothes. However, presenting around the lighter peak area of the color of the clothes without creases or curvy shapes could clothes. In the case of the Fig.4(e), the deep not be the solution for matching the perceived crease and its shadow (dark area) were found all color and the color of the clothes. The results over the clothing and they created a second peak indicated that the perceived color was not in the darker area of the histogram in addition always similar to the representative color of the to the expected peak of the clothing’s color. This clothes even when the subjects had a common case also showed that the perceived color was understanding of the colors. The perceived concentrated around the lighter peak area. colors of some of the samples were lighter than There were also samples whose the representative colors of the clothing samples. distributions of the perceived color and the color In this study, the perceived color of real clothes of the clothes both showed bimodal as shown in were not investigated. Therefore, it would be Fig.4(f). A larger peak was considered to be the interesting to carry out the investigation of most frequently appearing color for that observation point and perceived color of real particular clothing sample. The darker side of clothes to see whether the tendency would be the peak was large in this case. The perceived the same, i.e., whether the perceived color by a color spread widely but it did not appear subject will be a lighter color than the alongside the most frequently appearing color. representative color of a clothing image in real It spread from slightly lighter than the majority clothes. color toward the second peak area. The L* In this study, many of the creased and the differences of the perceived colors among the model-worn samples showed a wide spread of subjects was large. The L* differences were over data for both the colors of the clothes and the 20. Although the sample was not creased so subjects’ perceived colors. The differences were much and there was also a relatively large non- large enough for the subjects to perceive colors creased area at the center of this clothing area, of the clothes differently. The reasons were not many of the subjects did not judge the color from only because of the creases and curvy shapes of the central and non-creased area. In this sample, the clothes, but also illumination. The images the strongly illuminated area was found around used in this study were all extracted from the right side of the clothes. Thus, the color of existing online shops. Therefore, the the right side of the clothes was clearly lighter. relationship between the perceived color and Some of the subjects focused around that lighter these factors could not be analyzed, such as the color area. Therefore, the subjects’ perception degree of creasing and curvy shapes, the showed quite large variation. intensity and angle of the illumination and the size of illuminated or shaded area. Therefore, 4. Conclusions further study is necessary by using samples The observation point and the perceived with systematic changes of these factors to color for clothes were investigated similar to the identify the characteristics of disagreement or online shopping environment. Various types of agreement between observation point and clothes were used but only single-colored clothes perceived color among subjects, and also the were selected as samples. The results of the relationship between the color of the clothes and study indicated the importance of the the perceived color distributions. presentation methods. Presenting the clothes with no creases or few creases were appropriate

( 540 ) 繊消誌 61

References 1) FOREVER21; 1) CIE 015:2004 Technical Report, https://www.forever21.co.jp/ , 3rd Edition. CIE, Vienna 2) Topshop; (2004) http://www.topshop.com/?geoip=home 2) Jon Y. Hardeberg; Acquisition and 3) URBAN OUTFITTERS; Reproduction of Color Images: Colorimetric http://www.urbanoutfitters.com/urban/inde and Multispectral Approaches, x.jsp Dissertation.com, USA, p.32-37 (2001) 4) ZARA; http://www.zara.com/jp/ 3) Berns, RS.; Methods for Characterizing 5) UNIQLO; http://www.uniqlo.com/jp/ CRT Displays, Displays, 16 (4):173-182 6) Mont–bell; https://www.montbell.jp/ (1996) 7) GU; http://www.gu -japan.com/ 8) American Eagle; Appendix http://www.aeo.jp/top/CSfTop.jsp The websites of the online shops used in this study.

衣服画像の色認識

―視点と認識色―

(2017 年 9 月 25 日受付;2018 年 2 月 28 日受理)

北口 紗織*# 田中 沙英* 佐藤 哲也* 廣澤 覚** 早水 督**

*京都工芸繊維大学 **京都市産業技術研究所

―― 要 旨 ―――――――――――――――――――――――――――――――――――――― 私たちの生活の中では,単色のモノであっても,その見え方は必ずしも単色とは限らない.店 でディスプレイされている服は,その形や照明の影響などで,明るいところ,暗いところが出てく る.そのような状況であっても,私たちは容易に単色の服と判断することが可能である.しかし, どこを見て,どの様な色と認識しているのか,また,人々が同様の色として認識しているのかどう かはわからない.そこで,本研究では,衣服画像を用いて,服のどこを見て色を判断し,その色を どのような色と判断するのか調査を行った.実験では,オンラインショッピング環境を再現し,単 色の服には限定したが,様々なタイプ,提示方法の衣服画像を用い,評価実験を行った.その結果, しわを少なく提示したものに被験者間の色の共通認識が見られた.しかし,衣服の主な色と認識色 が必ずしも一致するものではなかった.また,しわのあるものや,モデルが着用したものは,色判 断の場所や認識色が被験者間で大きく異なる場合があるということが示された.

キーワード:色認識,デジタル画像,衣服,オンラインショッピング ――――――――――――――――――――――――――――――――――――――――――――

Vol.59 No.7(2018) ( 541 )