Multimedia Tools and Applications (2022) 81:19013–19033

https://doi.org/10.1007/s11042-020-10090-6

A knowledge-supported approach for garment pattern

design using fuzzy logic and artificial neural networks

Zhujun Wang 1,2,3,5 & Yingmei Xing 1 & Jianping Wang 2,5 & Xianyi Zeng 4 &
Yalan Yang 2 & Shuo Xu 2

Received: 8 May 2020 / Revised: 31 August 2020 / Accepted: 15 October 2020 /

Published online: 29 October 2020
# Springer Science+Business Media, LLC, part of Springer Nature 2020

Abstract
In the clothing industry, garment pattern design serves as a significant middle link
between fashion design and manufacturing. With the advent of advanced multimedia
techniques, like virtual reality, 3D modeling, and interactive design, this work has
become more intuitive. However, it is still a tremendous knowledge-based work that
relied on the experienced patternmakers’ know-how. For enterprises, it will take a long
time to cultivate a patternmaker from an abecedarian to an expert. Moreover, while facing
fierce competition in the market, enterprises still have to endure the pressures and risks
led by the turnover of experienced patternmakers. In this context, we put forward a
knowledge-supported garment pattern design approach by learning the experienced
patternmakers’ knowledge based on fuzzy logic and artificial neural networks. Based
on this approach, we created a knowledge-supported pattern design model, consisting of
several sub-models following the garment styles, considering the properties of fabrics and
fitting degree. The inputs of the model are the feature body dimensions, while the outputs,
namely the pattern parameters, can be generated following the fabric and fitting degree
selected. Through performance comparison with other models and the actual fitting test,
the results revealed that the present approach was applicable. Our proposed approach can
not only support the non-expert patternmakers or abecedarians to make decisions when
developing the patterns by reducing the difficulties of patternmaking but help the
enterprises to reduce the dependencies on the experts, hence promoting the efficiency
and reducing risks.

Keywords Garment pattern design . Knowledge modeling . Fuzzy logic . Artificial neural
networks . Anthropometric measurements

* Jianping Wang
wangjp@dhu.edu.cn

Extended author information available on the last page of the article

19014 Multimedia Tools and Applications (2022) 81:19013–19033

1 Introduction

With the advances of the novel informatics and multimedia technologies, such as virtual
reality, artificial intelligence, and deep learning, the clothing industry has succeeded in a rapid
growth worldwide [27]. It is reported that clothing production has almost doubled during the
past two decades [8]. However, many consumers still argue that it is difficult to find the
garments exactly meet their personalized needs in garment fit, fashion styles, colors, and
fabrics, etc. This phenomenon leads to the high recall rete of the apparel, which brings great
pressures and risks for the clothing enterprises [6]. Providing a satisfactory garment fit is
considered as an essential and critical issue, since it is the fundamental demands for con-
sumers. No matter how attractive and beautiful the garment is, the consumer will not make the
purchase if it cannot fit the consumer body appropriately. Therefore, it is a key issue for the
clothing industry to resolve is to concentrate on innovating and developing a new design
approach so as to fulfill personalized garments for consumers with better efficiency and
garment fit. In order to address this issue, mass customization (MC), integration of made-to-
measure and mass production, has become a trend for enterprises in the clothing sector to
deliver personalized products and services of high quality with acceptable cost rapidly in an
optimized way [29, 37]. With the concept of MC, the enterprises should offer garments more
efficient and fitting under controllable costs. However, more efficient and fitting, as well as the
controllable costs, can hardly be realized simultaneously by the conventional garment pattern
design approaches.
Garment patterns refer to the paper or cardboard templates based on which a whole piece of
fabric could be cut into different pieces before being sewn to form a garment [23, 31]. Garment
pattern design, also named as pattern cutting, pattern drafting or structure design, etc., plays a
significant role in clothing design and manufacturing, influencing the garment fit, production
efficiency and costs [21]. Traditionally, there are two main modes of garment pattern design:
mass production and made-to-measure. These approaches exist the following protruding
shortcomings. First, in mass production, in order to shorten the response time to the market,
the patterns are developed using the human body parameters based on a standard sizing system
and the graduation rules instead of taking measuring of the target consumers. Thus, garment
patterns made in this way may suit the standard human body shape, but can hardly fit well to
all kinds of body shapes. Second, in made-to-measure, patterns are designed based on the
anthropometric data measured from the target consumers, so the ill garment fitting can be
avoided, nonetheless, it is time-consuming and costly. Third, the existing pattern design
methods depend heavily on patternmakers’ expert know-how [20, 22], so that the product
qualities fluctuate from different pattern makers. For enterprises, it will take a long time to
cultivate a patternmaker from an abecedarian to a knowledgeable expert [23]. Otherwise, the
enterprises have to endure the pressures and risks led by the turnover of experienced
patternmakers. Once the experienced patternmakers resign, it will cause immeasurable losses
for the enterprises. Therefore, an effective garment pattern design approach is needed for the
clothing industry to overcome the shortages abovementioned.
To resolve the issues of production efficiency and garment fit, many researchers have
attempted to design garment patterns using 3D-to-2D flattening technology, with the devel-
opment of advanced multimedia technologies, such as 3D virtual garment design, 3D virtual
try-on, and 3D animation. Tao et.al. proposed a novel garment pattern design concept by
conceiving the virtual clothing directly on a mannequin morphology in cyberspace with
consideration of the ease allowance between the body shape and the garment, which
Multimedia Tools and Applications (2022) 81:19013–19033 19015

significantly reduces the production time by eliminating the process of 2D patternmaking [28].
Sebastien et.al. provided a methodology integrating 3D scanning and 3D-to-2D flattening
technology to define, quantify and control the ease allowance in 3D garment pattern making
for ready-to-wear or customized garments [30]. A new 3D flattening-based garment design
process was proposed by Pascal and Hong et.al. to provide form-fitting garments for disabled
peoples with atypical body shapes [2, 9–12]. Liu et.al. also applied 3D-to-2D flattening
technology to develop upper cycling clothes [19]. A 3D interactive garment pattern making
technology was put forward by Liu et al. to develop garment patterns efficiently in a “What
you see is what you get” way [20]. However, there are some drawbacks existing in the
abovementioned works. First, the method is restricted to simple styles and can hardly be
implemented in complex styles. Second, the method is limited to tight-fitting garments with
little ease allowance between the human body and garment. Third, the fabric properties
affecting the garment appearances are ignored in the process of flattening. Fourthly, the
operation of this method is rather complicated and time-costing and is not suitable for
industrial production. Therefore, this method has been rarely applied in real industrial
production.
In this context, we propose a machine learning-based garment pattern design knowledge-
supported model (see Fig. 1) to predict and control garment patterns parameters by learning
data from knowledgeable and experienced patternmakers’ know-how, aiming at making the
patterns more efficient and fitting, and reducing the dependencies on the experts. The proposed
model is constructed based on fuzzy logic (FL) and artificial neural networks (ANN) technol-
ogies, which takes the advantages of the FL and ANN in identifying, learning and generalizing
abilities of nonlinear relationships in a dataset with limited dimensions and sizes. The inputs of
the model are the feature body dimensions, while the outputs, namely the pattern parameters,
can be generated following the garment styles, the fabrics and fitting degrees selected. The
main contributions of our works are as follows:

& Provide a feasible and valid solution to resolve the contradictory issue between better
garment fit and production efficiency for the clothing industry in the context of mass
customization;
& Offer a generalized design approach which can be easily spread and utilized for other types
of garments;
& Powerfully support for patternmakers (especially the abecedarian) to make decisions in
personalized garment pattern design by reducing the operation difficulties;
& Largely release the heavy dependencies on the experienced experts for the enterprises and
reduce the risks led by the knowledgeable and experienced experts’ turnover;
& Promote to form a new collaborative garment design process and system by integrating the
proposed knowledge-supported model into a commercial 3D clothing design software.
This new design process and system can effectively promote the level of mass

Fig. 1 Machine learning-based garment pattern design knowledge-supported model

19016 Multimedia Tools and Applications (2022) 81:19013–19033

customization, aiming at providing personalized fashion products and services for con-
sumers more intuitively and efficiently;
& Easily extend the idea proposed in our study to other human-centered product design. A
corresponding knowledge-supported human-centered product design model can be con-
structed using the same methodology proposed to model the relationship between human
body dimensions and product parameters firstly. And then, the relevant knowledge-
supported collaborative garment design process and decision support system can also be
developed.

The remaining sections of this study are structured as follows. In the section two, the
framework and implementation process were expounded mainly. The section three discussed
how to construct the knowledge-supported model for garment pattern design. The section four
analyzed and evaluated the present knowledge-supported model by using actual fitting
techniques. Lastly, the conclusion and future works were present in the section five.

2 Methodology

2.1 Related works

Presently, artificial intelligence (AI)-based machine learning techniques, such as artificial

neural networks (ANN), deep learning, fuzzy logic, and genetic algorithm, has attracted
uprising interests and been applied to many tasks. For example, a convolutional neural
network integrated by an attention module was present by Ling et.al. for deep face recognition
[16]. Bin et.al. put forward an innovative multi-modal fusion method to incorporate the
geographical presentation from street maps into the real estate appraisal model with a deep
neural network [1]. An Attention-Based Modality-Gated Networks (AMGN) was put forth by
Huang et.al. to exploit the interrelationship between the modalities of images and texts and
extract the discriminative features for multimodal sentiment analysis [14]. Besides these
applications, as a kind of typical AI-based technique, due to the ability to identify nonlinear
relationships in a dataset, fuzzy logic model has been applied intensively in quantities of fields
of the clothing industry, including involving garment design [9, 11], garment evaluation [33,
34], apparel size selection [26], fashion knowledge base development [15, 36], and recom-
mendation system development [7, 35], etc.. In the field of garment pattern design, Chen Y.
et al. present a method of automatic pattern generation, based on the estimation of ease
allowance using fuzzy logic and sensory evaluation [4, 5]. However, However, the fuzzy
logic model has the drawback of lacking the abilities to acquire knowledge from the dataset
and tuning the parameters of membership functions adaptively.
Compared with the fuzzy logic technique, the ANN has the advantages of learning and
generalizing ability [25]. Due to this, researchers have paid attention to apply the ANN in
garment pattern design. In [3], Chan et al. presented a three-layer back-propagation neural
network model to predict shirt pattern parameters aimed to design patterns to fit a wide range
of body morphology. In 2009, Hu et al. developed a hybrid system composed of neural
network and immune co-evolutionary algorithm to predict garments’ fitness and find out
optimal sizes [13]. Xing et al. put forward a new approach for auto-generating the sleeve
pattern parameters directly from anthropometric measurements by modeling BP ANN in 2014
Multimedia Tools and Applications (2022) 81:19013–19033 19017

[32]. In 2017, Liu et al. construct a backpropagation artificial neural network model to estimate
pattern making-related body dimensions by inputting few key human body dimensions [17]. In
2019, another study by Wang et al. put forth a new predictive model for estimating body
dimensions using radial basis function artificial neural networks [31]. However, there are some
shortcomings to the methods abovementioned. First, these studies focused only on a fixed
garment style rather than a series of garment styles that restricted the realistic implementation
in industrial production. Second, the factors influencing the garment pattern parameters are
considered inadequately, such as the consumers’ preference, fabric properties, and craftsman-
ship. Third, there is no regard for the exploitation of the proficient patternmakers’ know-how
for non-experts or abecedarians.
To date, little research little attention was concentrated on employing a hybrid method of
fuzzy logic and artificial neural networks to exploit the experienced patternmakers’ know-how
for personalized pattern design. In this context, we proposed a new knowledge-supported
methodology of personalized garment patternmaking by constructing a hybrid model of fuzzy
logic and artificial neural networks. The integration of FL and ANN gives a human-machine
friendly knowledge representation framework for acquiring, representing and using the knowl-
edge of the domain expert, this protruding advantage of which makes it suitable for estimating
the garment pattern parameters where the exact transfer functions cannot be well modeled, and
training data sets are limited.

2.2 Framework of the knowledge-supported model for garment pattern design

For the classical pattern design approach, after acquiring the information from consumers in
terms of anthropometric measurements, garment styles, preferences, and usage, pattern makers
will perform the parameterization of the information for patternmaking by using their own
experts’ know-how. The unique knowledge is represented implicitly in the parameters called
ease allowances, by adding which on the body dimensions the patternmakers can obtain the
pattern parameters. To exploit the expertise of patternmakers and release the dependencies on
patternmakers’ knowledge, we propose an alternative solution to estimate the garment pattern
parameters by constructing a patternmakers’ knowledge-supported model using fuzzy logic
and artificial neural networks techniques through learning the experimental data regarding the
influencing factors sufficiently.
This paper concentrates on constructing the pattern designers’ knowledge-supported model
aimed to combine the patternmakers’ expert know-how to generating the pattern parameters
automatically and rapidly, hence improving the efficiency of pattern designing, as well as the
fitness for consumers. The learning data are quantitatively extracted from the personalized
garment patterns finely designed by the experienced and skillful patternmakers with a career of
over ten years in this domain. From the learning data obtained, a model based on FL and ANN
is set up. The model has two primary features. First, it will be capable of estimating the
garment pattern parameters automatically by mimicking the patternmakers with sufficient
consideration of the influencing factors. Second, it will be friendly and conveniently imple-
mented by the practitioners without adequate experience and skill, especially for the
abecedarians.
We created the overall architecture of the knowledge-supported model initially. Then, the
model is split into several sub-models, and each sub-model has a one-to-one relationship with
the pattern parameters, with diverse inputs and outputs. Figure 2 illustrates the general
framework of the proposed model and structure of the sub-model. In Fig. 2, Fbdij represents
19018 Multimedia Tools and Applications (2022) 81:19013–19033

Fig. 2 The general framework of the knowledge-supported model

the jth feature body dimension for the ith type garment, and Ppik is denoted as the kth pattern
parameters for the ith type garment.

2.3 Construction of the knowledge-supported model

The knowledge-supported model proposed in this research is constructed by the following

steps:

Step 1: Layer 1 was the input layer, including one or two nodes in this study. The inputs and
outputs in the input layer were expressed as follows:
ð1 Þ
inputi ¼ Bdi ; i ¼ 1; 2 ð1Þ


ð1Þ ð1Þ ð1Þ
output i ¼ f i inputi ¼ inputi ; i ¼ 1; 2 ð2Þ

ð1Þ
Where, outputi was the output of the ith node in Layer 1.

Step 2: Layer 2 was also named the membership function layer. The outputs in Layer 2 were
calculated according to formula (3).


ð2Þ ð1Þ
output j ¼ μ j outputi ; j ¼ 1; 2; …; m ð3Þ
Multimedia Tools and Applications (2022) 81:19013–19033 19019

ð2 Þ
Where, output j was denoted as the outputs of the jth node in Layer 2; μj was the membership
function of the jth node in Layer 2.

Step 3: Layer 3 was knowns as the rule layer. Each node of Layer 3 was labeled Π in Fig. 3,
which was produced by the nodes of outputs in each group, as illustrated in Fig. 6.
The outputs of this layer were computed following formula (4) and (5).


ð3Þ ð23Þ ð2Þ
input k ¼ ∏k ωjk output j ð4Þ


ð3Þ ð3 Þ ð3Þ
outputk ¼ f k inputk ¼ input k ð5Þ

ð3Þ ð23Þ
Where, output k referred to the outputs of the kth node in Layer 3; ωjk was the weighted
value between Layer 2 and Layer 3.

Step 4: Layer 4 was the output layer of the model. Only one fixed existed in this layer, which
computed the overall output as the summation of all inputs from Layer 3.


ð34Þ ð3Þ
inputðo4Þ ¼ ∑ ωko outputk ð6Þ
k


outputðo4Þ ¼ f o inputðo4Þ ¼ inputðo4Þ ð7Þ

Fig. 3 Flat patterns of pants

19020 Multimedia Tools and Applications (2022) 81:19013–19033

ð34Þ
Where, ωko was the output strength of the kth rule; output ðo4Þ referred to the outputs of the
model.

Step 5: After the output of the model was computed in the output layer, the error measure
was calculated and checked.
1 1
E¼ ∑ ðDi −Oi Þ2 ¼ ∑i e2 ð8Þ
2 i 2

Where E was defined as the cost function used to measure the error; Di referred to the desired
outputs of the model.
If the error of the model met the goal or were acceptably small, the model would cease.
Else, all the modifiable parameters would be tuned in the backward pass through the training
method until the sum of the squared errors between the model output and the desired output
reached the minimum. The steepest descent method was introduced to adjust the linked
weighted values in the model. The process of calculation based on the back-propagation
algorithm was expressed as follows.

Step 1: In the output layer (Layer 4), the error δðo4Þ from the backpropagation was computed
following formula (9).

∂E ∂E ∂e ∂Oi ∂outputðo4Þ
δðo4Þ ¼ − ¼ −    ¼e ð9Þ
∂input ðo4Þ ∂e ∂Oi ∂outputðo4Þ ∂inputðo4Þ

The iterated value in every epoch of the weighted value was computed:

ð34Þ ∂E ∂E ∂output ðo4Þ ∂inputðo4Þ ð3Þ

Δωko ¼ −η ð34Þ
¼ −η  ð4Þ
  ¼ ηδðo4Þ outputk ð10Þ
∂ωko ∂outputo ∂inputðo4Þ ð34Þ
∂ωko

Where η was the learning rate.

Hence, the weighted value in the output layer could be modified by formula (11).
ð34Þ ð34Þ ð34Þ
ωko ðN þ 1Þ ¼ ωko ðN Þ þ Δωko ðN Þ ð11Þ

Where N was the number of iteration epoch.

Step 2: In Layer 3, the formula (12) used to calculate the error δðo3Þ from the backpropagation
was expressed as follows.
ð3Þ
ð3Þ ∂E ∂E ∂outputðo4Þ ∂input ðo4Þ ∂outputk ð34Þ
δk ¼ − ð3Þ
¼− ð4Þ
   ¼ δðo4Þ ωko ð12Þ
∂input k ∂output o ∂inputðo4Þ ∂output ðk3Þ ∂inputðk3Þ

ð23Þ
The linked weighted value ωjk was amended according to the outputs of layer 2 and the error
ð3Þ
δk in layer 3.
Multimedia Tools and Applications (2022) 81:19013–19033 19021

ð3Þ
ð23Þ ∂E ∂E ∂input k ð3Þ ð2Þ
Δωjk ¼ −η ð23Þ
¼ −η  ð3Þ
 ð23Þ
¼ ηδk output j ð13Þ
∂ωjk ∂output k ∂ωjk

ð23Þ ð23Þ ð23Þ

ωjk ðN þ 1Þ ¼ ωjk ðN Þ þ Δωjk ðN Þ ð14Þ

Where N was the number of iteration epoch.

At the beginning of the phase of training, training error between the model output and the
desired output was set to 0, whereas the maximum of iteration epochs was set to 1000 epochs.
If either of the goals reached, the training procedure of the model terminated.

3 Experiment

The general principles of approach proposed can be extended to other garment styles. For
simplicity, we elaborate it using a specific case of male’s pants in this study.

3.1 Acquisition of anthropometric data

Information on anthropometric data is a precondition of garment patterns design. Contributing

to the excellent performances on the accuracy, high resolution, and processing velocity, 3D
whole-body scanning systems exert more impact on guaranteeing good garment fitness for the
apparel industry [24]. Thus, we select randomly 172 males aged 20–48 years-old in China, and
then, perform an anthropometric experiment to acquire human body data by using a laser-
based 3D body scanner (Vitus Smart) in this study. To reduce the inevitable experimental
errors, we scanned each instance three times, and then adopted the mean values eventually. For
symmetrical body parts (i.e., leg), only the dimensions corresponding to the right leg were
measured, under the hypothesis that the human body is symmetry.
To evaluate our proposed approach close to reality, we split the data samples into two
groups.

& A learning dataset consists of 68 instances selected from the population. The present model
and the contrasted models are constructed based on this dataset.
& A testing dataset, composed of the remaining 104 instances, serves as the real demands of
the consumers.

3.2 Definition of the primary dimensional constraint parameters

Observe from the patterns of pants (see Fig. 3), we can easily find that the pants’ pattern is
composed of various straight and curve lines determined by various constraint parameters. Too
many parameters will dramatically affect the efficiency of patternmaking. Therefore, we should
identify the primary constraint parameters in pattern design, which have a significant impact on
garment comfort and appearance. In this phase, we took surveys on 15 experienced patternmakers
who have engaged in the development of pants over ten years. According to the survey results,
we selected seven pattern parameters as the key pattern parameters initially, and they were pants
19022 Multimedia Tools and Applications (2022) 81:19013–19033

length, knee length, crotch length, waist girth, hip girth, crotch girth, and knee girth. These pattern
parameters are constraint by the corresponding body dimensions. Some studies have demon-
strated that the human statue, waist girth, and hip girth have interrelationship firmly with other
body dimensions [17, 18]. Therefore, we selected the human statue, waist girth, and hip girth as
the primary dimensional constraint parameters in our study, namely feature body dimensions.
Hence, we defined the pattern parameters corresponding to the feature body dimensions as the
primary pattern parameters, while the rest pattern parameters as the secondary ones.

3.3 Definition of fit-enhanced pattern parameters

If we want to make a well-designed personalized garment pattern, all the pattern constraint
parameters should be under consideration.
Let Pp = {Pp1, Pp2, ⋯, Ppn} be aset of n constraint parameters in pants patterns;
Let Pppr ¼ Pppr pr
1 ; Pp2 ; ⋯; Ppk
pr
be a set of k primary constraint parameters in pants
patterns;  
Let Ppse ¼ Ppse 1 ; Pp2 ; ⋯; Ppl
se se
be a set of l secondary constraint parameters in pants
patterns;
Both Pppr and Ppse are the subset of Pp.
Let Bd = {Bd1, Bd2, ⋯, Bdn} be a set of n body dimensions corresponding to the pattern
parameters of pants;
 
Let etotal ¼ etotal
1 ; e2 ; ⋯; en
total total
be a set of n total ease allowances corresponding to the
constraint parameters
n in pants patterns; o
Let efactor ¼ efactor
1 ; efactor
2 ; ⋯; efactor
m be a set of ease allowances affected by m factors;
For a given pattern parameter Ppi, the relationship between pattern parameters Ppiand its
corresponding body dimension Bdi can be denoted as bellow:

Ppi ¼ Bdi þ etotal

i ; i ¼ 1; 2; ⋯; n ð15Þ

etotal
i ¼ efactor
1 þ efactor
2 þ ⋯ þ efactor
j ; j ¼ 1; 2; ⋯; m ð16Þ

In the apparel industry, for the sake of efficiency, the primary pattern parameters are calculated
by the formula (15) and (16) firstly. And then, the secondary pattern parameters are usually
computed by the formula (17).
Ppse ¼ f ðPppr Þ ð17Þ

This method has two main drawbacks. First, it is still involved for the non-experts, as it is difficult
for them to deal with etotal accurately. Second, the garment fit cannot be guaranteed at anywhere, as
Ppse is inferred from Pppr, not directly calculated from the corresponding body dimensions.
In the realistic process of pattern design, patternmakers determine the pattern parameters
following the characteristics of fabrics to actualize the specific garment style. In this context,
we employed 5 kinds of fabrics with various compositions and textile structures in our
experiment. Table 1 shows the fabric parameters applied in our study.
We defined the ease values for the body dimensions related to the key pattern parameters
following various fitting degree, in terms of loose fit, medium fit, and tight fit, based on the data
Multimedia Tools and Applications (2022) 81:19013–19033 19023

Table 1 The fabric parameters

No. Composition Structure Weight (g/m2) Thickness (cm) Bending Resistance B

(1e-6 N.m)

Warp Weft

1 100% cotton Woven 157 0.04 42.612 13.795

2 65% Polyester, 35% cotton Woven 223 0.04 24.04 16.19
3 100% wool Woven 288 0.07 17.66 11.77
4 100% Polyester Knit 120 0.06 1.01 0.22
5 87% Polyamide 13% Elasthane Knit 220 0.06 0.63 0.63

obtained by the experienced patternmakers and questionnaires, as demonstrated in Table 2. Hence,

we achieved the fit-enhanced pattern parameters at the conditions of various fabrics. Finally, we
acquired the experimental dataset, consisting of feature body dimensions and fit-enhanced pattern
parameters.

3.4 Definition of graphics constraint parameters

From the view of graphics, the smoothness of the curve lines is a critical criterion for evaluating
the garment patterns, influencing the appearances of the garments. Moreover, the smooth curve
lines will significantly reduce the difficulties in the process of production, hence improving the
productive efficiency. For the pants patterns, the side seams and the crotch lines around the
waist-to-hip area are the most curve lines, strongly related to the comfort and appearance. In
order to keep the curvatures of these curves in a controllable degree, we defined the graphics
constraint parameters in three positions at the waist lines (see Fig. 4). For a given waist girth and
hip girth, we should adjust the parameters of front pleats and back darts to comply with the
outline constraints.

3.5 Modelling the relationships between the feature body dimensions

and the pattern parameters

3.5.1 Modeling the relationships between the feature body dimensions and the pattern
parameters in the height direction

As the body dimensions corresponding to the pattern parameters in the height direction have a
close relationship with stature, we created three sub-models with the same single input to
predict pants length, crotch length, and knee length, respectively. We defined various mem-
bership functions for each sub-model, which were the dsigmf, trapmf, and gaussmf function for
sub-model 1, 2, and 3, respectively. The designed sub-models for the pattern parameters in the
height direction are shown in Fig. 5.

Table 2 Ease allowances for the specific body dimensions related to the key pattern parameters (Unit: cm)

Pattern parameters Waist girth Hip girth Crotch girth Knee girth Crotch length Fitting degree

Ease allowance I 2 10 12 10 1.5 loose

Ease allowance II 1 4 6 4 1 neutral
Ease allowance III 0 0 0 0 0 tight
19024 Multimedia Tools and Applications (2022) 81:19013–19033

Fig. 4 Graphics constraint parameters of pants patterns

3.5.2 Modeling the relationships between the feature body dimensions and the pattern
parameters in the girth direction

As the primary pattern parameters in the girth direction, pants waist girth and hip girth can be
calculated by adding ease to the corresponding human body dimensions intuitively. Therefore,
to simplify the model, we utilized waist girth and hip girth as the inputs for the sub-models to

Fig. 5 The sub-models for the pattern parameters in the height direction
Multimedia Tools and Applications (2022) 81:19013–19033 19025

predict pants’ waist girth and hip girth (see Fig. 6). The membership functions of these two
sub-models were both trimf functions.
As the secondary pattern parameters in the girth direction, both pants crotch girth and knee
girth have a close interrelationship with the primary pattern parameters. Thus, for predicting
the crotch girth and knee girth, we constructed the sub-models by choosing the waist girth and
hip girth as the inputs simultaneously (see Fig. 7). The membership function for the crotch
girth sub-model was denoted as the trimf function, while that of the knee girth sub-model was
the gbellmf function.

4 Results and discussions

In order to realize and verify the proposed knowledge-supported system, we constructed and
trained the model for a specific fabric with a given fitting degree using the learning dataset
initially. Then, we extracted the corresponding testing data from the testing dataset to verify
our model. Afterward, the same procedures were extended to other fabrics with the various
fitting degree. Table 3 shows parts of the learning data and testing data in our experiment.
Ultimately, we introduced the metrics of root mean squared error (RMSE) and mean
absolute error (MAE) to assess our model and compare it with other models.

4.1 Performance comparison between the knowledge-supported model and other

classical models

After the model learning the experimental data, we obtained the structure parameters of the
model, in terms of the type and numbers of membership functions (MF), as shown in Table 4.
And then, we assessed the model using the testing dataset. Meanwhile, we compared the
performance of the models between our model and some classical models, involving the linear
regression, BP ANN, and RBF ANN model.

Fig. 6 The sub-models for the primary pattern parameters in the girth direction
19026 Multimedia Tools and Applications (2022) 81:19013–19033

Fig. 7 The sub-models for the secondary pattern parameters in the girth direction

Figure 8 shows the comparing results between various models. Through a point-to-point
comparison with other models, we can easily find that both of the RMSE and MAE for pants waist
girth and hip girth in the proposed model highlighted in red frame (see Fig. 8) are 0, equal to those of
linear regression (LR) model, meaning that the predict accuracy of our proposed model is the same
as the LR model. From the view of MSE (see Fig. 8a), for other pattern parameters, the values of our
models are the lowest compared with other models. For example, for pants length, represented by
blue histogram in Fig. 8a, the value of the proposed model is lower than 2, while all the values of the
other models are larger than 2, indicating that the performance of our model is optimal. From the
view of MAE (see Fig. 8b), the values of our models are also the lowest among the four models.
Hence, it can be drawn to the conclusion that the performance of the knowledge-supported model
proposed is better than the traditional LR model, and the classical nonlinear models.

4.2 Application and verification of the knowledge-supported model in pattern design

To evaluate the model presented in this study, we selected an instance from the testing dataset
randomly, the feature body dimensions of which were as below: stature 171 cm, waist girth 92 cm,
and hip girth 100 cm. First, we chose fabric No. 2 (see Table 1), and defined the fitting degree with
Multimedia Tools and Applications (2022) 81:19013–19033 19027

Table 3 Parts of the learning data and testing data for fabric 1 with loose fit (Unit: cm)

No. Data Type Input Data Target Output Data

ST HWG HHG PL CL KL WG HG CG KG
1 Learning Data 164.85 80.93 90.30 99.34 27.21 61.28 82.93 100.30 62.50 47.00
2 169.45 87.95 97.40 106.83 31.22 67.03 89.95 107.40 69.20 48.80
3 174.00 89.01 99.30 108.61 26.41 65.51 91.01 109.30 68.22 51.40
4 176.00 74.50 91.85 109.43 30.10 67.77 76.50 101.85 62.30 46.60
5 163.64 88.50 91.60 101.34 27.70 62.52 90.50 101.60 63.30 46.40
6 156.31 71.00 83.80 96.95 26.95 59.95 73.00 93.80 62.90 47.10
7 164.11 92.84 96.00 100.20 28.66 62.43 94.84 106.00 68.60 48.20
8 183.21 82.90 90.90 116.41 26.40 69.41 84.90 100.90 66.40 46.20
9 159.50 65.50 85.15 96.25 22.40 57.33 67.50 95.15 63.40 45.60
10 175.45 71.91 91.00 108.00 29.00 66.50 73.91 101.00 65.00 46.50
11 Testing Data 161.65 73.90 87.20 100.43 29.42 62.93 75.90 97.20 63.90 45.30
12 169.44 82.65 90.30 100.65 29.70 63.18 84.65 100.30 66.30 46.60
13 171.05 80.60 90.30 105.53 27.68 64.61 82.60 100.30 63.40 46.40
14 165.21 77.90 85.20 104.45 31.02 65.74 79.90 95.20 58.40 45.20
15 166.31 87.12 91.60 100.83 30.56 63.70 89.12 101.60 64.80 47.50
16 175.57 80.30 94.50 108.34 29.11 66.73 82.30 104.50 67.40 47.20
17 171.85 68.95 88.95 105.81 28.46 65.14 70.95 98.95 61.90 45.40
18 167.00 75.20 92.40 104.31 31.76 66.04 77.20 102.40 62.80 47.60
19 168.51 73.60 91.00 103.95 27.40 63.68 75.60 101.00 66.60 46.20
20 160.65 80.21 95.51 101.45 29.30 67.38 82.21 105.51 69.80 48.40

Note: ST, HWG, and HHG represent human stature, waist girth and hip girth, respectively; PL, CL, KL, WG,
HG, CG, and KG refer to the pants length, crotch length, knee length, waist girth, hip girth, crotch girth, and knee
girth, respectively

ease allowance II (see Table 2). Then, two pants patterns of the instance were designed, using the
parameters from the knowledge-supported model and the experienced expert, respectively. Finally,
we evaluated the patterns by the actual try-on technique. Figure 9 illustrates the comparison of
actual try-on effects between the two pants. Pant A and B represent the pants designed following the
expert and the knowledge-supported model, respectively. Furtherly, we measured the garment ease
values at the randomly selected positions under the condition of actual draping. The results are
shown in Table 5. From Table 5, it can easily be seen that the ease values of the two pants at the
selected position are rather close, indicating that the proposed model is applicable.

4.3 Collaborative garment design process based on the proposed

knowledge-supported garment pattern design model

A new collaborative garment design process can be developed by integrating the proposed
knowledge-supported garment pattern model into a commercial 3D clothing design software.

Table 4 Structural parameters of the knowledge-supported model

Model Inputs Outputs MF type Numbers of MF

MF1 MF2

Sub-model 1 Stature Pants length gbell 5 /

Sub-model 2 Stature Pants crotch length trap 7 /
Sub-model 3 Stature Pants knee length tri 7 /
Sub-model 4 Waist girth Pants waist girth tri 5 /
Sub-model 5 Hip girth Pants hip girth tri 5 /
Sub-model 6 Waist girth, hip girth Pants crotch girth tri 4 3
Sub-model 7 Waist girth, hip girth Pants knee girth gauss 3 3
19028 Multimedia Tools and Applications (2022) 81:19013–19033

Fig. 8 Performance comparing results between various models

Multimedia Tools and Applications (2022) 81:19013–19033 19029

Fig. 9 Comparison of actual try-on effects

The general scheme of the knowledge-supported collaborative garment design process is

illustrated in Fig. 10. Concretely, the process which mainly includes five phases is described
as follows:

1. Acquiring the information from the target consumers, involving anthropometric data,
garment styles (such as shirt, pants, suits, and skirts), personal preference (such as fabrics,
color, and fitness), and range of use (such as time, places and occasions).

Table 5 Comparison of draping ease. (Unit: cm)

Pants A Pants B

Ankle 13.6 14.2

Calf 4.2 3.8
Knee 4 3.8
Middle thigh 4.8 5.2
19030 Multimedia Tools and Applications (2022) 81:19013–19033

Fig. 10 General scheme of the knowledge-supported collaborative garment design process

2. Designing the initial garment patterns based on the parameters generated from the
patternmakers’ knowledge-supported model by learning the experimental data.
3. Evaluating the 3D effects of the initial patterns through virtual try-on in a 3D virtual
environment in a collaborative way between the consumers and patternmakers. If the 3D
effects satisfy the consumers, then jump to phase 5 directly; If not, then go to phase 4.
4. Adjusting the patterns until facilitating the needs of the consumers in a manner of human-
computer interaction design, called “What you see is what you get” [20].
5. Manufacturing the real garments that satisfy both the consumers and pattern designers
based on the patterns well-adjusted in phase 4.

5 Conclusions and future works

In this study, we put forth a knowledge-supported garment pattern design approach by

exploiting knowledgeable and experienced patternmakers’ know-how based on fuzzy logic
and artificial neural networks. The fuzzy logic and artificial neural networks techniques have
been employed to learning the experienced and skilled patternmakers’ knowledge, and then
model the relationships between the feature body dimensions and key pattern parameters
following the garment styles, taking the properties of fabrics, and fitting degree into account.
The general principles of the present approach can be externed to all types of the garments. For
simplicity, to demonstrate the effectiveness of the proposed model, we took pants considered
as a universal clothing style in men’s daily life as research focus. The experimental results
showed that the overall performance of our model is better than the traditional models, such as
Multimedia Tools and Applications (2022) 81:19013–19033 19031

linear regression model, BP ANN and RBF ANN. The major contributions of this study can be
summarized as follows: (1) The proposed approach can provide personalized garments more
fitting and efficient, promoting the implementation of mass customization; (2) The model
proposed not only can help the patternmakers without adequate experiences in patternmaking
to develop personalized garment patterns, but also enable the enterprises to control the product
qualities avoiding influenced by the experiences and skills of the patternmakers; (3) Enter-
prises using the present model can release the strong dependencies of experienced
patternmakers, reducing the risks by the turnover of experienced experts; (4) A new collab-
orative garment design process can be formed by combining the proposed knowledge-
supported model with a commercial 3D garment design software, aiming at facilitating the
personalized requirements of the consumers more intuitively, accurately and efficiently.
Furthermore, the performance of the model put forth can be enhanced progressively by adding
new knowledge from the continuous implementation of the model.
Due to garment patterns are business secrets for the clothing enterprises, the dataset we can
obtained in this study is limited. Thus, more datasets should be collected from more companies
to train the proposed model in the future. Meanwhile, as the expansion of the learning datasets,
more advanced artificial neural networks like attention-based modality-gated networks, should
be considered and utilized. Moreover, we would explore the development a collaborative
personalized garment design decision support system by integrating the knowledge-supported
model in future research.

Acknowledgements The authors wish to acknowledge the financial support of the Key Research Project of
Humanities and Social Sciences in Anhui Province College (No. SK2016A0116 and SK2017A0119), the
Fundamental Research Funds for the Central Universities and Graduate Student Innovation Fund of Donghua
University (No. CUSF-DH-D-2020091), the Open Project Program of Key Laboratory of Silk Culture Heritage
and Products Design Digital Technology of Ministry of Culture and Tourism of China (No. 2020WLB07), the
European H2020 Research Program (Project: FBD_BModel, No. 761122), the Special Excellent Ph.D. Interna-
tional Visit Program of DHU, the Open Project Program of Anhui Province College Key Laboratory of Textile
Fabrics, Anhui Engineering and Technology Research Center of Textile (No. 2018AKLTF15), the Social
Science Planning Project in Anhui (No. AHSKQ2019D085), and the National Key Research and Development
Program of China (No. 2019YFF0302100).

References

1. Bin J, Gardiner B, Liu Z, Li E (2019) Attention-based multi-modal fusion for improved real estate appraisal:
a case study in Los Angeles. Multimed Tools Appl 78(22):31163–31184. https://doi.org/10.1007/s11042-
019-07895-5
2. Bruniaux P, Cichocka A, Frydrych I (2016) 3D digital methods of clothing creation for disabled people.
Fibres Text East Eur 5(119):125–131. https://doi.org/10.5604/12303666.1215537
3. Chan AP, Fan J, Yu WM (2005) Prediction of men’s shirt pattern based on 3D body measurements. Int J
Cloth Sci Tech 17(2):100–108. https://doi.org/10.1108/09556220510581245
4. Chen Y, Zeng X, Happiette M, Bruniaux P, Ng R, Yu W (2008) A new method of ease allowance
generation for personalization of garment design. Int J Cloth Sci Tech 20(3):161–173
5. Chen Y, Zeng X, Happiette M, Bruniaux P, Ng R, Yu W (2009) Optimisation of garment design using
fuzzy logic and sensory evaluation techniques. Eng Appl Artif Intell 22(2):272–282. https://doi.org/10.1016
/j.engappai.2008.05.007
6. Demir M, Nasibov E, Vahaplar A (2019) A fuzzy logic apparel size decision methodology for online
marketing. Int J Cloth Sci Tech 31:299–315. https://doi.org/10.1108/IJCST-06-2018-0077
7. Dong M, Zeng X, Koehl L, Zhang J (2020) An interactive knowledge-based recommender system for
fashion product design in the big data environment. Inf Sci 540:469–488. https://doi.org/10.1016/j.
ins.2020.05.094
19032 Multimedia Tools and Applications (2022) 81:19013–19033

8. Foundation EM (2017) A new textiles economy: redesigning fashion's future. https://www.

ellenmacarthurfoundation.org/publications/a-new-textiles-economy-redesigning-fashions-future
9. Hong Y, Bruniaux P, Zeng X, Curteza A, Liu K (2017) Design and evaluation of personalized garment
block for atypical morphology using the knowledge-supported virtual simulation method. Text Res J
88(15):1721–1734. https://doi.org/10.1177/0040517517708537
10. Hong Y, Bruniaux P, Zeng X, Liu K, Chen Y, Dong M (2017) Virtual reality-based collaborative design
method for designing customized garment for disabled people with scoliosis. Int J Cloth Sci Tech 29(2):
226–237. https://doi.org/10.1108/IJCST-07-2016-0077
11. Hong Y, Liu K, Bruniaux P, Dong M, Zhang J, Chen Y (2018) Application of 3D-TO-2D garment design
for atypical morphology: a design case for physically disabled people with scoliosis. Industria Textila 69(1):
59–64. https://doi.org/10.35530/IT.069.01.1377
12. Hong Y, Zeng X, Bruniaux P, Liu K (2016) Interactive virtual try-on based three-dimensional garment
block design for disabled people of scoliosis type. Text Res J 87(10):1261–1274. https://doi.org/10.1177
/0040517516651105
13. Hu Z-H, Ding Y-S, Yu X-K, Zhang W-B, Qiao Y (2009) A hybrid neural network and immune algorithm
approach for fit garment design. Text Res J 79(14):1319–1330. https://doi.org/10.1177/0040517508100726
14. Huang F, Wei K, Weng J, Li Z (2020) Attention-based modality-gated networks for image-text sentiment
analysis. ACM Trans Multimed Comput Commun Appl 16(3):article 79. https://doi.org/10.1145/3388861
15. Ling X, Hong Y, Pan Z (2020) Development of a dress design knowledge base (DDKB) based on sensory
evaluation and fuzzy logic. International Journal of Clothing Science and Technology. https://doi.
org/10.1108/IJCST-02-2020-0016
16. Ling H, Wu J, Huang J, Chen J, Li P (2020) Attention-based convolutional neural network for deep face
recognition. Multimed Tools Appl 79(9):5595–5616. https://doi.org/10.1007/s11042-019-08422-2
17. Liu K, Wang J, Kamalha E, Li V, Zeng X (2017) Construction of a prediction model for body dimensions
used in garment pattern making based on anthropometric data learning. J Text Inst 108(12):2107–2114.
https://doi.org/10.1080/00405000.2017.1315794
18. Liu K, Wang J, Tao X, Zeng X, Bruniaux P, Kamalha E (2016) Fuzzy classification of young women’s
lower body based on anthropometric measurement. Int J Ind Ergon 55:60–68. https://doi.org/10.1016/j.
ergon.2016.07.008
19. Liu K, Wang J, Zhu C, Yan H (2016) Development of upper cycling clothes using 3D-to-2D flattening
technology and evaluation of dynamic wear comfort from the aspect of clothing pressure. Int J Cloth Sci
Tech 28(6):736–749. https://doi.org/10.1108/IJCST-02-2016-0016
20. Liu K, Zeng X, Bruniaux P, Tao X, Yao X, Li V, Wang J (2018) 3D interactive garment pattern-making
technology. Comput Aided Des 104:113–124. https://doi.org/10.1016/j.cad.2018.07.003
21. Liu K, Zeng X, Tao X, Bruniaux P (2019) Associate Design of Fashion Sketch and Pattern. IEEE Access 7:
48830–48837. https://doi.org/10.1109/ACCESS.2019.2906261
22. Liu Y-J, Zhang D-L, Yuen MM-F (2010) A survey on CAD methods in 3D garment design. Comput Ind
61(6):576–593. https://doi.org/10.1016/j.compind.2010.03.007
23. Liu K, Zhu C, Tao X, Bruniaux P, Zeng X (2019) Parametric design of garment pattern based on body
dimensions. Int J Ind Ergon 72:212–221. https://doi.org/10.1016/j.ergon.2019.05.012
24. Lunscher N, Zelek J (2018) Deep learning whole body point cloud scans from a single depth map. 2018
IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW):1208–12087.
https://doi.org/10.1109/CVPRW.2018.00157
25. Mas’ud A, Ardila-Rey J, Albarracín R, Muhammad-Sukki F, Bani N (2017) Comparison of the perfor-
mance of artificial neural networks and fuzzy logic for recognizing different partial discharge sources.
Energies 10:1060. https://doi.org/10.3390/en10071060
26. Nasibov E, Demir M, Vahaplar A (2019) A fuzzy logic apparel size decision methodology for online
marketing. Int J Cloth Sci Tech 31(2):299–315. https://doi.org/10.1108/IJCST-06-2018-0077
27. Song D, Li T, Mao Z, Liu A-A (2019) SP-VITON: shape-preserving image-based virtual try-on network.
Multimed Tools Appl. https://doi.org/10.1007/s11042-019-08363-w
28. Tao X, Bruniaux P (2013) Toward advanced three-dimensional modeling of garment prototype from
draping technique. Int J Cloth Sci Tech 25(4):266–283. https://doi.org/10.1108/09556221311326301
29. Tao X, Chen X, Zeng X, Koehl L (2018) A customized garment collaborative design process by using
virtual reality and sensory evaluation on garment fit. Comput Ind Eng 115:683–695. https://doi.org/10.1016
/j.cie.2017.10.023
30. Thomassey S, Bruniaux P (2013) A template of ease allowance for garments based on a 3D reverse
methodology. Int J Ind Ergon 43(5):406–416. https://doi.org/10.1016/j.ergon.2013.08.002
31. Wang Z, Wang J, Xing Y, Yang Y, Liu K (2019) Estimating human body dimensions using RBF artificial
neural networks technology and its application in Activewear pattern making. Appl Sci 9(6). https://doi.
org/10.3390/app9061140
Multimedia Tools and Applications (2022) 81:19013–19033 19033

32. Xing Y, Wang Z, Li T, Ye H (2014) An innovative approach for auto-generating the sleeve pattern sizes by
artificial neural network model using MATLAB. Textile Bioengineering and Informatics Symposium in
conjunction with Asian Protective Clothing Conference (TBIS-APCC 2014): 667–674
33. Zeng X, Koehl L (2003) Representation of the subjective evaluation of the fabric hand using fuzzy
techniques. Int J Intell Syst 18(3):355–366. https://doi.org/10.1002/int.10092
34. Zeng X, Ruan D, Koehl L (2008) Intelligent sensory evaluation: concepts, implementations, and applica-
tions. Math Comput Simul 77(5):443–452. https://doi.org/10.1016/j.matcom.2007.11.013
35. Zhang J, Liu K, Dong M, Yuan H, Zhu C, Zeng X (2020) An intelligent garment recommendation system
based on fuzzy techniques. J Text Inst 111(9):1324–1330. https://doi.org/10.1080/00405000.2019.1694351
36. Zhang J, Zeng X, Liu K, Yan H, Dong M (2018) Jeans knowledge base development based on sensory
evaluation technology for customers’ personalized recommendation. Int J Cloth Sci Tech 30(1):101–111.
https://doi.org/10.1108/IJCST-03-2017-0036
37. Zhu X-j LH, Rätsch M (2018) An interactive clothing design and personalized virtual display system.
Multimed Tools Appl 77(20):27163–27179. https://doi.org/10.1007/s11042-018-5912-x

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

Affiliations

Zhujun Wang 1,2,3,5 & Yingmei Xing 1 & Jianping Wang 2,5 & Xianyi Zeng 4 & Yalan Yang 2 &
Shuo Xu 2
1
College of Textiles and Garments, Anhui Polytechnic University, Wuhu 241000 Anhui, China
2
College of Fashion and Design, Donghua University, Shanghai 200051, China
3
Key Laboratory of Silk Culture Heritage and Products Design Digital Technology, Ministry of Culture and
Tourism, Zhejiang 310018 Hangzhou, China
4
GEMTEX Laboratory, Ecole Nationale Superieure des Arts et Industries Textiles, 59056 Roubaix, France
5
Key Laboratory of Clothing Design & Technology, Donghua University, Ministry of Education,
Shanghai 200051, China