Defining the Visual Complexity of Learning Management Systems Using Image Metrics and Subjective Ratings

Research has demonstrated that students’ learning outcomes and motivation to learn are influenced by the visual design of learning technologies (e.g., learning management systems or LMS). One aspect of LMS design that has not been thoroughly investigated is visual complexity. In two experiments, postsecondary students rated the visual complexity of images of LMS after exposure durations of 50– 500 ms. Perceptions of complexity were positively correlated across timed conditions and working memory capacity was associated with complexity ratings. Low-level image metrics were also found to predict perceptions of the LMS complexity. Results demonstrate the importance of the visual design of learning technologies and suggest that additional research on the impact of LMS visual complexity on learning outcomes is warranted. Defining the Visual Complexity of LMS Image Metrics and Subjective Ratings Des mesures d'images de bas niveau ont également été trouvées comme étant indicateurs des perceptions de la complexité des SGA. Les résultats démontrent l'importance de la conception visuelle des technologies d'apprentissage et suggèrent que des recherches supplémentaires sur l'impact de la complexité visuelle des SGA sur les résultats d'apprentissage sont justifiées.


Introduction
Postsecondary institutions in Canada (and around the world) continue to invest in online learning technologies and online course development (Bates et al., 2017) and, in doing so, increase students' access to education and reduce their education-related costs (Coates, James, & Baldwin, 2005). Test scores, student engagement, and perceptions of learning (Nguyen, 2015) provide evidence that delivering courses online may be as effective as traditional face-to-face delivery of content and learning activities (Ladyshewsky, 2004;Means, Toyama, Murphy, & Baki, 2013;Tallent-Runnels et al., 2006). These effects, however, may be attenuated by a number of factors (Lim, Ayesh, Stacey, & Chee, 2013), including the poor visual design of learning management systems (LMS), which can disorient learners and overload their cognitive resources (Eveland Jr. & Dunwoody, 2001). An LMS is a web-based software application that is used to organize and administer content and track and report progress for learning and training purposes (Ellis, 2009;Mabila, Gelderblom, & Ssemugabi, 2014). Describing the appearance of LMS accurately and determining how visual design influences student learning is important for making evidence-informed design modifications. Describing the visual complexity of LMS in efficient, meaningful, and reliable ways is a foundational step in the broader investigation of the impact of LMS design on student learning outcomes. Therefore, we explored whether low-level image metrics predicted judgements of the visual complexity of LMS and whether these judgements were influenced by viewer characteristics. We also sought to confirm the stability of first impressions of the visual complexity.

Visual Complexity of Learning Technologies
Early webpage design consisted of arranging static blocks of text with few graphics on a page (Berners-Lee, n.d.), as visual design was largely constrained by the state of hardware and software technology at the time. The persistent belief that functionality and efficiency were paramount to advancing networking technologies may also have slowed the development of more sophisticated and attractive webpage designs (Moshagen & Thielsch, 2010). Researchers soon realized, however, that visually pleasing designs could enhance the perceived usability, credibility, user satisfaction, and trustworthiness of the information presented on consumer-focused websites (Cyr, Head, & Larios, 2010;Cyr, Head, Larios, & Pan, 2009), which could ultimately impact behaviour (e.g., purchase intentions; Brunner-Sperdin, Scholl-Grissemann, & Stokburger-Sauer, 2014).
One aspect of visual aesthetics that strongly influences visual attention and behaviour is complexity (Wang, Yang, Liu, Cao, & Ma, 2014). Low-level image metrics, such as feature congestion (Corchs, Ciocca, Bricolo, & Gasparini, 2016;Madan, Bayer, Gamer, Lonsdorf, & Sommer, 2018), the numbers of text and image areas (Reinecke et al., 2013), and the numbers of words, links, fonts, and graphics (Geissler, Zinkhan, & Watson, 2006;Ivory, Sinha, & Hearst, 2001;Michailidou, Harper, & Bechhofer, 2008) have been used to predict viewers' perceptions of visual complexity of consumerbased webpages. Past research on website complexity has drawn on Berlyne's (1960) theory of stimulus complexity. Berlyne suggested that a stimulus becomes increasingly complex when the number of distinct elements increases, and the number of elements seen as a unit decreases. The theory goes on to predict that there is an optimal level of complexity-that too little or too much complexity decreases communication effectiveness and influences consumer attitudes and intensions negatively (Berlyne, 1960; see also Geissler et al., 2006).
Research also shows that reliable judgements about various aspects of webpages, including complexity, can be made after extremely brief exposure durations (e.g., 50 ms; Lindgaard, Fernandes, Dudek, & Brown, 2006;Tractinsky, Cokhavi, Kirschenbaum, & Sharfi, 2006), which is particularly problematic when first impressions are poor as these impressions are robust even in the face of evidence to the contrary (Lindgaard et al., 2006). Moreover, the strength of the relationships between objective metrics and subjective ratings may be influenced by several individual difference factors, including working memory capacity (Madan et al., 2018;Walhout et al., 2015) and familiarity with the stimuli presented (Forsythe, Mulhern, & Sawey, 2008;Gartus & Leder, 2017;McDougall, Curry, & de Bruijn, 1999) or the specific tasks assigned (Nadkarni & Gupta, 2007). Indeed, viewers describe simple webpages as placing few decision-making demands on them, whereas complex webpages require switches in one's focus of attention that tax cognitive resources (Harper, Michailidou, & Stevens, 2009). Further, consumer-focused websites with a high degree of complexity, but that lack structure and order, typically result in significant levels of frustration and wasted time (Lazar, Bessiere, Ceaparu, Robinson, & Shneiderman, 2003;Tuch, Bargas-Avila, Opwis, & Wilhelm, 2009). Greater complexity combined with a highly structured or ordered spatial arrangement increases viewer preferences for web interfaces (Deng & Poole, 2010).
Design criteria for learning technologies has largely been based on ease of use rather than the evaluation of aesthetics (Liu, 2003); however, evidence is accumulating to demonstrate that visually pleasing learning technologies enhance learning (Pomales-García, Liu, & Mendez, 2005) and motivation to learn (Zain, Tey, & Goh, 2011). For example, ratings of perceived attractiveness of an e-learning module have been found to be positively correlated with the number of content relevant words recalled (Pomales-García et al., 2005). A limitation of this work is that it was unclear whether the aesthetics of the content or the web interface that the modules were displayed in were being evaluated. Nevertheless, these results suggest that judgements about visual aesthetics should not be ignored and are important for understanding how technology can be further developed to support student learning.

The Present Study
We describe two experiments conducted to replicate, extend, and supplement the findings of previous research (viz., Lindgaard et al., 2006;Reinecke et al., 2013;Tractinsky et al., 2006), and to provide a foundation to explore the impact of visual design of LMS on student behaviour and learning outcomes in future research. Although efforts have been made to define the complexity of consumerfocused websites, it is unclear whether the findings of previous research apply to LMS user interfaces. Fundamental differences might be expected as the purpose and use of LMS and consumer-based websites are quite different. Thus, our goals for the present experiments were to define visual complexity of LMS by examining whether low-level image metrics predicted viewer's subjective complexity ratings; determine whether individual difference factors (e.g., age, education level, LMS familiarity, working memory capacity) influence complexity judgements; and verify that subjective complexity ratings correlate across various exposure durations, suggesting that first impressions of LMS are stable.

Method
Participants. University undergraduate students (N = 29; 17 women; M age = 20.8 years, SD = 3.8, Range: 18-36) with normal or corrected-to-normal vision and hearing, and fluent speakers of English were recruited from the University of Manitoba, Winnipeg, Canada, to participate in the present study. Twenty-three participants were right-handed. The sample consisted of individuals from a broad range of geographic and educational backgrounds. Participants received a $20 gift card following informed consent. The study procedures were approved by the university's Research Ethics Board.

Materials and procedures.
The study was conducted in individual sessions in a soundattenuated, well-lit, windowless laboratory in a single 1.5 hr session. Each session began with a detailed explanation of the procedures before written consent was obtained. Participants then completed general information, and computer proficiency (CPQ) questionnaires, and two subtests of the Wechsler Adult Intelligence Scale-Fourth Edition (WAIS-IV; Wechsler, 2008) in one of four test orders, such that the four experimental tasks they were asked to complete alternated with these questionnaires and WAIS-IV subtests.
General information questionnaire. An 11-item questionnaire was administered to collect information about age, gender, and geographic and educational background.

CPQ.
The CPQ is a 33-item paper-pencil, self-report questionnaire designed to assess the level of computer proficiency in adults in six domains: Computer Basics, Printing, Communication, Internet, Scheduling, and Multimedia (Boot et al., 2015). Participants rated their perceived ability to perform a specific task on a 5-point scale (1 = Never tried to 5 = Very easily). We created an LMS domain to measure proficiency in using LMS. Domain scores were computed by averaging the responses to items in each domain. The six original domain scores were summed to produce the Total CPQ scores. Reliability of Total CPQ scores and domain scores was reported as high for the original questionnaire (Cronbach's α = .86-.98) (Boot et al., 2015).

Digit Span and Coding.
Scores generated from the administration of the Digit Span and Coding subtests of the WAIS-IV (an individually administered measure of cognitive ability for individuals aged 16-90 years; Wechsler, 2008) were used to estimate working memory capacity, and psychomotor and visual motor speed, respectively. These constructs are known to be related to performance on computerbased tasks (Ownby & Czaja, 2008). The Digit Span subtest has three parts: Forward (individual tries to repeat digits forward); Backward (individual tries to repeat digits backward); and Sequencing (individual tries to repeat digits in ascending order). In the Coding subtest, individuals use a key to copy symbols that are paired with numbers within time limits. The normative mean scaled score for each subtest is 10 (SD = 1).

Experimental tasks.
The stimulus set consisted of 400 screenshots of various LMS pages from trial versions and existing courses of the top 26 LMS used in post-secondary education in North America. LMS pages (above-the-fold or content that fills the screen) were captured at a resolution of 1920 x 1024 pixels in 24-bit true colour. Screenshots included course content to ensure that the images were representative of the visual display that learners would typically view. Objective measurements of complexity (see Table 1) were extracted from each of the 400 images using existing algorithms; the source code was obtained online at https://homes.cs.washington.edu/~reinecke/publications.html. These metrics were previously defined by Reinecke et al. (2013), Hasler and Suesstruck (2003), and Zheng, Chakraborty, Lin, and Rauschenberger (2009). The values for number of leaves classified as text area and vertical and horizontal balance were zero for our sample of images and could not be entered in our analyses. Predicted complexity values were calculated using the remaining image metrics (Reinecke et al., 2013) and were used to classify the 400 screenshots into three equal-sized categories for Experiment 2: low (lower 33% of values), medium, and high complexity images (upper 33% of values) (Gelman & Park, 2008) (Figure 1). We also noted JPEG (Joint Photographic Experts Group) file size. Table 1 Definitions of the image metrics extracted from the images of learning management systems (LMS)

Image metrics Definition
Colour Hue Average pixel value of the purity of colours. Values range from 0 to 180.

Saturation
Average pixel value of the intensity of colours. Values range from 0 to 255.

Value
Average pixel value of the relative lightness or darkness of colours. Values range from 0 to 255.

Colourfulness
Weighted sum of the trigonometric length of the standard deviation in a*b* [green (-) to red (+)/blue (-) to yellow (+)] space and the distance of the centre to the neutral axis. Values can begin at 0 (not colourful) and have no upper limit, but values of 109+ have been interpreted as extremely colourful a .

Image metrics Definition
Space-based decomposition

Number of leaves
Recursive division of an image into N evenly spaced regions (leaves) until a region has no visible space divider or is too small. This is a continuous variable beginning at 0.

Number of image areas
Area identified as a separate image using an algorithm. Adjacent images may count as one image. This is a continuous variable beginning at 0.

Number of text groups
Number of horizontal groups of text characters. This is a continuous variable beginning at 0.
Text area and non-text area Number of regions classified based on whether they contain text or non-text information.

Number of quadtree leaves
Recursive division of an image into subparts based on the amount or type of information until divisions can not longer be made. This is a continuous variable beginning at 0.

Symmetry
Degree to which the arrangement of the leaves on the flips side of the horizontal and vertical axes are similarly arranged. Values range from 0 (less symmetrical) to 1 (more symmetrical).

Balance
Degree to which the top and bottom, and right and left parts of an image have similar optical weights (e.g., a similar number of leaves). Values range from 0 (less balanced) to 1 (more balanced).

Equilibrium
Degree to which the quadtree leaves center around the midpoint of an image. Values range from 0 (less centering) to 1 (more centering).
Note. Adapted from "Predicting users' first impressions of website aesthetics with a quantification of perceived visual complexity and colorfulness," by Reinecke et al., 2013  Stimuli were presented on a 24-in. widescreen LCD monitor with a 2000-MHz Pentium II computer, using Experiment Builder software (SR Research, ON, Canada). For two of the four experimental tasks, participants viewed each of the 400 screenshots for 500 ms. This was followed by a screen asking them to rate (with a keypress) the visual complexity (or colourfulness as part of a separate investigation) by means of Likert-type scales (1 = not at all to 7 = extremely) (Reinecke et al., 2013;Tuch et al., 2009). Participants rated the screenshots in two separate blocks and took short breaks after every 50 trials. Ten practice trials were administered prior to the start of each task so that participants could become familiar with the task and could ask clarifying questions.
The procedures for the other two experimental tasks were similar to those described above except that participants viewed a subset of screenshots (n = 60) that varied in levels of complexity (simple, neutral, complex) while gaze behaviours were sampled (500 Hz, spatial resolution = 0.5°) using the EyeLink 1000 Plus eye-tracking system (SR Research, ON, Canada). An eye-tracker was employed to collect indices used to make inferences about visual attention (Holmqvist et al., 2015). Each screenshot remained on the screen until the end of the participant's first fixation, and then participants rated its complexity (or colourfulness as part of a separate investigation) by means of Likert-type scales. A break followed every 20 trials. Prior to administration of the two eye-tracking experiments, a 9-point calibration/validation procedure was performed. A chin rest was utilized to limit head movement and drift correction was performed on each trial to manage eye gaze drift due to fatigue and changes in head position.

Results
Mean ratings of visual complexity from participants' ratings were computed for each of the 400 screenshots. Examination of various regression diagnostics revealed that six images had undue influence on the model in the preliminary stages of our analyses and were removed prior to running subsequent analyses. Next, we entered the image metrics (predictors from Table 1, with the exception of number of leaves classified as text area and vertical and horizontal balance for reasons described above) and mean subjective complexity ratings for each screenshot (outcome) into two separate multiple linear regressions with backward elimination to find the most parsimonious model. The metrics that did not contribute significantly to the prediction of the subjective ratings were removed in an iterative fashion. The final complexity model indicated that six image metrics (see Table 2) explained 46.6% of the variance of the subjective complexity ratings [R 2 = .47, F(6, 387) = 58.13, p < .001].  (2003) Next, we calculated the complexity of each screenshot using the unstandardized coefficients from our final regression model and from a previous model (Reinecke et al., 2013), and examined their relationships with subjective complexity ratings (see Table 3). The correlation coefficients were significantly different from each other (Z = 2.09, p = .04, two-tailed), perhaps reflecting fundamental differences in the purpose and use of LMS and consumer-focused webpages. As an additional test for operationalizing visual complexity, we ran correlations between subjective complexity ratings and JPEG file size and found a significant association between these two variables (p < .001), which may suggest that JPEG file size provides a relatively good estimate of LMS complexity without the need to collect subjective complexity ratings or using image metrics and may be a more practical proxy for complexity, particularly for LMS developers. Table 3 Pearson product-moment correlation coefficients between participants' subjective ratings of complexity, JPEG file size, and predicted complexity ratings Negative correlations between digit span scaled scores and complexity ratings for each complexity level (low, medium, high; r(29) = -.43, -.41, -.40, p < .05, for all three correlations, respectively) were evident. Other participant characteristics (i.e., age, years of education, CPQ Total and the LMS domain scores, coding subtests scaled scores) were not associated with subjective complexity ratings.
The mean subjective complexity ratings from the 400 trial (500 ms exposure duration) and the 60 trial (exposure duration limited to the end of the first fixation) conditions were highly correlated, r(58) = .755, p < .001, suggesting reliable complexity ratings regardless of exposure duration. The effect of complexity (low, medium, high) on mean fixation durations was significant, F(2, 57) = .10, p = .91. On average, participants fixated the images for 213 ms (SD = 63).

Experiment 2
In Experiment 1, participants perceived the LMS screenshots as similarly complex whether they viewed the images for 500 ms or to the end of their first fixation (M = 213 ms, SD = 63). To verify that first impressions of complexity are similar across various exposure durations, we presented a subset of images from Experiment 1 to a new group of participants in Experiment 2 and asked them to rate complexity after various presentation times while their gaze behaviours were recorded.

Method
Participants. Undergraduate students at the University of Manitoba, Winnipeg, Canada (N = 16; 9 women) with normal or corrected-to-normal acuity, colour vision, and hearing participated in Experiment 2. Participants' mean age was 22.1 years (SD = 3.2 years; Range: 18-30 years) and they had an average of 3.2 years of post-secondary education (SD = 2.4; Range = 1-8 years). The sample represented individuals from a wide range of geographical and educational backgrounds and all participants spoke English fluently. Participants received a $20 gift card following written informed consent. Experiment 2 was approved by the university's Research Ethics Board.

Materials and procedure.
Participants completed the general information questionnaire, the CPQ, and the Digit Span and Coding subtests counterbalanced in four task orders. After each task, one block of the experiment was administered. As in Experiment 1, participants were seated at a desk and viewed the experimental stimuli on a high-precision 24-in. widescreen LCD monitor with a Pentium II computer, using Experiment Builder software (SR Research, ON, Canada). Prior to the eye-tracking experiments, a 9-point calibration/validation was performed. Drift correction was performed on every trial to ensure that data was adjusted for movement.
Sixty screenshots, classified as low (n = 20), medium (n = 20), and high (n = 20) complexity were selected from the sample of 400 screenshots in Experiment 1 (see Figure 2). Each participant completed three blocks of 10 trials (10 screenshots from each level of complexity). During a given block, the screenshot remained on the screen for 50 ms, 150 ms, 500 ms, or until the end of the participant's first fixation (M = 235 ms, SD = 98). Participants then rated the complexity of the screenshot by means of Likert-type scales. Experimental tasks and blocks were counterbalanced. Each screenshot was viewed by at least three participants and mean ratings of complexity were calculated for each screenshot.

Results
First, we examined the mean fixation durations across Exposure Duration conditions to verify that participants attended longer to the stimuli that were presented for longer durations. The results of a 3 (Complexity: low, medium, high) x 4 (Exposure Duration: 50 ms, 150 ms, 500 ms, End of first fixation) ANOVA, with repeated measures on the last factor, indicated a main effect of Exposure Duration, F(1.3, 72.9 Next, mean subjective complexity ratings were submitted to a 3 (Complexity: low, medium, high) x 4 (Exposure Duration: 50 ms, 150 ms, 500 ms, end of first fixation) ANOVA, with repeated measures on the last factor. A main effect of Complexity was evident, F(2, 56) = 4.08, p = .02, η p 2 = .13.
Follow-up tests (LSD) confirmed that the mean ratings were significantly lower for low than for high complexity images (p = .006). There was no main effect of Exposure Duration and no interaction between the two variables. Together, these findings suggest stability in viewers' perceptions of complexity across exposure durations.

General Discussion
In the present research, we found that image metrics could be used to predict students' ratings of complexity of LMS user interfaces, and that this model was distinct from models reported in previous studies (Reinecke et al., 2013). For example, our model was unique in that value (a colour metric) and horizontal symmetry (a quadtree decomposition metric) were significant (negative) predictors of complexity (see Figure 3). Symmetry (or asymmetry), in addition to the number of elements, has been argued to be an important dimension of complexity (Nadal, Munar, Marty, & Cela-Conde, 2010). The human visual system is uniquely tuned to detecting symmetry in the environment (Wagemans, 1997).
When the environment appears more symmetrical, viewers perceive it as less complex (Attneave, 1957) and more aesthetically pleasing (Jacobsen & Höfel, 2003) thereby eliciting positive affect (Pecchinenda, Bertamini, Makin, & Ruta, 2014). Although some contend that vertical symmetry is easier to detect than horizontal symmetry, thus is more important for judgements about attractiveness (Huang et al., 2018), overall mirror symmetry (average symmetry across four axes) may be more predictive of several types of viewer judgements (citation/citations). For example, decreased mirror symmetry along with increased GIF (Graphics Interchange Format) image compression (RMSGIF) predict increased perception of complexity for a set of black and white abstract patterns (Gartus & Leder, 2017). Our findings from the present study suggest that the degree of horizontal symmetry above the fold must be used carefully and intentionally when designing LMS user interfaces; however, symmetry may not be important when determining whether the interfaces appear professional or unprofessional (Zheng, Chakraborty, Lin, & Rauschenberger, 2009). Slight decreases in hue and value and increases in colourfulness predicted higher ratings of LMS complexity. Our finding that value was a significant predictor of subjective complexity is a new finding but should not be surprising. Value is defined as the relative lightness or darkness of a colour, and contrasting values help to define specific aspects of a stimulus. When the overall value of a stimulus is low, boundaries between aspects of the stimulus are not easily seen by the viewer and may contribute to increased perceptions of complexity. Webpage colour, in general, strongly influence viewers' satisfaction and trustworthiness ratings, which may in turn influence their intention to purchase (or eloyalty). A blue colour scheme is universally linked to trust, whereas yellow is perceived as aversive causing feelings of hostility and distrust leading to greater or diminished e-loyalty, respectively (Cyr et al., 2010). Some research suggests that LMS interfaces with a black (or other dark colour fonts) on white backgrounds reduces viewers' processing time and increase their memory for words, and that overall colour schemes of blue, green, pink, and orange are preferred (Tharangie, Irfan, Marasinghe, & Yamada, 2008). Further study on the effects of colour, hue, and value in LMS is warranted.
Consistent with previous findings (Lindgaard et al., 2006;Reinecke et al., 2013;Tractinsky et al., 2006), the results from our experiments demonstrated that complexity ratings were consistent across exposure durations ranging from 50 to 500 ms. Consistency in judgements about stimuli across time has been referred to as the "mere exposure effect" (Bornstein & D'Agostino, 1992;Zajonc, 1968). When first impressions of visual aesthetics are positive, a "halo effect" can occur whereby good first impressions influence perceptions of usability (Sonderegger & Sauer, 2010). The opposite is true for very poor first impressions (Lindgaard et al., 2006) that leave viewers feeling unimpressed and may lead to distrust (Cyr et al., 2009). While we did not ask our participants about whether the levels of complexity were appreciated or appealing to them, one might predict that very simple or very complex LMS pages would be unappealing but for different reasons (e.g., boring or overwhelming, respectively; Berlyne, 1960). Moreover, the consequences of poor first impressions of learning technologies may be perceptions of reduced reliability or trustworthiness of the course content, and the subsequent avoidance of use and engagement with the LMS. This would be detrimental to learning.
Lastly, we found negative correlations between Digit Span scaled scores (Wechsler, 2008) and complexity ratings, suggesting that individuals with a reduced working memory capacity perceived LMS as more complex and perhaps more difficult to process than those with a greater working memory capacity. Working memory allows one to temporarily hold and process task relevant information (Baddeley, 2012). Moreover, the ability to chunk information and having familiarity with the stimuli can increase working memory capacity (Shen, Popov, Delahay, & Reder, 2018). Thus, one might expect that experience with LMS and mean complexity ratings for screenshots of LMS would be negatively correlated, however, we did not find this to be true in our study. One possible reason for this null finding may be that the university students participating in our research had limited experience with all the LMS from which our stimuli were derived.

Limitations and Future Research
Although investigations to define the visual complexity of consumer-focused websites (Harper et al., 2009;Ivory et al., 2001;Reinecke et al., 2013) and other types of visual stimuli (Corchs et al., 2016;Goldberg, 2014;Madan et al., 2018) have been conducted, to our knowledge, we are the first to examine the visual complexity of LMS using image metrics and students' ratings of complexity. However, we must acknowledge several limitations of our research. First, although we used a large stimulus set, the number of participants in both experiments was relatively small, which reduces the generalizability of our findings. Therefore, it is important that additional research be conducted to replicate and extend our findings.
A second limitation of our research is that the images we used were screenshots of LMS that contained course content in addition to the visual design of the LMS itself. Students would rarely (if ever) view entirely blank course shells, however, we acknowledge that this decision may have added a potential confound. Additionally, if content included images of salient objects or people, students' attention may be preferentially drawn to these aspects of the LMS screenshot. Regardless of age, we have a strong tendency to attend to faces more than the background when viewing scenes presented on a computer screen (Stoesz & Jakobson, 2014). Moreover, the presence of human images on consumerbased websites has been linked to higher levels of perceived warmth and social presence and greater image appeal (Cyr et al., 2009). Unfortunately, very few of our stimuli included images of people's faces and we could not test whether this had an impact on complexity ratings. Thus, the presence of human faces is an important consideration for online teaching and learning environments and suggests that online courses include images or videos of the instructor teaching. Future research studies could involve the development of one course placed within various LMS so that content is kept constant and the impact of the visual design of specific LMS could be examined directly. Additional factors such as the presence of text-based instruction and instruction from a person in a video could also be examined.
Another limitation of our research is that we did not examine the impact of the complexity or other facets of visual aesthetics (e.g., craftsmanship; Moshagen & Thielsch, 2010) on learning outcomes. Direct relationship between visual complexity and learning outcomes are likely to be moderated or mediated by several other factors. For example, researchers have shown that feelings of frustration, anger, and boredom are problematic as negative affect can dramatically reduce task performance and impede learning in traditional face-to-face educational settings (Christianson, 1992;Sharp, Hemmings, Kay, & Sharp, 2017). A clear understanding of the relationships between LMS engagement, attitudes towards such technologies, affect when using LMS, and then how those relationships impact student learning outcomes is an important consideration for future research (Coates et al., 2005). Such investigation would likely require various pillars of assessment including self-report, physiological assessment, and test scores to measure learning of content.
Finally, we examined the concept of complexity and examined relationships to viewers' individual difference factors (e.g., working memory capacity), however, we did not ask participants to provide information about their preferences for the visual design of an LMS. Personalization by the system being used and customization by the user may improve the user experience and increase benefits to retailers, for example, by increasing buying and customer loyalty (Smith, 2017). Social media platforms and other applications also allow users to modify the user interface (e.g., white vs. black background; modern vs. standard views); these ideas could be translated to LMS. Many LMS already allow instructors to customize the student view by reducing or adding the number of tools displayed, so it may be possible to add the ability for students to customize their view of the LMS. Learners may also benefit from personalization of a learning path based on their performance on certain tasks as they proceed through an online course (Niknam & Thulasiraman, 2020). The impact of personalization of the visual aspects of the LMS itself or how this variable interacts with the learning path approach or other teaching and learning strategies on student learning requires more investigation. This research may reveal, however, that limits to customization are necessary to avoid creating an overly complex learning environment (Berlyne, 1960;Geissler et al., 2006).

Conclusions
Objective characteristics of webpages, including image metrics generated from algorithms, can be used to predict more complex psychological principles and human behaviour (Harper et al., 2009;Ivory et al., 2001), which can be incredibly useful when planning studies to examine how the visual and navigational designs of LMS affects the student experience and learning outcomes. Limited empirical work has been done in the area of the visual perception of LMS; thus, examining postsecondary students' perceptions of the visual complexity of LMS establishes the existence of visual design for LMS as a distinguishable and measurable construct in the field of education. Combining behavioural and physiological methodologies is also unique in education research and these methodologies may prove useful to further examine postsecondary students' perceptions of LMS and their overall online learning experiences, and how these relate to academic outcomes.