Modeling and Analyzing the Impact of Mosaic Art on the

Psychological Well-Being of Primary School Students

Suyan Zhang

, Xin Dong

, Yan Feng

, Li Chen

, Guanghui Huang

and Jianhua Zhang

Faculty of Humanities and Arts, Macau University of Science and Technology, Avenida Wai Long, Macao, China

Department of Engineering Science, Macau University of Science and Technology, Avenida Wai Long, Macao, China

School of Creative Media, City University of Hong Kong, Kowloon Tong, Hong Kong, China

Medical Engineering Technology and Data Mining Institute, Zhengzhou University, 100 Science Avenue, Zhengzhou, China

Keywords: Deep Learning, Data Analysis, Art Therapy, Expression Therapy, Mosaic Art, Mental Health Assessment.

Abstract: Contemporary children's mental health has emerged as a focal point of societal concern, particularly given

that children are in the foundational stages of physical and psychological development. The exploration of

children's mental health through expressive art therapies, including art therapy, has attracted widespread

attention. First, we conduct mosaic art classes for 50 students at a primary school in Henan Province, China,

collect students' works, and test the students' psychological status according to the mental health scale.

Subsequently, we categorize the artworks into positive and negative types, combine them with various deep

neural network models to classify the student works, and analyse the correlation between colors, brightness,

and student psychological states in the works, aiming to provide a theoretical model for student mental health

assessment. Finally, we found that the ResNet model achieved an accuracy of 76% in the artwork

classification task. Brightness in student works cannot be a representative factor of psychological states,

whereas the role of color (yellow, blue, green, brown) was particularly prominent. Through this study, we

conclude that color in artistic expression has potential value for the mental health of primary school students,

providing scientific evidence and theoretical support for contemporary children's mental health through

expressive therapy.

1 INTRODUCTION

As the Chinese education system progressively shifts

towards a focus on quality education, mental health

issues among elementary school students have

garnered widespread attention from various sectors of

society. Children's psychological resilience and

coping skills are relatively underdeveloped, and

prolonged exposure to unregulated emotional states

can result in adverse emotions such as aversion to

learning, low self-esteem, anxiety, depression, and

more severe psychological issues like personality

disorders, potentially harming their future academic

and personal lives (Zhang et al, 2022). Consequently,

addressing children's mental health has become an

imperative. Originating in Europe in the 1940s, art

therapy employs visual arts as a form of

psychological treatment, with expressive therapy

serving as a crucial component. Research has

demonstrated that expressive therapy effectively

alleviates specific emotional states, overcomes

communication barriers, and steers individuals

towards positive changes (Hogan, 2001).

Mosaic art, a significant form of expressive

therapy, utilizes materials like stones, shells, enamel,

ceramics, and glass to create artworks. A distinctive

feature of mosaic art is that it comprises many

independent small pieces, yet visually forms a unified

whole (Dunbabin, 1999; Hyatt, 2007). This design

concept resembles the pixel concept in computing,

facilitating the application of deep learning models in

classifying mosaic works (Kim and Jeong, 2019).

These models can distinguish not only the style of

artworks but also classify them on an emotional level

(Cetinic et al, 2019). This approach meets the need to

analyze children's artworks and their psychological

states. Currently, the analysis of artworks in

expressive therapy primarily relies on questionnaire

surveys and the clinical experience of therapists. We

aim for computer models to provide a more objective

assessment tool. In this study, we selected fifth-grade

students from a primary school in Henan Province as

108

Zhang, S., Dong, X., Feng, Y., Chen, L., Huang, G. and Zhang, J.

Modeling and Analyzing the Impact of Mosaic Art on the Psychological Well-Being of Primary School Students.

DOI: 10.5220/0012922100004536

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Data Mining, E-Learning, and Information Systems (DMEIS 2024), pages 108-112

ISBN: 978-989-758-715-3

the research subjects. We collected stress data from

students through questionnaires and combined it with

mosaic artworks created by the students themselves.

We then utilized computer models such as ResNet,

VGG, and AlexNet for data analysis to explore the

potential impact of artistic expression on the mental

health of elementary school students.

2 MATERIALS AND METHODS

2.1 Course Design

This study involved conducting a mosaic art course

for 50 fifth-grade students at a primary school in

Henan Province, China. Prior to the course, students

completed the Psychological Health Diagnostic Test

(MHT), which was developed according to the

standards set by Professor Zhou Bucheng in 2009 and

tailored specifically to the contemporary context of

Chinese children (Zhang et al, 2022). The MHT

comprises sections on learning anxiety and validity

testing. The learning anxiety section contains 15

questions, whereas the validity testing section

includes 10 questions. According to the guidance

accompanying the scale, a score below 3 on the

learning anxiety scale indicates good mental health,

whereas a score above 7 suggests anxiety.

Concerning the validity scale, which reflects the

confidence level of student scores, a score of 7 acts as

the threshold, with lower scores signifying higher

confidence.

Students are required to independently create

mosaic artworks following specific guidelines: They

must use tiles to assemble a butterfly pattern on a

circular base measuring 15.5 cm in diameter. The

tiles, which may be triangular, square, or rectangular,

should range from 1 to 3 cm in size and feature colors

selected from blue, green, red, purple, or additional

color palettes. Upon completion, the students'

artworks are photographed and archived as 970x970

pixel images.

2.2 Samples and Preprocessing

2.2.1 Experimental Samples

Initially, after the class, 47 artworks and

questionnaires are collected, and those artworks that

do not comply with the specified requirements are

excluded, resulting in 27 valid artworks.

Subsequently, Python 3.11.5 was employed, along

with the OpenCV and PIL libraries, to develop code

for image analysis, including color classification and

grayscale processing. Specifically, the edges of the

figures were removed, and the color of each pixel was

analyzed, with the resulting color proportions being

reported. Finally, the data were subjected to

inspection and color refinement by art professionals,

who categorized the colors into ten types: yellow,

blue, green, red, purple, brown, pink, black, white,

and other.

2.2.2 Database Samples

The data are derived from a public dataset based on

the International Affective Picture System (IAPS)

images from 2010, where the emotional attributes of

the samples encompass happiness, anger, awe,

satisfaction, disgust, excitement, fear, sadness,

among others. Emotions are categorized as either

positive or negative, specifically excluding images

with ambiguous emotional expressions such as "awe"

(Machajdik and Hanbury, 2010). The dataset consists

of 250 images depicting positive emotions and an

equal number depicting negative emotions, divided

into training and testing sets at a ratio of 4:1.

2.3 Model Methods

The deep learning models used in the experiment

were ResNet, VGG, and AlexNet. Since ResNet

showed the most effective performance in subsequent

experiments, we focused on ResNet50, ResNet101,

and ResNet152. These three models vary in their

number of convolutional layers, specifically 50, 101,

and 152 layers, respectively. The model is

distinguished by a network depth of more than 100

layers and features a residual module, which

enhances model accuracy by increasing depth. We

utilized the model functions provided by Torchvision

in PyTorch and employed the Cross Entropy Loss

function. We compared various approaches in

selecting the optimizer and learning rate, and

ultimately chose the SGD optimizer, with the learning

rate set at 0.005. The AlexNet model also utilizes the

model functions provided in PyTorch. The AlexNet

architecture comprises 5 convolutional layers, 3 fully

connected layers, and dropout layers (Deng et al,

2009). The loss function, optimizer, and learning rate

are the same as those used with ResNet. The VGG

model improves on AlexNet by using multiple

consecutive 3 × 3 convolutional kernels instead of

AlexNet's larger ones, thereby better preserving the

Modeling and Analyzing the Impact of Mosaic Art on the Psychological Well-Being of Primary School Students

109

image's properties (Demir, 2020). The model includes

19 hidden layers, comprising 16 convolutional layers

and 3 fully connected layers. PyTorch offers the

VGG16 function for the 16-layer convolutional

neural network. The selected loss function and

optimizer are Cross Entropy Loss and SGD,

respectively, with the learning rate set at 0.005.

Furthermore, all three models ultimately output two-

dimensional vectors to classify figures into positive

and negative categories.

In the task of evaluating the importance of colors,

we employed the Random Forest algorithm. The

model was deployed in Python, utilizing the

RandomForestClassifier function provided by scikit-

learn to classify images based on color and ascertain

the importance of various colors using features

returned by the random forest.

Figure 1: Model design. Student paintings are directly used

as input for the model, outputting two-dimensional

classification results.

Figure 2: Accuracy of 5 models. Student works are divided

based on their confidence score. The grey is the result of the

public dataset. Student works are divided based on their

confidence score. The works with high confidence are

orange, and all works are blue.

3 RESULTS

3.1 Deep Learning Fitting Results

Initially, we utilized a public dataset of

positive/negative images and applied deep learning

models to train the dataset. The training set,

consisting of 400 images, was divided into batches of

50 and trained 100 times. The predicted results for the

test set are displayed in Figure 2. The ResNet model

demonstrates strong performance on both the public

dataset and in the classification of student artworks.

Specifically, ResNet50 achieves the highest

accuracy, reaching 77% on the test set and attaining

over 70% accuracy in classifying student artworks.

Conversely, AlexNet achieves an accuracy of 68% on

the test set but does not meet the experimental

expectations in classifying student artworks.

3.2 Grayscale Analysis Results

To ascertain whether image brightness serves as an

indicator of psychological states, we converted the

images to grayscale and retrained the models.

Grayscale images reflect only brightness information

and exclude color information. The results of the

models retrained with grayscale images are displayed

in Figure 3.

Grayscale images do not adequately fit the

models, yielding a prediction accuracy on the test set

of only about 60%. For the ResNet model, the highest

accuracy achieved in classifying student artworks

was only about 54%. Moreover, when considering

confidence levels, ResNet101 and ResNet152 were

unable to distinguish negative from student samples.

Similarly, the predictive performance of the VGG

model was poor, misclassifying all student artworks

as positive samples. The AlexNet model's accuracy in

classifying student data was less than 50%.

Consequently, the parameter of image brightness may

not be relevant to this study on psychological

conditions. Further exploration is necessary to

comprehend the reasons behind these results.

3.3 Color Analysis Results of Works

After eliminating the influence of image brightness,

we explored the relationship between color and

psychological states. In section 2.2.1, we employed a

hybrid approach combining algorithmic and manual

methods to categorize colors into ten categories: red,

blue, green, yellow, purple, pink, brown, black, white,

and others. Given that ResNet152 demonstrates the

most robust predictive performance for student

artworks (considering confidence levels), we

integrated the color data analysis from the 27 student

artworks with the ResNet152 predictions. The

random forest algorithm was employed to perform

DMEIS 2024 - The International Conference on Data Mining, E-Learning, and Information Systems

110

nonlinear fitting of the data, with the importance of

each color feature displayed in Figure 4. The analysis

indicated that green, brown, yellow, and blue were the

most significant colors, collectively accounting for

50% of the overall importance.

Figure 3: Classification results of grayscale model.

Figure 4: Random forest filters important colors in images.

Table 1: The significance of color representation.

Color Positive

significance

negative

significance

Yellow

Bright and warm

(Pastoureau, 2019),

happy (Boyatzis and

Varghese, 1994)

Hunger, frustration,

anger (Gehred, 2020)

Blue

Positive, calm

(Boyatzis and

Varghese, 1994),

wise, loyal (Gehred,

2020)

Impulsiveness,

anxiety (Ran et al.,

2017, Korkmaz et al.,

2016), introverted

sadness, depression,

and cold (Güneş and

Olguntürk, 2020)

Green

Relaxation, Comfort

(Boyatzis and

Varghese, 1994,

Mammarella et al.,

2016), Nature,

Symbolizing mystery,

demons (Pastoureau,

2019), jealousy, envy,

and misfortune

Harmony (Gehred,

2020)

(Güneş and

Olguntürk, 2020)

Brown

Reliable, stable,

friendly,

comfortable, and

safe (Gehred, 2020)

Negativeness (Güneş

and Olguntürk, 2020,

Madden et al., 2000),

mourning (Gehred,

2020)

3.4 Retrospective Analysis of the

Correlation Between Color and

Psychology

Building on the analysis results from section 3.3, we

further validated the psychological implications of

yellow, blue, green, and brown through a

comprehensive retrospective literature review. The

findings of the literature review are presented in

Table 1. The analysis reveals that most colors possess

both positive and negative connotations, while warm

and cool color tones do not exhibit clear tendencies

towards positivity or negativity (Demir, 2020).

Neutral colors display distinct emotional

associations: white is typically linked with positivity,

and black with negativity. Furthermore, warm and

cool color palettes may exhibit emotional biases as

the shades deepen or lighten (Hanada, 2018). Taking

yellow as an example, although it evokes associations

with warmth and brightness, some research suggests

that it also symbolizes hunger (Eiseman, 1998).

Additionally, the color green has dual positive and

negative effects, with its negative connotations often

tied to cultural customs, such as in Japanese culture

where green is considered ominous, and in

Christianity where it signifies demons.

Studies have reported associations between colors

and depression-related psychological disorders;

specifically, patients with depression exhibit a higher

preference for blue and a reduced preference for

yellow (Ran et al., 2017). The psychological

significance of colors is dynamic; although brown is

visually associated with the earth and connotes

reliability, it can also evoke negative and mourning

emotions in various contexts (Gehred, 2020). The

retrospective analysis demonstrates that colors play a

significant role in psychology, influenced by factors

such as geographical location, upbringing

environment, and the temporal characteristics of the

subjects. Consequently, our study, which focuses on

mosaic art, confirms that the colors of mosaic tiles

can reflect the growth environments of elementary

Modeling and Analyzing the Impact of Mosaic Art on the Psychological Well-Being of Primary School Students

111

school students, with yellow, blue, green, and brown

being particularly significant.

4 CONCLUSIONS

Psychological issues among primary and secondary

school students are becoming increasingly prominent,

necessitating the development of objective tools to

assist schools and teachers in promptly identifying

and addressing these issues. Traditional

psychological surveys, characterized by their

extensive number of questions, long update cycles,

and rigid formats, are impractical as tools for daily

and frequent monitoring. By employing expressive

therapy within art therapy, psychological monitoring

can be seamlessly integrated into art classes using

computer models to analyze students' artworks. This

approach significantly reduces the professional

demands on teachers regarding psychology and can

effectively detect changes in students' psychological

states within a short period.

In this study, we conducted mosaic art classes for

fifth-grade students at a primary school in Henan

Province and utilized deep learning models to analyze

the correlation between the students' artworks and

their psychological states. The classification results

from this study indicate that the discriminative

models demonstrate robust performance and offer

significant reference value. Furthermore, we

discovered that yellow, blue, green, and brown play

pivotal roles in the classification of artwork states,

indirectly reflecting students' psychological states

through their color classifications. However, a

limitation of this study is the absence of long-term

tracking data on students. As academic pressure

increases, the nature of students' artworks may

evolve. This study has preliminarily established

significant correlations between mosaic colors and

the psychological conditions of elementary school

students. Additionally, this research contributes

scientific evidence and theoretical support for the use

of expressive therapy in contemporary children's

mental health.

REFERENCES

Boyatzi, C. J.& Varghese, R. 1994. Children's emotional

associations with colors. The Journal of genetic

psychology, 155, 77-85.

Cetinic, E., Lipic, T. & Grgic, S. 2019. A deep learning

perspective on beauty, sentiment, and remembrance of

art. IEEE Access, 7, 73694-73710.

Demir, Ü. 2020. Investigation of color ‐ emotion

associations of the university students. Color Research

& Application, 45, 871-884.

Deng, J., Dong, W., Socher, R., Li, L.J., Li, K. & Feifei, L.

2009. Imagenet: A large-scale hierarchical image

database. IEEE conference on computer vision and

pattern recognition. Ieee, 248-255.

Dunbabin, K. M. 1999. Mosaics of the Greek and Roman

world, Cambridge University Press.

Eiseman, L. 1998. Colors for your every mood: Discover

your true decorating colors, Capital Books.

Gehred, A. P. 2020. Canva. Journal of the medical library

association: JMLA, 108, 338.

Gunes, E. & Olgunturk, N. 2020. Color ‐ emotion

associations in interiors. Color Research &

Application, 45, 129-141.

Hanada, M. 2018. Correspondence analysis of color–

emotion associations. Color Research & Application,

43, 224-237.

Hogan, S. 2001. Healing arts: The history of art therapy,

Jessica Kingsley Publishers.

Hyatt, E. 2007. Art of the earth: Ancient arts for a green

future, AuthorHouse.

Kim, H. & Jeong, Y.-S. 2019. Sentiment classification

using convolutional neural networks. Applied Sciences,

9, 2347.

Korkmanz, S., Ozer, Ö., Kaya, Ş., Kazgan, A. & Atmaca,

M. 2016. The correlation between color choices and

impulsivity, anxiety and depression. European Journal

of General Medicine, 13.

Machajdik, J. & Hanbury, A. 2010. Affective image

classification using features inspired by psychology and

art theory. Proceedings of the 18th ACM

international conference on Multimedia. 83-92.

Madden, T. J., Hewett, K. & Roth, M. S. 2000. Managing

images in different cultures: A cross-national study of

color meanings and preferences. Journal of

international marketing, 8, 90-107.

Mammarella, N., Domenico, A., Palumbo, R. & Fairfield,

B. 2016. When green is positive and red is negative:

Aging and the influence of color on emotional

memories. Psychology and aging, 31, 914.

Pastoureau, M. 2019. Yellow: The History of a color,

Princeton University Press.

Ran, M., Tan, X., Jiang, G., Chen, X., Yang, H. & Yuan, G.

2017. Probe into Color Preference in Inpatients with

Depressive Disorders. Journal of Psychiatry and Brain

Science, 2.

Zhang, X., Dai, Z., Antwi, C. O. & Ren, J. 2022. A cross-

temporal meta-analysis of changes in left-behind

children’s mental health in China. Children, 9, 464.

DMEIS 2024 - The International Conference on Data Mining, E-Learning, and Information Systems

112