Adding texture to color: quantitative analysis of color emotions

What happens to color emotion responses when texture is added to color samples? To quantify this we performed an experiment in which subjects ordered samples (displayed on a computer monitor) along four scales: Warm-Cool, MasculineFeminine, Hard-Soft and Heavy-Light. Three sample types were used: uniform color, grayscale textures and color textures. Ten subjects arranged 315 samples (105 per sample type) along each of the four scales. After one week, they repeated the full experiment. The effect of adding texture to color samples is that color remains dominant for the WarmCool, Heavy-Light and Masculine-Feminine scale (in order of descending dominance), the importance of texture increases in that same order. The Hard-Soft scale is fully dominated by texture. The average intra-observer variability (between the first and second measurement) was 0.73, 0.66 and 0.65 for the uniform color, grayscale texture and color texture samples, respectively. The average inter-observer variability (between an observer and the other observers) was 0.68, 0.77 and 0.65, respectively. Using some 25,000 observer responses, we derived analytical functions for each sample type and emotion scale (except for the Warm-Cool scale on grayscale textures). These functions predict the group-averaged scale responses from the samples’ color and texture parameters. For uniform color samples, the accuracy of our functions is significantly higher (average adjusted R = 0.88) than that of functions previously reported. For color texture, the average adjusted R=0.80. Introduction There is growing interest in the understanding of human affective feelings in response to seeing colors. The so called ‘color emotions’ (i.e. emotional responses to color) involved in published studies do usually not refer to basic human emotions like happiness, surprise or fear. Rather, they capture the response on an associated affective dimension specified by the investigators and may therefore be anything. Color emotion studies recently published (e.g. [1]-[3]) have focused on the selection of emotional scales and how they interrelate (by means of factor analysis). Regression analysis shows the relationships of these scales with the underlying color appearance attributes (lightness, chroma and hue). Additionally the question whether color emotions can be regarded as culture specific or universal has been studied [1], [3]-[6]. Roughly summarizing the published studies on color emotion, the common finding is that the color emotions are reasonably well described by a small number of semantic factors, like for instance colour weight, colour activity and colour heat in [1]. Of the perceptual attributes, lightness and chroma are most frequently reported as being the relevant parameters for quantitative prediction of the color emotions, although hue cannot be ignored in scales like warm-cool. So far, the role of texture in color emotion has received only little attention. An early study by Tinker [7] showed that surface texture, as represented by coated paper or cloth, had little or no effect upon apparent warmth or affective value of colors. Kim et al. [8] used color and texture features to predict human emotions based on textile images. Erhart & Irtel [9] indicate that surface structure can change the emotional effect of colored textile samples, depending upon the color. More recently, Simmons & Russell [10] reported that the addition of texture can significantly change the perceived unpleasantness of colors, depending on the texture class. This paper reports on our study into the effect of texture on color emotion. Our systematic experiments build upon-, but differ in a number of ways from previous studies. Most importantly, instead of only studying uniformly colored samples, we also use samples with grayscale textures and color textures. These textures were synthesized to prevent biased responses such as reported in [10]. Second, we introduce a method in which all samples (shown on a computer display) maintain visible during experimental trials so that they can be ordered conveniently along an emotion scale. Third, our subjects performed the full experiment twice, with at least one week in between the first and second measurement. This allows quantification of the intra-observer variability over time, on which we are the first to report. We believe that repeatability information is at least as important as the information obtained from more observers. Finally, we sampled the available color gamut of our display in a very systematic manner to optimally cover both lightness, chroma and hue. We analyze our data in terms of rank correlations within subjects and between subjects and provide quantitative descriptions. We derive color and texture emotion formulae that predict the group-averaged responses on the emotion scales from the samples’ color and texture descriptors. Methods One of the problems we encountered in a pilot experiment is that when samples are shown one after the other, subjects tend to forget what responses they gave on the emotion scales for similar samples shown earlier in the trial. This leads to an unnecessary increase in variability in the subjects’ responses, and therefore lower intraand inter-observer correlations. To overcome this we designed our experiment in such a way that all samples maintain visible during a trial. We asked our subjects to order 105 square samples horizontally along an emotion scale labeled with opposite word pairs, using the computer mouse to drag samples from their initial location on the top of the screen. Samples could be dragged to any position on the screen to allow the subjects to keep an overview of the arrangement of the samples. Subjects knew that only the CGIV 2010 Final Program and Proceedings 5 horizontal position of the samples on the scale would be analyzed. Four emotion scales (Warm-Cool, Masculine-Feminine, Hard-Soft and Heavy-Light), in four different trials, were used. There were three conditions differing only in the type of samples. In the Uniform Color (UC) condition, uniformly colored samples were used that were systematically selected from the sRGB color gamut. In the Grayscale Texture (GT) condition the samples had a texture created in luminance, but not in the chromatic domain. Textures were generated using Perlin noise [11]. The third condition (Color Texture, CT) added a single color to the grayscale textures. Examples of results obtained for these three conditions are shown in Figure 1.