Package design gives a strong first impression to customers. Package design includes various factors. In this research, we focus on color patterns and fonts included in packages. The impressions on 45 sensory scales are analysed based on the semantic diferential method with respect to diferent font and color patterns using two-way ANOVA. Then a regression model is build to predict impressions by support vector regression using extracted features. As a result of ANOVA, main efects were found between 45 scales and colors, and 3 scales and fonts. Also, mutual interactions were found in 2 scales. In impression prediction, our regression model marks high coeficients of determination (> 0.56) on validation data which means our model is efective in predicting impressions.
on only colors or on only colors and layout. As far as the authors know, there is no research investigating afective When we see package design of products, various im- efects with both fonts and colors which are essential pressions and feelings including cute, natural, and cool and important elements in package design. Regarding are evoked. The evoked impressions and feelings are af- researches on prediction of impressions evoked from fected by multiple design elements such as colors, fonts, package design, most of them use only one indicator and layouts in the design. Therefore, it is important to suck as favorability. No research addressing quantitative clarify the relationship between design elements and af- prediction of impressions with multiple indicators has fective efects for package design which is used to give been reported. impressions connected to the products and to increase In this research, we attempt to quantitatively visualize costumers’ buying motivation. efects of elements in fonts and colors used in package
There are many researches reporting on design ele- design on evoked impressions from it, and to predict ments so far: psychological and feeling efects caused by impressions of package design with multiple indicators single color[1, 2], relationship between two-color combi- using support vector regression (SVR). Impression prenations and afective efects[ 3, 4], mapping of relation- diction using machine learning technique can estimate ship between afective words and color combinations[ 5, impression evaluation which is hardly indiscernible with6], and analysis and investigation on relationship among out large-scale market surveys. As a result, quick and afective efects, colors, and shapes including fonts[ 7, 8, flexible customization of package design meeting cus9, 10, 11, 12]. tomer needs will be available.
Regarding package design, many researches about analysis and investigation on impression evaluation and buying motivation for existing package design have been 2. Methods of subject experiments reported[13, 14]. Recently not only evaluation of impressions but also prediction of impressions using deep In this research, subject experiments were conducted learning and heat-map visualization of regions in pack- to analyse efects of fonts and colors on impressions of age images that strongly afects impressions have been package design and to collect data for training of SVR for attempted[15, 16]. However, most of the previous re- impression prediction of package design. The procedure search reporting on impressions of package design focus to generate the image data set used in the experiments is as follows. 1. Selection of original package images. 2. Selection of fonts and conversion of fonts in the
original package images. 3. Selection of color combination and conversion of color combinations in the images converted in the previous step.
stances used in the subject experiments under the conditions that information on the products and brands is hard to be inferred and the products are intuitively purchased. The main elements of the packages are color combinations and letters. Six packages are selected as shown in Table 1.
the previous research [17] and included with Microsoft Windows under the condition that they have large diference in font features which is described in Sec. 3.1. The selected fonts are shown in Table 2. Only the letters in the selected package shown in Table 1 are converted into the selected fonts, considering that original layouts and sizes of letters are not changed. Thus, 24 in total package images are generated. to the 24 sets of images with the converted fonts mentioned in Sec. 2.2. There are 6 kinds of package desings, four kinds of fonts, and 4 kinds of color combinations. Therefore, 96 images were generated in total. The generated images for a certain design are shown in Table 4.
164 subjects took part in the experiments: 118 men and 2.3. Selection and conversion of colors 46 women. The age ranges from 20 to 67 years old: the average is 41.5 and standard deviation is 11.6. The subColor combinations are selected based on the image scale jects is instructed to answer 7 point Likert scales for 45 system developed by Nippon color and Design Research adjective pairs shown in Table 5 . Two of the adjective Institute Inc.[5]. There are four axes, soft, hard, warm, pairs are related to buying motivation and the others are and cool in the system. The color combination in each selected referring to the reference [17]. The category of axis is selected as a representative. Table 3 shows the the products was not taught to the subjects in advance selected color combinations. The package images with in order to decrease bias for evaluations generated by diferent color combinations are generated by applying subjects’ knowledge. the selected color combinations using k-means clustering Each subject evaluated only one instance for each original package shown in Table 1. 16 images for one original package were evaluated with 16 diferent subjects. A data set of 16 images for each original package were evaluated 9 to 11 units of 16 subjects. The order of presentation of images were shufled as much as possible according to the package designs, fonts, and color combinations in order to reduce the order efects.
1 Contrast (one dimension)
Deference between the brightness of the letters and background. 2 Line width (one dimension)
Ratio of line region and background region. 3 Circularity (one dimension)
Complexity of convex hull 4 Center of gravity (two dimensions)
Coordinate of center of gravity of convex hull 5 Gradient (one dimension)
Gradient using robust estimation [18] 6 Aspect ratio (one dimension)
Ratio of hight and width of bounding rectangle 7 Edge feature value (20 dimensions) Feature values of edge calculated using 20 kinds of 3 × 3 of mask patterns
1 means the deference between the brightness of 90% of data obtained in the subject experiments for the letters in the package and the average brightness 96 kinds of generated package designs are used for the of the background over the whole images. The values training of the SVR model. 10% of data are used as test are calculated using letters written only in HGSTE Kaku data to evaluate the accuracy of the model. Five-fold Gothic U in the gray scale images of the pacakges shown cross validation is conducted. The Gaussian kernel is in 1. The higher the contrast value, the higher 1 is. used as the kernel for the SVR model. Each of 20 values The wider the line width is, the higher 2 is. 3 gets are prepared for hyper-parameters C and and tuned higher as the letter is circular shape. 5 gets closer to 90 with grid search to obtain a combination of precise hyperdegree as the letter tilts. 6 gets closer to 1 as the letter parameters. is set to the reciprocal of the number of get vertically longer. From 2 to 7 are the averages of features. values calculated with several alphabet characters mainly used such as for the name of products in each package.
For example, regarding package A, the six characters, B, A, L, T, E, and R, in the four types of fonts were used to calculate the values of 2–7. The averages of the values are used as the font features for package A.
is calculated as follows. Top three colors which occupy the package are divided into ten areas based on each area ratio. The values of RGB included in the each area are the color features. The example of extraction of color features is shown in Fig. 2
Fig. 3. Features of fonts and colors are input into SVR and 45 scales of adjective pairs are output. The output obtained by inputting features of both fonts and colors into the SVR model are compared with each output obtained by inputting features of either fonts or colors. In this research, standardization of features is conducted as reprocessing.
the analysis target if the variance of all the answered scales are less than 0.5. Two-way ANOVA with two factors of font and color to each of 45 adjective pair scales was conducted to quantitatively estimate the efects of fonts and colors on the impressions of package design. As an example, the results of the ANOVA analysis to the scale of "Masculine – Feminine" are shown in Table 6. The null hypothesis are that the impressions of package design are not afected by both fonts and colors and that there is no efects by the interaction between fonts and colors. As a result, each of fonts and colors afects the impressions while interaction between fonts and colors were not observed.
The summary of all the scales are as follows. The significant diferences were observed on all of the 45 scales regarding colors. The significant diferences were observed on only three scales, "Masculine – Feminine," "Regular – Irregular," and "Young – Old" regarding fonts. The significant diferences were observed on two scales, "Regular – Irregular" and "Young – Old" regarding the interaction between fonts and colors. As a resutl, certain efects were observed in all the scales of adjective pairs.
the color combination for package A. The color combination is composed of blown and black in the background and beige in the font. The value of 1 resulted in 32.99 when Warm is used as the color combination for package A. The color combination is composed of orange and yellow in the background and red in the font. The value of 2 gets highest when HGSTE Kaku Gothic U is used as the font. The value of 3 gets highest when Cooper Black is used as the font. The value of 6 is highest to HGSTE Kaku Gothic U and close to 0.5 when Cooper Black and Broadway are used. The value of 6 gets highest to Lucida Caligraphy. It seems that the features overall correspond to intuition.
Regarding the model trained with features both fonts and colors, correlation coeficients more than 0.7 were the SVR model with the features of fonts and colors is obtained in 11 scales of the 45 adjective pairs, and cor- also presented. relation coeficients between 0.4 and 0.7 were obtained In the subject experiments, the scores for the scales in 18 scales. The former 11 scales include the two scales, of 45 adjective pairs were collected and it was shown "Masculine – Feminine" and "Violent – Gentle," which that these scales can be used to explain the efects of are included in the three scales in which significant dif- fonts and colors on impressions of package design. 45 ferences were observed in the results of ANOVA on the adjective pairs for the factor of colors, 3 adjective pairs factor of fonts. for the factor of fonts, and 2 adjective pairs for interaction
The accuracy averages of 11 scales whose correlation between fonts and colors were observed. coeficients were more than 0.7 are summarised in Ta- In the evaluation of the SVR model, it is confirmed bles 7 –9. The average of correlation coeficients to the that the features to train the model are valid for the pretest data using both features showed the highest value. diction of impressions of package design. Meanwhile, Meanwhile, the average of correlation coeficients to the improvement of the font features and review of the setest data using font features showed the lowest value. lected fonts should be considered for future work because low correlation coeficients were observed between the 4.4. Discussions and conclusions adjective pairs and the font features. For example, the adjective pairs including "Calm – Upset," "Strong – Week," In this paper, investigation on the efects of elements in "Violent – Gentle," and "Showy – Modest" are expected package design, fonts and colors, on the impressions of to be improved by adding features including "balance," package design is reported. Moreover, the results of the "ratios of line widths," and "area ratios of background attempt to predict impressions of package design using and font." Fonts which are not sophisticated too much
bers JP20H05957. have potential to be efective for improvement because the used font in this research are too suitable to packages.
Investigation on overfitting of the SVR model is requred to improve the accuracy of the model. By comparing the results of ANOVA for all the data and for the data of only pacakge A, the following factors should be considered:layout of design elements, reality and existence of illustrations, and complexity in package. The authors will attempt to develop a practical system including the proposed prediction system of impressions by addressing consideration of other features besides fonts and colors. [10] K. Ishibashi, K. Miyata, Proposal of font search method based on visual similarity, (in Japanese), Transactions of Japan Society of Kansei Engineering 12 (2013) 77–85. [11] Y. Iwata, M. Iwata, S. Tano, Analysis of relation among font shape, emotion and Kansei for emotional font generation, (in Japanese), Transactions of Japan Society of Kansei Engineering 3 (2003) 7– 16. [12] Y. Morohara, K. Kondo, S. Shimada, N. Sato, Automatic picking of index colors in textile pictures for designers, (in Japanese), Journal of Information
Processing 36 (1995) 329–337. [13] M. Saito, H. Ushioda, Y. Wada, Efects of package color on taste impression by appearance of plastic bottled green tea, (in Japanese), Transactions of Japan Society of Kansei Engineering 8 (2009) 361– 368. [14] M. Yoshida, M. S. Nobuaki Minato, Package design of gluten-free foods using semantic diferen[1] T. Oyama, Psychological efects of color, (in tial method, (in Japanese), Transactions of Japan Japanese), Journal of the Illuminating Engineering Society of Kansei Engineering 20 (2021) 101–110.
Institute of Japan 46 (1962) 452–458. [15] H. Shinohara, T. Ishiguro, S. Nakamura, T. Yamasaki, [2] T. Oyama, Y. Tanaka, J. Haga, Color-afection and Prediction of favorability rating on beer-can packcolor-symbolism in japanese and american students, age designs using convolution neural network and (in Japanese), the Japanese Journal of Psychology visualization by class activation mapping, in: Pro34 (1963) 109–121. ceedings of the Annual Conference of JSAI, (in [3] K. Ito, Afective efects o two-color combinations Japanese), 3N4J1002, 2019.
within the same hue, (in Japanese), Journal of Color [16] B. Xia, H. Sakamoto, X. Wang, T. Yamasaki, User Science Association of Japan 28 (2004) 3–15. preference prediction of packaging designs using [4] K. Ito, T. Oyama, Afective efects of tow-color deep learning, in: Proceedings of the Annual Concombinations between diferent hues, (in Japanese), ference of JSAI, (in Japanese), 1M3GS1302, 2020. Journal of Color Science Association of Japan 29 [17] Y. Shimizu, R. Doizaki, M. Sakamoto, A system to (2005) 291–302. estimate an impression conveyed by onomatopoeia, [5] Nippon Color and Design Research Institute (in Japanese), Transactions of the Japanese Society Inc., Image scale system, Accessed: Jan 14, for Artificial Intelligence 29 (2014) 41–52. 2023. URL: http://www.ncd-ri.co.jp/Introduction/ [18] Y. Koyanagi, Robust estimation methods and its ImageScaleSystem.html. applicatoni to data analysis, (in Japanese), Opera[6] H. Miyauchi, Stimulating five senses with colors tions Research as a Management Science Research and images, (in Japanese), Departmental Bulletin 23 (1978) 299–304.
Paper of Mukogawa Women’s University 29 (2019) 19–41. [7] M. Fukushima, T. Kitahara, S. Eshita, H. Fukuhara,
Analysis of relationships among color, word, and shape: evaluating its utility for the design process, in: Proceedings of the Annual Conference of JSAI, (in Japanese), 1J4-GS-9a-04, 2021. [8] T. Oyama, Colors, forms, motion and phonetic sounds and their afective and symbolic efects, (in Japanese), Japanese Psychological Review 54 (2011) 456–472. [9] T. Honda, N. Hirose, S. Mori, Changes of afective properties of japanese characters with their color-font combinations, in: IEICE technical report, volume 111, (in Japanese), 2011, pp. 127–132.