Abstract: - It is considered that children’s normal growth depends on their ability to use their fine motor skills.

Deficits in fine motor skills in preschool children can interfere with even basic daily activities. Research also

links these problems to future challenges. Therefore, early identification of preschool children’s fine motoric

abilities is considered essential. However, the assessment of the development of fine motor skills is considered to

be a rather complex process. Complex and time-consuming methods are used for their reliable assessment, which

also requires the presence of educational experts. The aim of this study is to investigate whether it is possible to

create a simple and useful tool for assessing fine motor skills in preschool children, based on convolutional neural

networks. For this purpose, a comparative study between 5 state-of-the-art CNN architectures is carried out, to

investigate their accuracy in assessing fine motor skills. Drawings of Greek students from public kindergartens

were used to train the investigated CNN models. The Griffiths II and the Eye Coordination Scale were used

to assess the developmental age of preschool children. The findings demonstrate that, although challenging,

automatic and precise detection of fine motor skills is feasible if a larger dataset is used to train deep learning

models.

Key-Words: - Convolutional Neural Networks, Deep Learning, Fine Motor Skills, Preschool, Griffiths Test,

Assessing Development

Received: June 15, 2022. Revised: February 16, 2023. Accepted: March 14, 2023. Published: April 12, 2023.

1 Introduction

Motor development refers to the change and enrich-

ment of motor behaviour throughout life. The devel-

opment of motor skills in children can be divided into

two main categories: gross motor skills and fine mo-

tor skills. Gross motor skills refer to skills and move-

ments that involve whole-body movements and large

muscle groups to perform, while fine motor skills re-

fer to precise movements that use smaller muscles,

such as writing, drawing, tying shoelaces, etc. Dur-

ing the first few years of life, children grow rapidly

and develop both gross and fine motor skills. Sat-

isfactory development of fine motor skills is cru-

cial for children’s well-being. Several studies have

shown that early difficulties in gross and fine mo-

tor skills have significant effects on academic per-

formance and psychosocial maladjustment, [1]. They

are also associated with reading development in pri-

mary school, [2], numeracy, [3], arithmetic, [4], men-

tal imagery, [5], and working memory impairments,

[6]. Various factors can prevent children from de-

veloping their fine motor skills to the full. One ex-

ample is media use, which may have a negative im-

pact on the development of fine motor skills in early

childhood, [7]. In addition, obesity could also lead

to lower fine motor precision performance, [8], while

more generally the impact of modern society has also

been studied as a factor that could potentially af-

fect fine motor development, [9]. However, parental

support for pre-school children, even if it is mini-

mal, was considered to be really important to miti-

State-of-the-art CNN architectures for assessing Fine Motor Skills: A

comparative study.

1KONSTANTINOS STRIKAS, 2NIKOLAOS PAPAIOANNOU, 2IOANNIS STAMATOPOULOS,

2ATHANASIOS ANGEIOPLASTIS, 2ALKIVIADIS TSIMPIRIS, 2DIMITRIOS VARSAMIS,

1PARASKEVI GIAGAZOGLOU

1Department of Physical Education and Sport Sciences

Aristotle University of Thessaloniki

Serres, GREECE

2Department of Computer, Informatics and Telecommunications Engineering

International Hellenic University

Serres, GREECE

WSEAS TRANSACTIONS on ADVANCES in ENGINEERING EDUCATION

DOI: 10.37394/232010.2023.20.7

Konstantinos Strikas, Nikolaos Papaioannou,

Ioannis Stamatopoulos, Athanasios Angeioplastis,

Alkiviadis Tsimpiris, Dimitrios Varsamis,

Paraskevi Giagazoglou

E-ISSN: 2224-3410

Volume 20, 2023

gate the consequences of fine motor impairments and

to help children to reach their maximum potential in

the development of fine motor skills, [10]. The same

seems to be true for physical activity, which is con-

sidered to be beneficial for children, especially when

provided regularly in a formal setting, [11], as well

as for touch typing interventions, [12]. Therefore,

the assessment of preschool children’s fine motor

skills is considered essential, considering the above-

mentioned effects on children’s development. The

Griffiths Scales No II is considered to be one of the

most reliable developmental screening tests, [13]. It

consists of six subscales to better assess motor devel-

opment. In the bibliography, other reliable screen-

ing developmental tests such as Movement-ABC 2,

[14], DSNR, [15], and pegboard tasks, [16], are also

widely used. Recently, a new method using deep

machine learning techniques, specifically convolu-

tional neural networks (CNN), has been proposed for

the assessment of fine motor skills in preschool chil-

dren, called FineMotorSkillsCNN, [17]. Convolu-

tional neural networks (CNN) are important tools for

image recognition and classification tasks in various

fields, such as smart agriculture, [18], medicine, [19],

and self-driving cars, [20], probably because of their

powerful image processing capability. Due to the

stack of convolutional layers, CNNs can automati-

cally extract features. Depending on the depth of the

layers, CNNs could extract from low-level features,

such as edges and dark spots, to high-level features,

such as an object. Therefore, due to the success and

high performance of CNNs in complex image pro-

cessing problems, various methods have been pro-

posed. In this paper, an evaluation of the performance

of different state-of-the-art CNN architectures such

as Efficient-Net, [21], ResNet, [22], VGG16, [23],

and MobileNet, [24], for the assessment of fine mo-

tor skills of preschool children is presented, in order

to investigate whether it could become a simple and

useful tool for teachers and parents to detect possi-

ble impairments early and to help children reach their

full potential in the development of fine motor skills.

The remainder of the paper is organized as follows:

Section 2, presents the basic elements of the theory

and the methodology used, while Sections 3 and 4,

present the results of the comparative study and the

conclusions respectively.

2 Methods and materials

2.1 Dataset

The original dataset used in this study consists of 1601

images representing pictures of a man or a woman

drawn by 442 children from 20 different preschool

units. The dataset was divided into training and test

sets. The training set, which was used to train the dif-

ferent CNN models, consists of 1121 images, while

480 images were used as the test set. As in [17] the

images were divided into six different classes accord-

ing to their developmental age, where class 0to class

5correspond to low and high developmental ages, re-

spectively ( Table 1).

Table 1: Griffiths II test scores and classes according

to Developmental age (DA).

DA Class

32-47 0

48-53 1

54-61 2

62-67 3

68-73 4

74-150 5

The classification of the pictures was carried out

by three educational experts on the basis of the Grif-

fiths II Scale D, which consists of six items for each

year, such as threading beads onto a lace, building a

tower of cubes, cutting with scissors, copying simple

geometric shapes and drawing a house and a person

freely. The developmental age (in months) was cal-

culated by multiplying the number of items passed by

two. The testing process started with simple tasks cor-

responding to a younger age and was stopped after six

consecutive failures in six different skills. The devel-

opmental age of each child was then determined. Fig-

ure 1, shows pictures randomly selected from each of

the six different developmental age classes.

(a) Class 0 (b) Class 1 (c) Class 2

(d) Class 3 (e) Class 4 (f) Class 5

Figure 1: Sample painting-class pairs from dataset.

WSEAS TRANSACTIONS on ADVANCES in ENGINEERING EDUCATION

DOI: 10.37394/232010.2023.20.7

Konstantinos Strikas, Nikolaos Papaioannou,

Ioannis Stamatopoulos, Athanasios Angeioplastis,

Alkiviadis Tsimpiris, Dimitrios Varsamis,

Paraskevi Giagazoglou

E-ISSN: 2224-3410

Volume 20, 2023

2.2 State-of-the-art CNN architectures

The CNN architectures employed, are described in

this chapter. The models studied were trained using a

transfer learning approach.

2.2.1 VGG16

Attempting to improve on the original Convolutional

Network Architecture proposed by Krizhevsky, [25],

VGG addresses an important aspect of Convolutional

Network Architecture, which is depth. VGG achieves

high performance in both localisation and classifi-

cation tasks by increasing the depth of the network

(16 layers) and using small 3×3 convolutional kernels

with convolutional stride 1and 2×2 pooling layers

with step size 2. Thus, the number of parameters in-

creases significantly, while the decision function be-

comes more discriminative.

2.2.2 ResNet

ResNet primarily addresses the degradation problem.

It has been found that as the depth of the networks

increases, the accuracy decreases rapidly. To deal

with this problem, ResNet introduced a deep resid-

ual learning framework that skips connections from

initial filters to later layers. Identity mapping is per-

formed on the shortcut connections. Based on the idea

that if the model becomes deeper by adding layers

constructed as identity mappings, the performance of

the model shouldn’t be worse than its shallower coun-

terpart, ResNet prompts the network to approximate

the residual function. Due to the small sample size of

our dataset, ResNet50 is used in this study.

2.2.3 MobileNet

MobileNet is designed for mobile and embedded

vision applications due to its low computational

cost without compromising accuracy. To build

lightweight deep neural networks, MobileNet differs

from traditional approaches regarding convolution ar-

chitecture. MobileNet is based on depth-separable

convolutions, which split the process of both filtering

and combining inputs into a new set of outputs into

two distinct processes, a filtering layer and a combin-

ing layer. This is in contrast to standard convolution,

which performs the entire process in one step. As a

result, the parameters are drastically reduced and the

model becomes less computationally intensive com-

pared to a network of the same depth using regular

convolutions.

2.2.4 EfficientNet

EfficientNet’s baseline network is generated by a

multi-objective neural architecture search, similar to

MnasNet, [26]. Focusing on model scaling, it was

found that performance can be improved by carefully

balancing network depth, width and resolution. To

achieve this, a compound scaling method was pro-

posed, that scales network depth, width and resolu-

tion uniformly, using a set of predetermined constant

scaling coefficients. Specifically:

d=αϕ

w=βϕ

r=γϕ

α∗β2∗γ2≈2

and

α≥1, β ≥1, γ ≥1

where d, w, r correspond to depth, width and res-

olution respectively, α, β, γ are constants that can

be identified using a simple grid search, while ϕis

a compound coefficient for uniformly scaling net-

work depth, width and resolution. EfficientNet con-

sists of 8different models, namely EfficientNet-B0

to EfficientNet-B7, which are obtained by scaling the

baseline mesh. In this study, EfficientNet-B0 and

EfficientNet-B1 are used.

2.3 Transfer learning

Through the use of the machine learning task known

as transfer learning, it is possible to transfer the

knowledge gained from the training of a neural net-

work in one problem to another task or domain, [27].

This method has been widely used to efficiently train

models with limited data sets, and to overcome cost

and time constraints. The use of pre-trained mod-

els instead of training neural networks from scratch

speeds up the process, since the training model uses

information from previous training processes and al-

ready understands the features of the problem under

investigation. In summary, transfer learning produces

more reliable and generalised models, while signifi-

cantly reducing the risk of over-fitting.

2.4 Experimental setup

The study was conducted on a system running Win-

dows 11 Pro, equipped with a 2nd Gen Intel(R)

Core(TM) i7-12700 2.10GHz, 16GB RAM and a

1000GB SSD. Each of the models studied was trained

for 15 epochs using the pre-trained model weights

and an additional 5 epochs using its own weights to

train on all three class scenarios. Adam was used as

the optimiser, with a learning rate of 0.0001 and a

loss function of categorical cross-entropy. Data aug-

mentation techniques were used to extend the train-

ing set. Several techniques such as Horizontal Ran-

domFlip, RandomRotation(0.2), RandomWidth(0.2),

RandomHeight(0.2) and RandomZoom(0.2) were

used to transform existing images. Tensorflow,

Numpy, Keras, Matplotlib and sklearn libraries were

WSEAS TRANSACTIONS on ADVANCES in ENGINEERING EDUCATION

DOI: 10.37394/232010.2023.20.7

Konstantinos Strikas, Nikolaos Papaioannou,

Ioannis Stamatopoulos, Athanasios Angeioplastis,

Alkiviadis Tsimpiris, Dimitrios Varsamis,

Paraskevi Giagazoglou

E-ISSN: 2224-3410

Volume 20, 2023

used for the training and evaluation of the CNN mod-

els investigated in this study.

3 Results

As mentioned above, the main work of this study is

to compare state-of-the-art CNN architectures for the

assessment of preschoolers’ fine motor skills. Due to

the small sample size, and in order to further inves-

tigate the accuracy of the CNN models under study,

the aforementioned classes of the dataset (see 2.1)

were also merged into two and three classes. Thus,

although it is assumed that the 6-classes case more

accurately represents the groups of preschoolers stud-

ied, the 2-classes and 3-classes cases are also studied.

In the case of two classes, the classes 0,1,2and 3,4,5

from Table 1 were merged into classes 0and 1respec-

tively. Similarly, in the case of three classes, classes

0,1were merged into class 0, while classes 2,3and

4,5were merged into classes 1and 2respectively.

The sample size for each case is shown in detail in

Table 3. The training and testing accuracies of the in-

vestigated methods are presented both in tabular form

(Table 2) and in graphical form (Fig. 2).

Table 2: Accuracy of the examined CNN architectures

for a total of 2,3 and 6 classes.

Number of Classes 2 3 6

Training Testing Training Testing Training Testing

EfficientNetB0 74.13% 68.94% 60.57% 38.12% 48.62% 22.92%

EfficientNetB1 75.74% 49.07% 63.07% 44.79% 46.74% 25.42%

MobileNet 70.21% 69.98% 48.53% 45.21% 30.95% 31.25%

ResNet 71.81% 55.69% 58.07% 53.33% 43.53% 30.00%

VGG16 70.21% 69.98% 50.76% 50.63% 29.88% 30.21%

Figure 2: Accuracy of the examined CNN archi-

tectures, where CNNmodel_x refer to the number of

classes.

The results suggest that as the number of classes

increase, the accuracy of the examined methods de-

crease. For the case of two classes, two meth-

ods namely MobileNet and VGG16 both achieved

the highest accuracy (69.98%) on testing data, while

EfficientNet-B1 had the lowest accuracy (49.5%).

Similarly to EfficientNet-B1, ResNet50 didn’t per-

formed well, with an accuracy equal to 55.69%. For

the case of three classes, the accuracy of all the ex-

amined methods significantly decreased. In con-

trast to the 2-classes case, where ResNet50 performed

poorly compared to the other methods, for the 3-

classes case ResNet50 achieved the highest accuracy,

namely 53.33%, while EfficientNet-B0 performed

worse, with an accuracy equal to 38.12%, although

it achieved the second highest accuracy on training

data (60.57%). It should be noted that for the classi-

fication task of 3classes, only ResNet50 and VGG16

achieved a performance of more than 50%. In the case

of six classes, none of the examined methods man-

aged to achieve an accuracy of more than 50%, neither

on the test data nor on the training data. The highest

accuracy was 31.25% achieved by MobileNet, while

the second highest was 30.21% achieved by VGG16.

Similar to the case of three classes, EffiecientNet-

B0 performed poorly compared to the other investi-

gated methods, achieving a classification accuracy of

22.92%.

Beyond the case of two classes, these results ini-

tially suggest a poor classification performance. This

could be due to the small sample size. Table 3 shows

that the number of images in the training and test sets

is not evenly distributed. Specifically, out of a total

of 483 images in the test set, only 12 correspond to

class 0, while 30 and 49 correspond to classes 1and

5respectively. The above results for the 6-classes

case indicate that ResNet50 achieved one of the high-

est performances, namely 30%. From its class distri-

bution, which can be seen from the confusion matrix

(Fig. 4: (b)), it can be seen that for class 3, out of a

total number of 143 images, ResNet50 correctly clas-

sified 123 of them (86%). For classes 2and 4, which

also contained a large number of images in both the

training and test sets, the model had a poor accuracy

of 3% and 11.6% respectively. Nevertheless, for the

second class, 86 out of 100 drawings were classified

in the next class and for the fourth class, 111 out of

146 drawings were classified in the previous class.

This doesn’t seem to be the case for the classes men-

tioned above, to which a small number of drawings

correspond (Table 3), because for class 0, the draw-

ings weren’t classified either in the correct class or in

a neighbouring class. The same seems to be true for

both the first and the fifth class. Therefore, this could

be an indication that a larger data set is needed to in-

crease the accuracy.

As already mentioned, in the case of six classes,

even for classes 2and 4, which contain a larger

number of drawings compared to the other classes,

ResNet50 had a poor accuracy of 3% and 11.6% re-

WSEAS TRANSACTIONS on ADVANCES in ENGINEERING EDUCATION

DOI: 10.37394/232010.2023.20.7

Konstantinos Strikas, Nikolaos Papaioannou,

Ioannis Stamatopoulos, Athanasios Angeioplastis,

Alkiviadis Tsimpiris, Dimitrios Varsamis,

Paraskevi Giagazoglou

E-ISSN: 2224-3410

Volume 20, 2023

spectively. For the same case, although EfficientNet-

B1 had the second worst performance in general

(25.42%), it is observed that for class 2(see Fig. 4),

out of a total number of 100 images, 34 were correctly

classified (34%), while for class 4the accuracy was

around 15%. In addition, many of the images in the

test set are classified into neighboring classes. There-

fore, with the exception of class 3, for the classes with

more images in the test set, EfficieNet-B1 shows a

higher classification performance. It’s worth noting

that for class 1, EfficientNet-B1 was the only one to

correctly classify one image and 17 out of 30 to the

correct class or to a neighbouring class, in contrast to

the other methods examined, for which, for the same

case, they did not classify any image, either to the

correct class or to a neighbouring class. Thus, de-

spite its poor accuracy, EffiecientNet-B1 would tend

to be considered more convincing than the other mod-

els examined, since it is considered to be more dis-

criminative. The accuracy of VGG16, ResNet50 and

MobileNet is significantly increased due to the fact

that they classify most of the images of the test set in

class 3, regardless of their actual class (see Fig. 4).

This leads to a higher classification performance in

general, since class 3contains 143 out of 480 images

of the test set and VGG16, ResNet50 and MobileNet

achieve at least 85% accuracy in this class.

Figure 4 shows the rest of the confusion matri-

ces of the investigated CNN models for the 6-classes

case. Compared to MobileNet and VGG16, it can

be observed that EfficientNet-B0 and EfficientNet-

B1 are more discriminative. In particular, for class

2case, MobileNet and VGG16 classify 91% and

97% of the drawings to the fourth class, while e.g.

EfficientNet-B1 classifies 34% to the actual class and

1%,40%,10% and 15% to classes 1,3,4,5respec-

tively.

The same seems to be true for all cases studied.

For the case of 3-classes, it is observed that the mod-

els didn’t correctly classify any of the images of class

1(see Fig. 5 and Fig. 6). Table 2 illustrates that

VGG16 had the second best accuracy, specifically

50.63%, but from Figure 5 it is observed that all the

images are classified to class 2. Similar to VGG16,

the MobileNet classifies most of the images to class

2and the rest of them to class 3, achieving 45.21%

accuracy. For the case of three classes, ResNet50

seems to perform better compared to the other meth-

ods, both in terms of accuracy and of the allocation

of the images. ResNet50 achieved the highest accu-

racy, namely 53.33%. Figure 6 shows that ResNet50

classified correctly 0out of 42 images of class 1,155

out of 243 images of class 2and 93 out of 195 im-

ages of class 3, contrary to VGG16, which achieved

the second highest accuracy, but for classes 1and 3

it classified correctly 0out of 42 and 195 images re-

spectively, while its performance improved since it

classified all of the 242 images of class 2correctly

(see Fig. 5). Similar to ResNet, the EfficientNet-B0

and EfficientNet-B1 appeared to be more discrimi-

native, since the images were allocated to different

classes, while achieving 38.12% and 44.79% classifi-

cation accuracy respectively.

Table 3: Sample size for each class of train and test

set, where Train_x and Test_x, refer to train and test

set respectively for the x-classes case.

Class Train_6 Test_6 Train_3 Test_3 Train_2 Test_2

0 27 12 98 42 334 142

171 30 569 243 787 338

2236 100 454 195 - -

3333 143 - - - -

4335 146 - - - -

5119 49 - - - -

Figure 3: Confusion matrix of test set of EfficientNet-

B1 for the 6-classes case. Columns and rows refer to

predictive and actual values respectively.

In summary, for the three cases studied, VGG16

achieved the highest average classification accuracy

on testing data with 50.27%, followed by MobileNet

with an average accuracy of 48.8%. On the testing

data, the worst average classification accuracy was

observed by EfficientNet-B1, which was limited to

an accuracy of 39.76%. In contrast, on training data,

EfficientNet-B1 outperformed the other CNN models

examined with an average classification accuracy of

61.85%, followed by EfficientNet-B0 with 61.10%.

In addition, MobileNet achieved a classification ac-

WSEAS TRANSACTIONS on ADVANCES in ENGINEERING EDUCATION

DOI: 10.37394/232010.2023.20.7

Konstantinos Strikas, Nikolaos Papaioannou,

Ioannis Stamatopoulos, Athanasios Angeioplastis,

Alkiviadis Tsimpiris, Dimitrios Varsamis,

Paraskevi Giagazoglou

E-ISSN: 2224-3410

Volume 20, 2023

(a) EfficientNet-B0 (b) ResNet50

Figure 4: Confusion Matrices of the examined CNN

architectures for the 6-classes case.

curacy of 49.89%, which was the worst performance

on training data.

4 Conclusion

This study compared five state-of-the-art CNN archi-

tectures to investigate their accuracy in assessing fine

motor skills in preschool children. The data set used

consists of drawings of children (male or female),

classified by experts according to the children’s de-

velopmental age. The Griffiths II and the Eye Co-

ordination Scale were used to assess the preschool

children’s developmental age. As for the results, on

testing data, besides that VGG16 achieved the high-

est average classification accuracy of 50.27% usually

classifies most of data on a specific class. On training

data, EfficientNet-B1 achieved the highest classifica-

tion performance with an average classification accu-

racy of 61.85%. For the case of 3-classes ResNet50

seems to classify the drawings best. For the case of

6-classes EffiecientNet-B1 seems to perform better

compared to the other models, since it allocates better

the images to the correct classes or to the neighboring

classes. In general, all the methods examined showed

poor classification accuracy. This may be due to the

small sample size. Confusion matrices confirm that

the classes with larger sample sizes improve the clas-

sification, in contrast to the classes with fewer draw-

ings in their image set. In summary, the results sug-

gest that although challenging, automatic and accu-

rate scoring of fine motor skills may be feasible when

(a) EfficientNet-B0 (b) EfficientNet-B1

Figure 5: Confusion Matrices of the examined CNN

architectures for the 3-classes case.

using a larger dataset.

5 Declarations

All authors declare that they have no conflicts of in-

terest.

References:

[1] M. Katagiri, H. Ito, Y. Murayama, M. Hamada,

S. Nakajima, N. Takayanagi, A. Uemiya,

M. Myogan, A. Nakai, and M. Tsujii, “Fine

and gross motor skills predict later psychoso-

cial maladaptation and academic achievement,”

Brain and Development, vol. 43, no. 5, pp. 605–

615, 2021.

[2] S. Suggate, E. Pufke, and H. Stoeger, “Chil-

dren’s fine motor skills in kindergarten predict

reading in grade 1,” Early Childhood Research

Quarterly, vol. 47, pp. 248–258, 2019.

[3] U. Fischer, S. P. Suggate, and H. Stoeger, “Fine

motor skills and finger gnosia contribute to

preschool children’s numerical competencies,”

Acta Psychologica, vol. 226, p. 103576, 2022.

[4] A. Asakawa and S. Sugimura, “Mediating pro-

cess between fine motor skills, finger gnosis,

and calculation abilities in preschool children,”

Acta Psychologica, vol. 231, p. 103771, 2022.

WSEAS TRANSACTIONS on ADVANCES in ENGINEERING EDUCATION

DOI: 10.37394/232010.2023.20.7

Konstantinos Strikas, Nikolaos Papaioannou,

Ioannis Stamatopoulos, Athanasios Angeioplastis,

Alkiviadis Tsimpiris, Dimitrios Varsamis,

Paraskevi Giagazoglou

E-ISSN: 2224-3410

Volume 20, 2023

Figure 6: Confusion Matrix of ResNet50 architecture

for the 3-classes case.

[5] P. Martzog and S. Suggate, “Fine motor skills

and mental imagery: Is it all in the mind?,” Jour-

nal of Experimental Child Psychology, vol. 186,

pp. 59–72, 2019.

[6] E. Michel and S. Molitor, “Fine motor skill au-

tomatization and working memory in children

with and without potential fine motor impair-

ments: An explorative study,” Human Move-

ment Science, vol. 84, p. 102968, 2022.

[7] P. Martzog and S. Suggate, “Screen media are

associated with fine motor skill development in

preschool children,” Early Childhood Research

Quarterly, vol. 60, pp. 363–373, 2022.

[8] I. Gentier, E. D’Hondt, S. Shultz, B. Deforche,

M. Augustijn, S. Hoorne, K. Verlaecke, I. De

Bourdeaudhuij, and M. Lenoir, “Fine and gross

motor skills differ between healthy-weight and

obese children,” Research in Developmental

Disabilities, vol. 34, no. 11, pp. 4043–4051,

2013.

[9] D. Gaul and J. Issartel, “Fine motor skill profi-

ciency in typically developing children: On or

off the maturation track?,” Human Movement

Science, vol. 46, pp. 78–85, 2016.

[10] S. W. Bindman, L. E. Skibbe, A. H. Hindman,

D. Aram, and F. J. Morrison, “Parental writ-

ing support and preschoolers’ early literacy, lan-

guage, and fine motor skills,” Early Childhood

Research Quarterly, vol. 29, no. 4, pp. 614–624,

2014.

[11] L. Dapp, V. Gashaj, and C. Roebers, “Physical

activity and motor skills in children: A differen-

tiated approach,” Psychology of Sport and Exer-

cise, vol. 54, p. 101916, 2021.

[12] H. McGlashan, C. Blanchard, N. Sycamore,

R. Lee, B. French, and N. Holmes, “Improve-

ment in children’s fine motor skills following

a computerized typing intervention,” Human

Movement Science, vol. 56, pp. 29–36, 2017.

[13] P. Giagazoglou, V. Tsimaras, E. Fotiadou,

C. Evaggelinou, J. Tsikoulas, and N. An-

gelopoulou, “Standardization of the motor

scales of the griffiths test ii on children aged 3–6

years in greece,” Child: care, health and devel-

opment, vol. 31, pp. 321–30, 06 2005.

[14] S. E. Henderson, D. Sugden, and A. Barnett,

“Movement assessment battery for children-2,”

Research in Developmental Disabilities, 1992.

[15] H. Ito, W. Noda, S. Nakajima, Y. Tanaka,

M. Hamada, and M. Katagiri, “Prediction

of psychosocial maladaptation in elementary

school based on development appraisal by nurs-

ery teacher: validation of the developmental

scale for nursery record (dsnr),” Japanese J Dev

Psychol, vol. 27, p. 59 – 71, 2016.

[16] J. Tiffin and E. Asher, “The purdue pegboard;

norms and studies of reliability and validity.,”

The Journal of applied psychology, vol. 32,

p. 234–247, 1948.

[17] K. Strikas, A. Valiakos, A. Tsimpiris,

D. Varsamis, and P. Giagazoglou, “Deep

learning techniques for fine motor skills as-

sessment in preschool children.,” International

Journal of Education and Learning Systems,

vol. 7, pp. 43–49, 2022.

[18] Y. Li, J. Nie, and X. Chao, “Do we really

need deep cnn for plant diseases identifica-

tion?,” Computers and Electronics in Agricul-

ture, vol. 178, p. 105803, 2020.

[19] J. Yang, L. Zhang, X. Tang, and M. Han,

“Codnnet: A lightweight cnn architecture for

detection of covid-19 infection,” Applied Soft

Computing, vol. 130, p. 109656, 2022.

[20] A. Gupta, A. Anpalagan, L. Guan, and A. S.

Khwaja, “Deep learning for object detection and

scene perception in self-driving cars: Survey,

challenges, and open issues,” Array, vol. 10,

p. 100057, 2021.

WSEAS TRANSACTIONS on ADVANCES in ENGINEERING EDUCATION

DOI: 10.37394/232010.2023.20.7

Konstantinos Strikas, Nikolaos Papaioannou,

Ioannis Stamatopoulos, Athanasios Angeioplastis,

Alkiviadis Tsimpiris, Dimitrios Varsamis,

Paraskevi Giagazoglou

E-ISSN: 2224-3410

Volume 20, 2023

[21] M. Tan and Q. V. Le, “Efficientnet: Rethink-

ing model scaling for convolutional neural net-

works,” 2019.

[22] K. He, X. Zhang, S. Ren, and J. Sun,

“Deep residual learning for image recognition,”

pp. 770–778, 06 2016.

[23] K. Simonyan and A. Zisserman, “Very deep con-

volutional networks for large-scale image recog-

nition,” arXiv 1409.1556, 09 2014.

[24] A. Howard, M. Zhu, B. Chen, D. Kalenichenko,

W. Wang, T. Weyand, M. Andreetto, and

H. Adam, “Mobilenets: Efficient convolutional

neural networks for mobile vision applications,”

04 2017.

[25] A. Krizhevsky, I. Sutskever, and G. Hinton,

“Imagenet classification with deep convolu-

tional neural networks,” Neural Information

Processing Systems, vol. 25, 01 2012.

[26] M. Tan, B. Chen, R. Pang, V. Vasudevan,

M. Sandler, A. Howard, and Q. Le, “Mnas-

net: Platform-aware neural architecture search

for mobile,” pp. 2815–2823, 06 2019.

[27] Y. Bengio, G. Guyon, V. Dror, G. Lemaire,

D. Taylor, and D. Silver, “Deep learning of rep-

resentations for unsupervised and transfer learn-

ing,” vol. 7, 01 2011.

Contribution of Individual Authors to the

Creation of a Scientific Article (Ghostwriting

Policy)

The authors equally contributed in the present

research, at all stages from the formulation of the

problem to the final findings and solution.

Sources of Funding for Research Presented in a

Scientific Article or Scientific Article Itself

No funding was received for conducting this study.

Conflicts of Interest

The authors have no conflicts of interest to declare

that are relevant to the content of this article.

Creative Commons Attribution License 4.0

(Attribution 4.0 International, CC BY 4.0)

This article is published under the terms of the

Creative Commons Attribution License 4.0

https://creativecommons.org/licenses/by/4.0/deed.en

_US

WSEAS TRANSACTIONS on ADVANCES in ENGINEERING EDUCATION

DOI: 10.37394/232010.2023.20.7

Konstantinos Strikas, Nikolaos Papaioannou,

Ioannis Stamatopoulos, Athanasios Angeioplastis,

Alkiviadis Tsimpiris, Dimitrios Varsamis,

Paraskevi Giagazoglou

E-ISSN: 2224-3410

Volume 20, 2023