Image Quality Metrics, Personality Traits, and Subjective Evaluation of Indoor Environment Images

Abstract

Adaptive lighting systems can be designed to detect the spatial characteristics of the visual environment and adjust the light output to increase visual comfort and performance. Such systems would require computational metrics to estimate occupants’ visual perception of indoor environments. This paper describes an experimental study to investigate the relationship between the perceived quality of indoor environments, personality, and computational image quality metrics. Forty participants evaluated the visual preference, clarity, complexity, and colorfulness of 50 images of indoor environments. Twelve image quality metrics (maximum local variation (MLV), spatial frequency slope (α), BRISQUE, entropy (S), ITU spatial information (SI), visual complexity (R_spt), colorfulness (M), root mean square (RMS) contrast, Euler, energy (E), contour, and fractal dimension) were used to estimate participants’ subjective evaluations. While visual clarity, visual complexity, and colorfulness could be estimated using at least one metric, none of the metrics could estimate visual preference. The results indicate that perceived colorfulness is highly correlated with perceived clarity and complexity. Personality traits tested by the 10-item personality inventory (TIPI) did not impact the subjective evaluations of the indoor environmental images. Future studies will explore the impact of target and background luminance on the perceived quality of indoor images.

1. Introduction

Today, people spend 90–95% of their lives in indoor environments [1]. Although several environmental factors (light, noise, temperature, and moisture) impact occupants’ wellbeing, comfort, and performance, visual factors are arguably the most critical and widely studied aspects of human-environment interaction. Humans are visual species—most of the information absorbed by the brain is visual [2], and most people, up to 85%, are visual learners [3]. Therefore, there is a growing number of studies and metrics to quantify the perceived quality of the architectural environment to improve the design and integration of systems, such as adaptive lighting and building systems [4,5,6]. It is possible to hypothesize that adaptive building systems can detect environmental information using sensors (including but not limited to position, temperature, force, humidity, proximity, sensors, photodiodes, and even cameras), and adjust lighting, HVAC, and other mechanical systems to improve the human condition in buildings, as shown in Figure 1.

Adaptive building systems require computational metrics to analyze visual inputs to estimate occupant subjective responses in real time. Computer science provides such metrics, which can be utilized in the development of adaptive building systems. A subfield of computer science, image quality assessment (IQA), uses computational metrics to assess the quality of digital images [7,8]. However, these metrics often utilize a single dimension of the image quality (overall quality) rather than distinct properties of the human experience of the visual environment, such as visual preference, visual complexity, visual clarity, and colorfulness. Here, past studies investigating the visual properties of images are discussed, and a visual experiment was conducted to evaluate the performance of IQA metrics in estimating the quality of indoor environmental images. The results are discussed in the context of novel approaches for future adaptive lighting systems.

1.1. Visual Preference

Past studies hint to several factors affecting the visual preference or attractiveness of natural and human-made environments. For example, Orland et al. focused on the preference of natural objects in outdoor environments, especially the factors guiding the scenic attractiveness of landscapes [9]. In this study, a series of colorful images featuring properties with no tree and with three different size classes of trees (small, medium, and large) were generated by digitizing and editing 28 pictures, yielding a total of 112 (4 × 28) experimental images. The results indicated that the landscaping was most important for attractiveness, while the influence of the architectural style on attractiveness was less significant; within the “age” factor, older houses were preferred for attractiveness. Similarly, Yabiku et al. conducted a study to conceptualize four distinct domains potentially influencing landscape preference: environmental attitudes, socialization, aesthetics, and demographics [10]. Face-to-face interviews were carried out utilizing images of computer-generated landscapes that ranged from low-water xeriscapes to lush patterns with the inhabitants of 29 apartments in metropolitan Phoenix, Arizona. The landscape preference was asked to be assessed by showing four images of a typical North Desert Village (NDV) residence, each with a different landscaping regime. The images were produced by digitally modifying a photo of a real NDV home to reflect the four experimental treatments: desert, xeric, mesic, and oasis. On a 4-point scale “dislike very much”, “dislike somewhat”, “like somewhat”, and “like very much”, respondents were asked how much they preferred the landscape in each image. The findings showed that inhabitants, although finding desert landscapes to be aesthetically beautiful, chose high-water-use landscapes over dry landscapes for their yards. Stronger environmental view did not lead to a preference for xeriscapes, but it did lead to concessions on the quantity of turf grass that was favored in lush settings.

Other studies investigated the visual preference of indoor environments. O’Rourke et al. used paired photos in a screen-based survey to evaluate how the physical settings of public hospitals and clinics affected indigenous people’s views and experiences of waiting for care [11]. To test the hypothesis that cultural and socioeconomic heterogeneity in relation to location influenced perceptions of healthcare facilities, two distinct indigenous groups were chosen. They found that colonization had a substantial impact on personal perceptions, which was unaffected by location or furniture arrangement and included cultural and societal components including racism, guilt, fear, and the need for seclusion. They also underlined the need of using a cross-cultural perspective to enhance design interventions for waiting room spatial and symbolic treatments. Others investigated factors beyond architectural or interior design arrangements, such as lighting. Newsham et al. presented a series of high-quality, computer-rendered images of a typical open-plan partitioned office to 40 participants [12]. Participants were shown an initial batch of 12 photographs and were asked to assess each one on a scale of 1 to 10 for attractiveness. The results indicated that an appealing image of an office area was the one that was bright but not too luminous (e.g., washed out), with some non-uniformity, but not too much. Although these studies found a preference trend for indoor environments, such as offices and healthcare centers, no attempt was made to estimate subjective evaluations using metrics.

1.2. Visual Complexity

Visual complexity is characterized as the level of detail or intricacy contained within an image [13]. Mathematical methods based on algorithmic information theory or Kolmogorov complexity theory can be used to examine visual complexity [14]. Visual complexity has been recognized as a convincing predictor of preference for artistic works. According to previous studies, there is a significant relationship between visual complexity and aesthetic preference [15,16,17]. Others found that perceptions of visual complexity are influenced by three types of complexity: the number and diversity of features, the structure of elements, and asymmetry [18].

Kaya and Erkip investigated the effect of height of the floor on residents’ spatial perceptions and their feelings of being cramped in a dormitory facility [19]. The dependent factors in this study were room size perception, feeling crowded, and contentment with the room, whereas the independent variables, were floor height and residents’ sex. Participants’ comments on room size, contentment with the room, degree of privacy in the room, and other factors were evaluated. The results illustrated that contentment with a dormitory room increased as the space appeared larger and the impression of solitude in the room grows; participants who lived on the top floor perceived their rooms as being more spacious and less crowded than residents who lived on the bottom floor, as expected. These studies might have been more comprehensive if the authors had transferred visual scenes from three dimensions to two-dimension images to test if the visual perception can be estimated by image quality metrics.

Another study was conducted with 60 undergraduate psychology students ranking a collection of nine images of the sky based on their inclinations and how intricate they estimate the skyline to be [20]. The results revealed that the level of silhouette complexity had the greatest impact on preference, arousal, and pleasure. The perception of complexity, visual preference, arousal, and enjoyment were all higher when a silhouette’s complexity was higher. Durmus conducted a visual experiment utilizing 16 abstract and natural images to evaluate the effectiveness of image quality measures and spatial frequency information [21]. A 5-point Likert-type scale was used by 44 volunteers without a domain-relevant background to assess the visual complexity of images shown on a tablet. The results showed that the root-mean-square error (RMSE) and R_spt (a pseudo-random benchmark designed to counter the possible misuse of image-quality analysis methods) were statistically significantly correlated to subjective visual complexity judgments. Perceptions of spatial complexity were unaffected by biological sex.

1.3. Visual Clarity

Visual clarity can be described as the effectiveness of transmitting visual information. Visual clarity keeps people focused and reduces cognitive strain stress. Several studies investigated the effects of visual clarity and color perception. For example, Hashimoto and Nayatani assessed the correlation between visual clarity and lightness perception [22]. In their experiment, participants examined color samples under two neutral-colored side-by-side booths. The left booth was the reference, which was illuminated by a cool white fluorescent lamp with the illuminance (E_R) kept at 2000 lx. The right (test) booth was illuminated by two kinds of fluorescent lamps with variable illuminance E_T adjusted by the observers. The left eye of the observer looked at the left booth, while the right eye of the observer looked at the right booth, simulating haploscopic vision. Six male participants with normal color vision took part in the trial. They were instructed to adjust the test booth’s illuminance E_T until there was an equal perception of brightness between the two booths. Each observer’s assessments were repeated three times on different days. The average estimations of six participants revealed that by switching from the reference to the test lamp, high saturated red, green, and blue samples were found to have E_R/E_T values higher than unity. In contrast, yellow samples had E_R/E_T values that were significantly lower than unity, which meant that participants perceived the yellow samples’ lightness to be less than they did under the reference lamp. As a result, the perception of lightness in colored objects differed from the visual clarity of a lighting environment. Although the hypothesis of the effect of color contrast on visual clarity is novel, the limited sample size was a weakness of this study.

The perception of blur and visual clarity of artwork illuminated by projection systems were also tested [23,24]. In a visual experiment, five lighting conditions were generated for three artwork of different complexity levels by altering the focal length of the projector lens; one was the reference focal length setting; others were blurred by changing the lens’s focal length to enlarge the circle of confusion. Twenty-one participants with normal color vision and good visual acuity rated the visual clarity of low, medium, and high complexity pictures (defined by spatial perceptual information metric, a common approach for quantifying the spatial intricacy of images [25]) using a mean opinion score scale (1—bad, 2—poor, 3—fair, 4—good, 5—excellent). The findings illustrated that when the circle of confusion was expanded by 3%, the acceptability of blurred lighted pictures was substantially reduced. With increasing visual complexity of the artwork, blur perception shifted as well.

1.4. Colorfulness

The perceived intensity of a color (being more or less chromatic) is called colorfulness [26]. Colorfulness (sometimes colloquially referred as saturation or chroma despite technical differences) is often easier to assess and quantify since it is less subjective compared to other assessments, such as visual preference. Palus investigated the relationships between the calculated colorfulness of an image and its perceptual attributes (hue, lightness, and saturation) [27]. Palus calculated the colorfulness of five images (chart, duck, landscape, characters, mountains), with a resolution of 640 × 480 pixels. Based on the original five images (with the saturation, hue and lightness levels of 0), the author changed the saturation of each image into −0.6, −0.4, 0.2, 0.2, 0.4, 0.6; shifted the hue from −200, −150, −100, −50, to 50, 100, 150, 200 in degrees; lightness from −50%, −40%, −30%, −20%, −10%, to 10%, 20%, 30%, 40%, 50% and computed their colorfulness, respectively, to investigate the correlation of the change of saturation, hue shift, and change of lightness to colorfulness, respectively. The results indicated the perceived colorfulness of the image was a function of the average calculated saturation and lightness of the image. Similarly, Gu and colleagues introduced a new contrast-altered image database to investigate the relationship between contrast and image quality [28]. The database included 400 contrast-altered versions of 15 natural pictures. Twenty-two observers without a domain-relevant background, who participated in the subjective viewing tests, were instructed to rate their overall impression on a continuous quality scale from 1 to 5. The test images were presented in randomized order for each observer. The results showed that the quality of “blue” contrast-altered images is higher than that of “red” natural images; contrast-alteration can raise the quality of natural images.

The perception of colorfulness was also studied in the context of the design of architectural spaces. In a research study, 54 male and 46 female junior high school students (aged 12 to 14) compared and rated colorful images of hallways and lobbies for units with adult-oriented decorating and child-oriented decoration [29]. In the photographic comparison task, the view of a hospital in four distinct color prints were applied to create two comparison tasks, two hallways, and two lobbies. Participants were instructed to list three things they liked and three things they disliked concerning each group of images. Additionally, there was space provided for participants to provide their own thoughts. The results indicated that adolescents favored the vivid hues associated with childhood but disapproved of its elements, such as teddy bears and balloons. Despite the vast literature on the effect of color properties of objects and scenes on subjective evaluations, there are not too many metrics to estimate scene colorfulness.

1.5. Personality

Some studies suggested that personality may influence subjective perceptions of the visual environment. For example, Rosenbluh et al. contrasted the romantic/classic aspect of art with the neuroticism component of personality [30]. Fifty college students between the ages of 19 and 24 selected one artwork from each of 11 pairs that had been assessed for romanticism by art experts, and they also filled out the Maudsley Personality Inventory, which measures neuroticism. The results showed that subjects with a preference for romantic art tended to score higher in neuroticism than those with a preference for classical paintings. There was a negative correlation between extraversion and preference for romantic paintings. Others studied the correlation between personality and landscape preferences and found that people who were rated highly for their “sense of responsibility” inclined to reject hostile, defoliated, or wintery landscapes, despite the fact that they were more legible; people who were rated less favorably for their “emotional stability” preferred landscapes with structural rhythms and periodic patterns [31].

On the other hand, other studies indicated that there was not a strong correlation between personality and subjective preference for images. Furnham and Avison investigated the correlation between personality and preference for surreal paintings [32]. Twenty slides of surreal and realistic artworks were assessed by 62 participants. The questionnaires were then completed evaluating the tolerance of ambiguity, sensation, and the qualities of the “Big Five” survey. The results showed that the tolerance to ambiguity was not as strongly associated with a preference for surreal art as researcher had predicted.

The conflicting results on the effect of personality on subjective evaluations of art are intriguing, but there is a lack of research studying the relationship between subjective evaluations of indoor architectural environments and personality. Additionally, past studies did not utilize computational models to estimate them. The present study aims to build on previous research examining the feasibility of image quality metrics to predict the subjective perception of images.

The modeling of data that is relevant to smart building systems is necessary for managing a smart building. To access the visual perception in this process, image quality metrics can be obtained and analyzed. Thus, we hypothesize that the visual preference, complexity, clarity, and colorfulness can be estimated using image quality metrics. The relationship between subjective evaluations themselves (e.g., visual preference vs. visual clarity) is also tested. In addition, a 10-item personality inventory (TIPI) [33] was used to assess if visual preference is linked with the personality dimension of optimism.

2. Materials and Methods

2.1. Visual Experiment

A visual experiment was designed and conducted to test IQA metrics, and Pennsylvania State University institutional review board (IRB) approved the study protocol (STUDY00017319). Fifty images were selected from an online database [34] representing 10 types of architectural spaces: offices, residential spaces, educational spaces, indoor sports areas, retail stores, industrial spaces, restaurants, healthcare facilities, museums, and houses of worship, as shown in Figure 2. Images were selected to represent a wide range of perceived visual complexity (simple vs complex) and colorfulness (achromatic vs colorful). The image qualities were calculated a priori to ensure a range of values (very high and very low) were provided for each tested IQA metric.

The images were scaled to the same height (709 pixels) and displayed in a randomized order on a calibrated Display ++ LCD monitor by the Cambridge Research Systems, Kent, UK. The display was set up on a desk in front of a neutral-colored wall. The participants sat in a chair that was 0.635 m away from the screen, resulting in a 49-degree horizontal field of view and 30-degree vertical field of view. Neutral white fluorescent light sources were used to illuminate the experiment location. A calibrated Konica Minolta CL-500A illuminance spectrophotometer (Konica Minolta Sensing Americas, Ramsey, NJ, USA) was used to measure illuminance and spectra. The average horizontal illuminance on the desk at 0.7 m from the floor was 285 lx ± 4 lx, while the average vertical illuminance was 82 lx ± 2 lx at the participants’ eye level (1.2 m from the floor). The color rendering index (CRI) R_a was 85 and R₉ was 12. The light source’s correlated color temperature was 3869 K ± 3 K, with a Duv of 0.0077 ± 0.0001. A Konica Minolta LS-100 luminance meter (Konica Minolta Sensing Americas, NJ, USA) was used to measure display brightness. The gray background brightness was 45 cd/m² ± 4 cd/m², the whitest point was 90 cd/m² ± 2 cd/m², and the darkest point was 1.7 cd/m² ± 0.9 cd/m².

Forty participants without a domain-relevant background (17 male and 23 females, an average age of 27) took part in this study conducted in the Penn State University Lighting Lab. The G*Power software [35] was used to determine the sample size needed for the experiment for an effect size of r = 0.40 (α = 0.05, 1 − β = 0.80). Participants had 20/25 vision or better as tested by the Snellen chart, good contrast sensitivity (1.8/1.95 or better) tested by Pelli-Robson test, and normal color vision which was tested by Ishihara isochromatic plates.

The experimental procedure was explained to the participants before the experiment began. The instructions were read from a script to reduce research bias. Participants were asked to judge the quality of 50 images in terms of visual preference, visual clarity, visual complexity, and colorfulness by pressing a number on a number pad. A definition of each concept was provided before the experiment. Their reaction time for each question was also recorded. An example of the interface with a 9-point scale and progress bar on the left top corner is shown in Figure 3. The visual preference scale ranged from “Extremely unpleasant” to “Extremely pleasant”. For the next three questions (visual complexity, visual clarity, and colorfulness), ranges of “Extremely simple” to “Extremely complex”, “Extremely unclear” to “Extremely clear”, and “Not colorful” to “Extremely colorful” were provided for each judgment. A separate “I do not know/I do not care” option was included to reduce bias and enhance the reliability of the scales [36].

2.2. Objective Image Quality Metrics

Digital images are commonly used to identify and test the intrinsic working principles of the human visual system in vision science and environmental psychology research. These principles and experimental methods can be applied to architectural sciences research to test hypotheses and evaluate the quality of architectural spaces through image quality characteristics. In the past, various computational image quality metrics have been developed to estimate observers’ perceptual quality ratings. Subjective and objective quality assessment methods are the two primary kinds of IQA methodologies [37]. Subjective IQA is the most reliable means of evaluating image quality in psychophysical studies while being time-consuming and costly. On the other side, objective IQA aims to develop computational models capable of predicting subjective image quality [38]. In this paper, we carried out a study to predict subjective perceptual quality evaluations through objective IQA.

Twelve image quality metrics were used to estimate four subjective appraisals of the visual quality of environments (visual preference, visual complexity, visual clarity, and colorfulness). Maximum local variation (MLV) is a metric for image sharpness, the overall clarity of an image in terms of focus and contrast [39]. It is characterized by the borders between various tones or color zones. If an image is sharp, it seems simple and clear, with precise borders and excellent detail. Images with low sharpness can appear blurry and lacking in detail. MLV can evaluate resolution and acutance, two essential characteristics that contribute to perceived sharpness. Higher scores correspond to high clarity of detail in an image.

Entropy (S) is a measure of visual information content that is interpreted as the average uncertainty of the information source [40]. The higher the entropy values, the greater the level of uncertainty in images.

A blind/referenceless image spatial quality evaluator (BRISQUE) evaluates the image quality without comparing it to a reference (i.e., undistorted version of the same image) [41]. It is a spatial domain no-reference image quality evaluation algorithm. A higher level of perceptual quality is indicated by a lower score of BRISQUE. It estimates possible losses of “naturalness” in the image due to the existence of distortions using scene statistics of locally normalized brightness coefficients, resulting in a holistic quality computation. Its low processing complexity makes it excellent for real-time applications.

The International Telecommunication Union (ITU) spatial information (SI) is derived by using standard deviation across the pixels in each Sobel-filtered frame and is dependent on the Sobel filter [25]. The SI is a metric used to quantify the spatial detail in an image. Spatially complicated images have a higher SI value.

The number of cycles inside each degree of the visual field is quantified by spatial frequency, which is a measure of the periodic grating [42]. Images with fine details have a high spatial frequency, while images with coarse details have a low spatial frequency. For natural and simple images, spatial frequency information is commonly utilized to measure contrast sensitivity and visual performance. Spatial frequency illustrates the basic activity level of images, as well as the level of clarity. As a result, images with higher spatial frequency would have more clarity [43]. The frequency distribution of an image is indicated by the amplitude spectrum. On log-log axes, the amplitude fall-off follows the formula $1/ f{ }^{\alpha }$. A steeper slope (α) indicates that there is a higher proportion of low-frequency (large scale) amplitude in an image [44]. According to a study by Ogawa and Motoyoshi, perceptions of unpleasantness drop as the spatial-frequency bandwidth increases, and natural settings exhibit a linear spatial frequency slope [45].

The detectability suprathreshold (R_spt) uses an adaptive thresholding approach to quantify the visual complexity of images. The visual complexity of digital images can be estimated using this suprathreshold detectability image complexity metric [21].

There is also a colorfulness metric quantifying the saturation of images. Hasler and Suesstrunk proposed the colorfulness metric (M) by integrating the mean and standard deviation of red-green and blue-yellow color components to quantify the overall colorfulness of natural images [46]. Researchers asked 20 non-expert observers to judge the colorfulness of 84 images on a 7-point scale ranging from “not colorful” to “very colorful”. The colorfulness metric M was developed using the data from the visual experiment that had a 95.3% correlation with subjective ratings. The results indicate that the value M = 0 (not colorful), 15 (somewhat colorful), and 33 (moderately colorful).

Root mean square (RMS) contrast is based on the standard deviation of brightness levels in the stimulus, whereas Michelson contrast is based on the peak luminance values of the input. Research was carried out to determine the characteristics of different contrast measures: Michelson contrast, RMS contrast, and energy of various spatial stimuli [47]. The results showed that RMS contrast was superior contrast metric for stimuli with complicated or aperiodic brightness distributions. Although Michelson contrast is straightforward and easy to use, it is not very useful for complex stimuli.

The Euler number (also known as the Euler characteristic) is equal to the sum of all the objects in the image minus all their holes [48]. It is a critical property of images that remains unaffected by a variety of image transformations, including translations, rotations, scaling, projections, and even non-linear deformations. It has traditionally been utilized in a wide range of applications, including signature verification [49] and malaria parasite detection in blood pictures [50].

The signal (gray level) is an energy unit, since when light falls onto a sensor with watts per square meter units, a pixel has a square meter area, and the illumination is captured for a certain number of seconds. The sum of all the gray levels of an image can define the energy of an image. The rate of change in the color, brightness, or magnitude of the pixels over local areas is the energy of images.

Image contouring is the technique of detecting the structural outlines of objects in an image, which may then be used to determine the form of an object. Many image analysis applications, including image segmentation, object identification, and classification, rely on image contour detection [51].

The fractal dimension is a ratio used in fractal geometry that provides a statistical measurement of complexity by comparing how the detail in a fractal pattern varies with scale [52]. Fractal dimension is a crucial element in fractal geometry that has a wide range of applications, including image processing. It indicates the information about the geometric structure, working as a high-level image processing approach for detecting picture characteristics including texture, roughness, smoothness, area, and solidity [53]. The fractal dimension metric explains greater variation in assessments of perceived attractiveness than visual complexity measurements alone, especially for abstract and natural images [13].

The data from the visual experiment were analyzed to determine the relationship between 12 image quality metrics for 50 images and participants’ responses to visual preference, complexity, clarity, and colorfulness for each image. The quality of each image was quantified using computational image quality metrics: MLV, spatial frequency slope (α), BRISQUE, entropy (S), SI, visual complexity (R_spt), colorfulness (M), RMS contrast, Euler, energy (E), contour, and fractal dimension. The image quality metrics summary is given in

Table 1. Summary of the image quality metrics for 50 images used in the experiment.

	Avg.	Max.	Min.	Std. Dev.	95% CI
MLV	0.130	0.190	0.070	0.020	(0.124, 0.136)
BRISQUE	26	50	14	7	(24,28)
Rspt	228	627	96	97	(201,255)
SI	68	109	40	13	(64, 72)
Spa. freq. slope	−1.050	−0.850	−1.340	0.130	(−1.086, −1.014)
Entropy (S)	8.00	7.93	6.48	0.30	(7.92, 8.08)
Colorfulness (M)	34	75	7	18	(29, 39)
RMS contrast	212	299	117	48	(199, 225)
Euler	−13	550	−1233	308	(−98, 72)
Energy (E)	282,159	568,800	48,241	122,408	(248,230, 316,088)
Contour	538	1354	48	291	(457, 619)
Fractal dimension	1.920	1.920	1.912	0.001	(1.9197, 1.9203)

.

2.3. Personality Test

The 10-item personality inventory (TIPI) [31] consists of 10 elements, each with a pair of descriptors ranging from 1 (strongly disagree) to 7 (strongly agree). As presented in

Table 2. Ten-item personality inventory.

Extraversion	Agreeableness	Conscientiousness	Emotional Stability	Openness to Experience
1. Extraverted, enthusiastic	7. Sympathetic, warm	3. Dependable, Self-disciplined	9. Calm, emotionally stable	5. Open to new experiences, complex
6. Reserved, quiet	2. Critical, quarrelsome	8. Disorganized, careless	4. Anxious, easily upset	10. Conventional, uncreative

, each of the “Big Five” personality dimensions (E—Extraversion, A—Agreeableness, C—Conscientiousness, ES—Emotional Stability, and O—Openness) was represented by two items, one representing the positive pole of the dimension and the other representing the negative pole.

Items 2, 4, 6, 8, and 10 from the TIPI scale are reverse-scored items. Items 1 and 6 contribute together to extraversion which evaluates social engagement and energy levels. The agreeableness trait (items 7 and 2) illustrates personal variances in a common desire for societal harmony. Conscientiousness (items 3 and 8) is an inclination to exercise self-control. It is related to impulse control, regulation, and direction. Emotional stability (items 9 and 4) is the inclination to feel unpleasant emotions including wrath, anxiety, or despair. Openness (items 5 and 10) is a broad appreciation of aesthetics, sensation, adventure, uncommon conceptions, imagination, curiosity, and a range of experiences. To analyze the TIPI data, the reverse-scored items were recorded firstly, and then the average values of the standard item and the recoded reverse-scored item were assessed. The internal consistency of the standard item and the recoded reverse-scored item was also tested using Cronbach’s alpha, where an alpha value of 0.7 or higher value was considered to be acceptable in terms of internal consistency.

3. Results

3.1. Correlation between Image Quality Metrics and Subjective Evaluations

Table 3. Spearman’s correlation coefficient (r_s) and p-values (p) between image quality metrics and subjective evaluations.

	Preference	Complexity	Clarity	Colorfulness
MLV	rs = −0.030 p = 0.839	rs = 0.363 p = 0.010	rs = 0.222 p = 0.121	rs = 0.262 p = 0.066
BRISQUE	rs = −0.060 p = 0.679	rs = −0.318 p = 0.024	rs = −0.075 p = 0.603	rs = −0.152 p = 0.293
Rspt	rs = −0.116 p = 0.424	rs = 0.153 p = 0.288	rs = −0.191 p = 0.184	rs = 0.096 p = 0.509
SI	rs = −0.080 p = 0.579	rs = 0.388 p = 0.005	rs = 0.337 p = 0.017	rs = 0.288 p = 0.042
Spa. freq. slope	rs = −0.067 p = 0.642	rs = 0.217 p = 0.129	rs = −0.168 p = 0.245	rs = 0.190 p = 0.186
Entropy (S)	rs = 0.144 p = 0.318	rs = 0.249 p = 0.081	rs = 0.304 p = 0.032	rs = 0.371 p = 0.008
Colorfulness (M)	rs = 0.083 p = 0.564	rs = 0.268 p = 0.060	rs = −0.034 p = 0.815	rs = 0.727 p < 0.001
RMS contrast	rs = 0.040 p = 0.784	rs = 0.044 p = 0.764	rs = 0.088 p = 0.543	rs = 0.163 p = 0.258
Euler	rs = −0.117 p = 0.420	rs = 0.295 p = 0.038	rs = −0.149 p = 0.300	rs = 0.256 p = 0.073
Energy (E)	rs = −0.002 p = 0.990	rs = −0.201 p = 0.163	rs = −0.007 p = 0.964	rs = −0.356 p = 0.011
Contour	rs = 0.072 p = 0.620	rs = 0.419 p = 0.002	rs = 0.191 p = 0.184	rs = 0.389 p = 0.005
Fractal dimension	rs = 0.172 p = 0.233	rs = −0.275 p = 0.053	rs = −0.103 p = 0.176	rs = −0.356 p = 0.011

shows the Spearman’s correlation coefficient (r_s) and p-values (p) between image quality metrics and subjective evaluations. To determine whether to reject the null hypothesis (H₀: there is no relationship between image quality metrics and subjective evaluations), p-values (p) were utilized in hypothesis testing. The statistical significance threshold was considered p ≤ 0.05. The Spearman’s correlation (r_s) determines if two variables have monotonic associations (whether linear or not).

The results indicated that subjective evaluations of complexity, clarity, and colorfulness could be estimated with an image quality metric, but not preference. The strongest predictor of perceptual colorfulness was metric colorfulness (M) (r_s = 0.727, p < 0.001). Perceptual colorfulness was statistically negatively correlated with energy (E) (r_s = −0.356, p = 0.011) and fractal dimension (r_s = −0.356, p = 0.011). Perceptual clarity had a low positive correlated with SI (r_s = 0.337, p = 0.017), and Entropy (S) (r_s = 0.304, p = 0.032). Visual complexity was positively correlated with Contour (r_s = 0.419, p = 0.002), SI (r_s = 0.388, p = 0.005), MLV (r_s = 0.363, p = 0.010), and Euler (r_s = 0.295, p = 0.038); while it was negatively correlated with BRISQUE (r_s = −0.318, p = 0.024).

3.2. Correlation between Subjective Evaluations

Table 4. Mann-Whitney U test p-values and effect size r between subjective evaluations.

	Complexity	Clarity	Colorfulness
Preference	p = 0.463 r = 0.073	p < 0.001 r = 0.433	p = 0.001 r = 0.324
Complexity		p < 0.001 r = 0.344	p < 0.001 r = 0.353
Clarity			p < 0.001 r = 0.613

shows the Mann-Whitney U test p-values and effect size r between subjective evaluations themselves. The results revealed that the subjective evaluation of clarity was highly correlated with colorfulness (p < 0.001, effect size r = 0.613); perceived clarity also had a correlation with preference (p < 0.001, effect size r = 0.433), and complexity (p < 0.001, effect size r = 0.344). Furthermore, colorfulness was statistically significantly correlated with complexity (p < 0.001) with a larger effect size (r = 0.613).

3.3. Correlation between Personality and Subjective Evaluations

Table 5. Internal consistency of standard items and the recoded reverse-scored items.

	Extraversion	Agreeableness	Conscientiousness	Emotional Stability	Openness to Experience
Standard item	1. Extraverted, enthusiastic	7. Sympathetic, warm	3. Dependable, self-disciplined	9. Calm, emotionally stable	5. Open to new experiences, complex
Recoded reverse-scored item	6. Reserved, quiet	2. Critical, quarrelsome	8. Disorganized, careless	4. Anxious, easily upset	10. Conventional, uncreative
Cronbach’s alpha	0.91	0.48	0.68	0.7	0.42

summarizes the Cronbach’s alpha values for each personality trait. The internal consistency of standard items and the recoded reverse-scored items illustrated that the standard item and recoded reverse-scored item of the personality test was not always consistent. According to the Cronbach test, extraversion and emotional stability had internal consistency.

The Spearman’s correlation coefficient (r_s) and p-values (p) between five personality and subjective evaluations are shown in

Table 6. Spearman’s correlation coefficient (r_s) and p-values (p) between five personality and subjective evaluations.

	Extraversion	Agreeableness	Conscientiousness	Emotional Stability	Openness to Experience
Preference	rs = 0.111 p = 0.496	rs = −0.181 p = 0.265	rs = −0.093 p = 0.570	rs = −0.043 p = 0.793	rs = 0.167 p = 0.304
Complexity	rs = 0.084 p = 0.608	rs = −0.065 p = 0.690	rs = 0.112 p = 0.492	rs = −0.120 p = 0.463	rs = 0.188 p = 0.188
Clarity	rs = −0.004 p = 0.979	rs = −0.184 p = 0.255	rs = −0.082 p = 0.616	rs = −0.255 p = 0.112	rs = 0.146 p = 0.370
Colorfulness	rs = 0.031 p = 0.850	rs = −0.134 p = 0.415	rs = −0.041 p = 0.805	rs = −0.092 p = 0.571	rs = 0.122 p = 0.452

. The investigation of the association between personality and subjective evaluations illustrated that there was no significant correlation in any of the perception evaluations or personality.

4. Discussion

4.1. Subjective Evaluations and Image Quality Metrics

The results from the visual experiment indicate that the subjective evaluations (with the exception of preference) can be estimated using computational metrics. Unsurprisingly, predicting the visual preference of scenes were challenging, likely due to the personal, cultural, and other interobserver differences. In addition, there were correlations between subjective evaluations themselves. For example, the perceived colorfulness was highly correlated with perceived clarity and complexity. Considering the complexity of visual and cognitive pathways, this result is not surprising. Past studies highlighted the existence of some of these relationships already, such as the colorfulness-preference [54,55], and color contrast-clarity [22].

The significance of this study lies in evaluating the potential of using image quality metrics to estimate subjective evaluations of indoor environments. While some of the metrics performed well, others did not estimate any subjective evaluations. For example, the perceptual colorfulness could be effectively estimated by the colorfulness metric (M), while the second relatively high correlation of the metrics and subjective evaluations occurred in visual complexity and SI. Some image quality metrics were contextually useful—perceptual colorfulness was easily predicted through image quality metrics since it was the most objective evaluation; while other image quality metrics were not very statistically significantly correlated with perceptual preference, clarity, or complexity. This was distinctly different from previous studies indicating that the image quality metrics can be applied to estimate subjective perception [6,56,57], the justification behind the conflicting finding could be the state-of-the-art no-reference IQA methods were less efficient in dealing with the artifacts, including incorrect high-frequency details introduced by super-resolution algorithms [58]. Besides, it is not possible to give every pixel in an image equal attention.

Regarding predicting perceptual complexity, the tested images did not include humans or abstract objects (e.g., Durmus’ visual complexity study [21] included lines, people, etc.), which is a potential limitation of the experiment. For visual clarity, computational metrics were developed for display-related problems (e.g., blur), but the tested images in this experiment were not modified to create display-related issues. Besides, none of the metrics performed well in predicting perceived preference, probably due to the cognitive complexity, sample size, cultural and personal difference between participants. Tourancheau and his colleagues tested the behaviors of four objective image quality metrics on three image databases and demonstrated that the performances of the quality metrics could fluctuate significantly depending on the database used for testing [59]. Consequently, it might be critical to estimate the subjective evaluations of the environment through image quality metrics during the design and development of adaptive building systems.

These findings will hopefully encourage researchers to further develop metrics for estimating the subjective evaluations of indoor environments. The computational metrics have the potential to be deployed in adaptive building lighting systems where the central intelligence can process virtual information collected from the environment through machine learning algorithms to make real-time decisions to increase occupant comfort, performance, and well-being.

4.2. Personality Traits and Subjective Evaluations

To analyze the personality traits of the participants, the reliability of the gathered data of the standard item and recoded reverse-scored item of the personality test for consistency was investigated. Nunnally recommends a Cronbach’s alpha value greater than 0.7 for internal consistency of a test [60]. For each of the five-personality traits, a Cronbach’s alpha test was used to check the reliability of the individual TIPI items. Analyses of the personality traits indicated that the standard item and recoded reverse-scored item of the personality test was not always consistent.

In addition, the subjective evaluations of the images were analyzed and correlated to the big five personality traits of the participants. The results illustrated that there was no correlation between the participants’ personality traits and their subjective evaluations of images. This finding paralleled other investigators’ work [61] who have found there was no discernible relationship between individuals’ personality characteristics and lighting preferences. Since the sample size was moderate (n = 30), the data were not normally distributed and each variable had several levels, therefore, no consequential conclusions could be obtained. More psychophysical research with different stimuli and large sample size can provide more insight into the effect of personality on subjective assessments of the visual environment.

4.3. Limitations

The presented study was deliberately limited to indoor environments to test the feasibility of estimating subjective evaluations of architectural spaces. Although the sample size (n = 40) was reasonable for an image assessment study, a larger sample size may be needed to assess the effect of personality traits on subjective evaluations of images.

Another limitation is the 2D nature of the visual assessments. In an experiment, Rogowitz and Rushmeier compared the perceived quality of animated 3D objects with their matching projections of 2D still images [62]. The results revealed that the 2D measurements of image quality were insufficient to accurately reflect the perceived quality of 3D objects. In accordance with the direction of the light, the perceptible quality of 2D images of 3D objects may change noticeably. In short, it is necessary to test the performance of image quality metrics to estimate the subjective evaluations of immersive environments, which might be affected by factors such as target or background luminance.

5. Conclusions

Adaptive building lighting systems can be designed to assess the quality of the built environment and adjust lighting to increase occupant comfort, wellbeing, and performance. Such advanced systems require incorporating computational image assessment metrics to estimate human response to visual stimuli. Although there are several existing metrics that can predict overall image quality assessment, a granular approach might provide more insight into estimating different aspects of perception, such as visual preference, visual clarity, visual complexity, and perceived colorfulness.

To test the performance of existing image quality metrics in estimating the perception of images, a visual experiment was conducted where 40 participants with normal color vision, acuity, and contrast evaluated 50 images of indoor spaces. Results indicate that perceived colorfulness was highly correlated with perceived clarity and complexity. Several computational image quality metrics highly correlated with the visual clarity, complexity, and colorfulness of indoor images. Although none of the metrics could predict visual preference, the results were overall very encouraging for the development of adaptive building systems. The use of metrics to predict the visual assessment of the environment can help design better buildings systems that can save energy, allocate resources, and maximize occupant comfort and well-being.

A secondary hypothesis was the effect of personality on subjective evaluations of the visual environment, especially between agreeableness and preference. Past studies on art perception and personality bore mixed results. In this study, no correlation was found between personality and subjective judgments of the images of indoor environments.

It is reasonable to assume that adaptive lighting systems can be designed to capture images of the built environment to estimate the overall quality using a camera [63,64,65]. Although it is not strictly a limitation for this study, technology used to capture images and the image processing might have an impact on the performance of such adaptive building systems. However, a preliminary study on the impact of image capturing devices on IQA metric estimation showed only small differences between a webcam, iPhone, and HDR camera, under two difference lighting conditions [66]. The constantly improving quality of image capturing devices, and the potential of optimizing the sensor and aperture control capabilities [67,68], can yield better results in the future. Another potential topic of contention is the use of AI for face recognition and privacy concerns [69]. Advanced techniques, such as face anti-spoofing [70], visual markers [71], and a facial embedding algorithm [72], can address some of these privacy concerns.

The results from the study and recent technological developments highlight the need for more research to characterize the variations in IQA in response to the illumination and image capturing device settings and modes. Future studies will investigate the effect of lighting levels on perceptual preference, clarity, complexity, and colorfulness in relation to lighting levels. The results in this experiment will be extended to immersive environments to assess the difference in subjective evaluations of immersive environments and images shown in displays.