Prediction of garment fit level in 3D virtual environment based on artificial neural networks

Abstract

This paper proposes a probabilistic neural network-based model for predicting and controlling garment fit levels from garment ease allowances, digital pressures, and fabric mechanical properties measured in a three-dimensional (3D) virtual environment. The predicted fit levels include both comprehensive and local fit levels. The model was set up by learning from data measured during a series of virtual (input data) and real try-on (output data) experiments and then simulated to predict different garment styles, for example, loose and tight fits. Finally, the performance of the proposed model was compared with the Linear Regression model, the Support Vector Machine model, the Radial Basis Function Artificial Neural Network model, and the Back Propagation Artificial Neural Network model. The results of the comparison revealed that the prediction accuracy of the proposed model was superior to those of the other models. Furthermore, we put forward a new interactive garment design process in a 3D virtual environment based on the proposed model. Based on interactions between real pattern adjustments and virtual garment demonstrations, this new design process will enable designers to rapidly, accurately, and automatically predict relevant garment fit levels without undertaking expensive and time-consuming real try-ons.

Keywords

Garment computer-aided design (CAD)garment fit prediction probabilistic neural networks ease allowance digital clothing pressure

In the context of fierce international competition in the textile/clothing industry, adaptation and development of innovative technologies (internet of things, big data, artificial intelligence (AI), etc.) have become key to the success of most industrial companies. One important trend for textile/apparel innovations is mass customization, permitting the delivery of relevant, small-series garment products that meet consumers’ personalized fashion and functional requirements and their body shapes. Moreover, instead of brick-and-mortar stores, people in all countries have become increasingly familiar with online purchasing of garments in e-shops, due to several advantages such as convenience, variety, good discounts, and time-savings.¹ E-shopping transactions accounted for 14.1% of all retails transactions in 2019, and this figure is expected to rise to 22% in 2023.² In this context, providing data-based intelligent services to help professional designers to interact with online consumers for their personalized design will be significant for enhancing the quality of e-shopping platforms and creating a “business to business to consumer” business model, as well as supporting and affecting consumers’ purchase decisions. One of the fundamental factors influencing purchase decisions is garment fit. No matter how amazing the appearance of the garment, if the garment fit is poor, this will not lead to a purchase.

In the clothing industry, fit is defined as how the garment conforms to the body and how it appears on the body.³ Compared with other design attributes, such as fashion styles, silhouettes, fabrics, and colors, fit is one of the most significant purchasing criteria for consumers. Unfortunately, the return rate (close to 30%) due to poor garment fit remains high in garment e-shopping.^4,5 This situation requires companies to process the returned garments, incurring additional time and cost.⁶ A report highlighted that over 50% of consumers admitted that finding the correct size is one of the biggest challenges when shopping for apparel.⁷

When shopping at brick-and-mortar stores, the majority of consumers aim to find well-fitting garments through multiple actual try-ons. However, in an e-shop, finding the garment that best fits a specific consumer is very difficult due to the inability to try on a garment. Consumers can estimate fit using their previous experiences, however, this is not reliable because different brands typically design apparel to meet the needs of their respective target populations, potentially using different sizing systems, which may lead to a poor-fitting garment. To address this issue, over the last few years, commercially available three-dimensional (3D) virtual try-on software systems, such as Lectra 3D prototype, Clo 3D, 3D Runway Creator, VStitcher, and Vidya, have been developed for garment design and fit evaluation.^8-9 These systems allow the simulation of real garment designs and the creation of new 3D garment products by integrating two-dimensional (2D) pattern design, 3D parametric mannequin modeling, 3D virtual sewing, and fabric properties, based on sophisticated geometric and mechanical modeling and simulation techniques. They provide the opportunity for remote consumers or professionals to visualize, communicate, and evaluate virtual garments through a web platform by showing both static and dynamic draping and fitting effects on a specific 3D digital mannequin and selected fabrics.^10,11 Therefore, efficient evaluation of the fitting effects of 3D garment virtual try-on is a crucial and challenging issue to be solved. This will enable optimization of the design process by predicting and controlling garment fit levels precisely and automatically.

Currently, the usual approach for garment fit evaluation is visual analysis implemented by expert know-how based on photographs or videos.^12-14 However, the limitations in relation to accuracy and reliability of this subjective method are obvious. Thus, several researchers have attempted to utilize advanced technologies in this regard, such as 3D human body scanning.⁵ First, scans of body shapes with and without garments are constructed. These are then superimposed on the same coordinate system. Sequentially, the 2D and 3D parameters, such as linear distance, contact area, and volume of the air gaps between the body and clothed scan are calculated precisely. Therefore, garment fit can be analyzed quantitatively using these objective indicators. However, this method has three main shortcomings. The first is cost, due to the requirement for expensive equipment; the second is the complexity of operation; the third is that the process is time-consuming. These three factors have hindered its application in online apparel shopping. In e-shopping scenarios, consumers commonly make purchase decisions from numerous garments in a short time. Thus, it is essential to develop an approach that can predict garment fit level rapidly, precisely, and automatically. Liu et al. proposed a method of automatic garment fit evaluation via learning the digital pressure data.¹⁵ This work enabled prediction of garment fit based on virtual try-on for e-shopping. However, the indicator, namely digital clothing pressure, used for fit evaluation could not accurately predict the fit level of loose garments, as some parts of a loose garment are not in contact with the human body and the clothing pressure may be close to zero.¹⁵ Aside from digital clothing pressure, fashion designers conventionally evaluate garment fit based on ease allowance, which refers to a metric indicating the dimensional difference between the human body and garment.¹⁶ However, typical ease allowances may not reflect the fit levels produced by different fabrics. Furthermore, garment styles are diverse and ever-changing: neither digital pressure nor typical ease allowance are suitable for assessing the fit levels across all kinds of garments. In this context, we propose a machine learning (ML)-based garment fit level prediction model (see Figure 1) aimed at predicting and controlling garment local and comprehensive fit levels via learning data from garment ease allowance, digital pressure, and the mechanical properties of fabric measured in a 3D virtual environment.

Figure 1.

Machine learning-based garment fit level prediction model.

Currently, AI-based ML techniques, such as artificial neural networks (ANN), fuzzy logic, and genetic algorithms, have been intensively applied in the fashion industry^17-19 in fashion product design,^8-9,20 garment and fabric product evaluation,^21-25 fashion forecasting,²⁶ garment comfort estimation,^27,28 apparel manufacturing,^29-31 fashion retailing,^32-34 and garment supply chain management.^35,36 However, little attention has been paid to predicting garment fit level based on the virtual try-on using an ANN technique.

The probabilistic neural network (PNN), originally proposed by Specht, is a particular type of ANN model based on the well-known Bayesian decision rules.^37,38 PNN is acknowledged to be a powerful tool for tackling pattern recognition and classification problems in the textile and apparel sector,³⁹ including textile classification and evaluation,^40-42 textile defect detection,^43,44 fabric color grading,⁴⁵ pattern simulation,⁴⁶ machine fault diagnosis,⁴⁷ human body shape recognition,⁴⁸ garment recommendation systems,⁴⁹ and manufacturing plant location selection.⁵⁰ The advantages of PNN are as follows: 1. the PNN model has an inherently parallel structure, which guarantees convergence to an optimal classifier with increasing size of representative training dataset; 2. the number of parameters for constructing and optimizing a PNN model are relatively few; 3. the PNN is not trained iteratively, which can shorten the learning process dramatically; 4. the learning dataset for a PNN can be supplemented and removed without extensive retraining; and 5. the PNN is less sensitive to noisy data from outliers, which can harm several other types of ANN. Contributing to these advantages, we adopted PNN to predict garment fit level in this work.

The sections of this paper are organized as follows. In Section 2, the general research scheme and implementation procedure are introduced. In Section 3, we describe how to procure the learning data for constructing the garment fit level model in detail. In Section 4, we discuss how to establish a garment fit level prediction model for both loose- and tight-fitting garments. In Section 5, we analyze and evaluate the performance of the proposed model. Finally, the conclusions are presented in Section 6.

General research scheme and formalization

General research scheme

The process for garment fit level prediction is composed of four sequential phases (see Figure 2). In Phase I, we collected the output learning data (i.e., sensory evaluations on garment fit) by performing a series of real try-on experiments, as well as human body data. In Phase II, we acquired the input learning data, including the garment ease allowance values, digital pressures, and fabric properties measured in a 3D virtual environment by performing a series of virtual garment try-ons. In Phase III, a PNN-based garment fit level prediction model was set up and trained using the input and output learning data. In Phase IV, we applied and verified the proposed model to predict garment fit level.

Figure 2.

General scheme.

Formalization

The concepts and data involved in this paper were formalized as follows:

Let $F = \{f_{1}, f_{2}, \dots, f_{n}\}$ be the set of fabrics involved in this study.

Let ${F P}_{i} = \{{f p}_{i}^{1}, {f p}_{i}^{2}, \dots, {f p}_{i}^{q}\}$ be a vector of the fabric mechanical properties of the fabric, $m_{i},$ measured in the 3D virtual environment, where ${f p}_{i}^{q}$ is the fabric mechanical properties, q, of the fabric $f_{i} (i \in \{1, \dots, n\})$ .

Let $G_{i} = \{g_{i}^{1}, g_{i}^{2}, \dots, g_{i}^{m}\}$ be the set of actual garments used in this study, where $g_{i}^{m}$ is the m sized garment in the fabric $f_{i}$ .

Let $L (g_{i}^{j}) = \{l^{1} (g_{i}^{j}), l^{2} (g_{i}^{j}), \dots, l^{k} (g_{i}^{j})\}$ be the set of feature lines on the patterns of the garment $g_{i}^{j} (i \in \{1, \dots n,\}, j \in \{1, \dots, m\})$ , where $l^{k} (g_{i}^{j})$ is the feature line, k, on the patterns of the j sized garment in the fabric $f_{i}$ .

Let $K P (g_{i}^{j}) = \{{k p}^{1} (g_{i}^{j}), {k p}^{2} (g_{i}^{j}), \dots, {k p}^{k} (g_{i}^{j})\}$ be the set of key positions on the patterns of the garment $g_{i}^{j} (i \in \{1, \dots n,\}, j \in \{1, \dots, m\})$ , where ${k p}^{k} (g_{i}^{j})$ is the key position, k, on the patterns of the j sized garment in the fabric $f_{i}$ .

Let $S = \{s_{1}, s_{2}, \dots, s_{p}\}$ be the set of subjects involved in this study.

Let $D P (g_{i}^{j}, s_{l}) = \{{d p}^{1} (g_{i}^{j}, s_{l}), {d p}^{2} (g_{i}^{j}, s_{l}), \dots, {d p}^{k} (g_{i}^{j}, s_{})\}$ be the set of digital pressures of the garment, $g_{i}^{j},$ worn by the subject, $s_{l} (l \in \{1, \dots, p\}),$ where ${d p}^{k} (g_{i}^{j}, s_{l})$ is the digital pressure at key position k of the j sized garment in the fabric, $f_{i}$ , worn by the subject $s_{l}$ .

Let ${B P}_{l} = \{b p_{l}^{1}, {b p}_{l}^{2}, \dots, {b p}_{l}^{k}\}$ be the set of feature body positions in relation to the body of the subject, $s_{i}$ , where ${b p}_{i}^{k}$ is the feature body position, k, of the subject, $s_{l}$ , in which each ${b p}_{i}^{k}$ corresponds to one $l^{k} (g_{i}^{j})$ or ${k p}^{k} (g_{i}^{j})$ .

Let $L M (g_{i}^{j}, s_{l}) = \{{l m}^{1} (g_{i}^{j}, s_{l}), {l m}^{2} (g_{i}^{j}, s_{l}), \dots, {l m}^{k} (g_{i}^{j}, s_{l})\}$ be the set of feature line measurements of the garment, $g_{i}^{j}$ , worn by subject, $s_{l},$ in the 3D digital environment, where ${l m}^{k} (g_{i}^{j}, s_{l})$ is the measurement of the feature line, k, of the j sized garment in the fabric, $f_{i}$ , worn by subject $s_{l}$ .

Let ${B M}_{l} = \{{b m}_{l}^{1}, {b m}_{l}^{2}, \dots, {b m}_{l}^{k}\}$ be the set of human body measurements at the feature body positions of the subject $s_{l}$ , where ${b m}_{l}^{k}$ is the body measurements at feature body position k of the subject, $s_{l}$ , in which each ${b m}_{l}^{k}$ corresponds to one ${l m}^{k} (g_{i}^{j}, s_{l})$ .

Let $E (g_{i}^{j}, s_{l}) = \{e^{1} (g_{i}^{j}, s_{l}), e^{2} (g_{i}^{j}, s_{l}), \dots, e^{k} (g_{i}^{j}, s_{l})\}$ be the set of the ease allowances of the garments, $g_{i}^{j}$ , worn by subject, $s_{l},$ in the 3D digital environment, where $e^{k} (g_{i}^{j}, s_{l})$ is calculated following equation (1).

e^{k} (g_{i}^{j}, s_{l}) = {l m}^{k} (g_{i}^{j}, s_{l}) - {b m}_{l}^{k}, i \in \{1, \dots n,\}, j \in \{1, \dots, m\}, l \in \{1, \dots, p\}

(1)

Let $S D (g_{i}^{j}) = \{{s d}^{1} (g_{i}^{j}), {s d}^{2} (g_{i}^{j}), \dots, {s d}^{k} (g_{i}^{j})\}$ be the set of sensory descriptors of the garment, $g_{i}^{j}$ , where ${s d}^{k} (g_{i}^{j})$ is the sensory descriptor of the feature position k of the j sized garment in the fabric $f_{i}$ .

Let $EVA (g_{i}^{j}, s_{l}) = \{{eva}^{1} (g_{i}^{j}, s_{l}), {eva}^{2} (g_{i}^{j}, s_{l}), \dots, {eva}^{k} (g_{i}^{j}, s_{l})\}$ be the set of fit level values of the sensory descriptors of the garment, $g_{i}^{j}$ , given by the subject, $s_{l}$ , where ${eva}^{k} (g_{i}^{j}, s_{l})$ is the fit level value of the sensory descriptor of the feature position k of the j sized garment in the fabric, $f_{i}$ , given by the subject $s_{l}$ .

Let $pEVA (g_{i}^{j}, s_{l}) = \{{peva}^{1} (g_{i}^{j}, s_{l}), {peva}^{2} (g_{i}^{j}, s_{l}), \dots, {peva}^{k} (g_{i}^{j}, s_{l})\}$ be the set of fit level values of the garment, $g_{i}^{j}$ , calculated by the proposed models, where ${peva}^{k} (g_{i}^{j}, s_{l})$ is the predictive fit level value of the key position k of the j sized garment in the fabric, $f_{i}$ , worn by the subject $s_{l}$ .

Let $W = \{w^{1}, w^{2}, \dots, w^{k}\}$ be the set of weight values of the feature positions of the garments.

Let $F L = \{1, 2, 3, 4, 5\}$ be the set of the values of the garment fit level.

Acquisition of the learning data

The main objective of this study was to put forward a garment fit level prediction model. This work was supported by The European Horizon 2020 (H2020) research program. The international fashion brand company involved in this program is oriented to the Chinese market. Thus, we have described our approach in a scenario of consumers in China. However, the general principles of our proposed method would be suitable for consumers in different countries and regions.

Experiment preparation

Subject recruitment

In this study, a sensory experiment based on real try-on was undertaken to quantitatively characterize the garment fit level according to different garment positions. Considering the requirements of the fashion brand company involved in H2020, such as their target markets, populations, regional distribution, and job distribution, we initially recruited 200 males aged 18–35 randomly. Then, 17 male subjects with the representative figure types were selected and invited to join this experiment. Their body sizes were 155/76A, 155/80A, 155/84A, 160/80A, 160/84A, 160/88A, 165/84A, 165/88A, 165/92A, 170/84A, 170/88A, 170/92A, 175/84A, 175/88A, 175/92A, 180/88A, and 185/92A, respectively. The notation “170/88A” indicates that the figure type is A, and the height and bust girth are 170 cm and 88 cm, respectively. These sizes account for the vast majority of the population of China, according to the China National Standard (GB/T 1335.1-2008).⁵¹

To guarantee precision, we performed the experiments as follows. Firstly, we invited a professional with over 20 years’ experience studying and developing sizing systems to participate in our research. With her in-depth knowledge and rich experience, we were able to guarantee precision of the 17 male subjects, following the China National Standard.

Secondly, considering the target population of the fashion brand company, based on this standard, we chose Type A as our research focus due to its high prevalence in Chinese society.

Thirdly, having searched for the national standard, we determined height and bust girth ranges. Height ranged from 155 to 185, and the bust girth ranged from 76 to 92. The height and bust girth intervals were 5 and 4 cm, respectively. Based on these data, we determined the sizes and the combinations involved in our research.

Fourthly, we strictly excluded subjects with the specific body size, by comparing the detailed body measurements (such as neck girth, arm length, back length, height of the cervical point when standing, and the height of cervical point when sitting, and so forth) with the corresponding body measurements defined by the China National Standard.

Therefore, the body dimensions of the selected subjects were exactly in line with the China National Standard as far as possible.

Anthropometric measurement and 3D mannequin construction

To measure the anthropometric data precisely and expeditiously, we collected the anthropometric measurements of the subjects using Vitus Bodyscan (VITRONIC Corp, Wiesbaden, Germany). Next, the 3D mannequins (see Figure 3) corresponding to the subjects were generated using CLO 3D software.

Figure 3.

The mannequins used for virtual try-on created from the 17 subjects.

Table 1.

Key pattern parameters of shirts (cm)

Shirt size	Garment length	Bust girth	Collar girth	Shoulder width	Sleeve length	Waist girth	Hip girth
38	70	98	38	43.4	60	92	94
40	72	102	39	44.6	60.5	96	98
42	74	106	40	45.8	60.5	100	102
44	76	110	41	47	60.5	104	106
46	78	114	42	48.2	61	108	110

Table 2.

Key pattern parameters of leggings (cm)

Legging size	Legging length	Waist girth	Hip girth	Knee girth	Calf girth
XS	82.5	58	76	27	25
S	84	62	80	28	26
M	85.5	66	84	29	27
L	87	70	88	30	28
XL	88.5	74	92	31	29

Table 3.

Fabric information

	Type	Compositions	TH (cm)	SW1	SW2	SH	BW1	BW2	BR1	BR2	BS1	BS2	DE
Shirt fabric 1	Woven	100% cotton	0.23	55	58	14	23	29	90	90	30	30	7
Shirt fabric 2	Woven	97% cotton, 3% spandex	0.56	10	15	8	9	10	0	0	0	0	21
Shirt fabric 3	Woven	55% linen, 45% cotton	0.36	27	27	9	38	38	80	80	80	80	27
Shirt fabric 4	Woven	100% cotton	0.6	43	40	7	32	49	90	90	20	60	17
Shirt fabric 5	Woven	100% Silk	0.14	3	3	0	8	10	0	0	80	80	0
Leggings fabric 1	Knit	55% cotton, 45% rayon	0.43	11	12	10	6	15	80	80	20	20	10
Leggings fabric 2	Knit	100% polyester	0.9	37	34	8	12	24	90	90	30	30	19
Leggings fabric 3	Knit	68% viscose, 27% nylon, 5% elastane	0.86	15	15	15	31	18	0	0	80	80	30
Leggings fabric 4	Knit	100% polyester	0.64	46	28	21	43	48	85	85	80	80	13
Leggings fabric 5	Knit	100% polyester	0.5	27	27	9	38	38	80	80	80	80	27

Table 4.

The meaning of the fabric properties measured in the 3D virtual environment

Abbreviation	Meaning	Abbreviation	Meaning
TH	Thickness	BR1	Bending ratio–weft
SW1	Stretch–weft	BR2	Bending ratio–warp
SW2	Stretch–warp	BS1	Buckling stiffness–weft
SH	Shear	BS2	Buckling stiffness–warp
BW1	Bending–weft	DE	Density
BW2	Bending–warp

Experimental garments

Both loose- and tight-fitting garments were taken into account. The shirt is a classic, indispensable garment in men’s daily life, which can be matched with a wide range of garments from formal to leisure wear, such as suits, vests, and jeans. Staying healthy and active has recently increasingly become a concern for public all over the world. The constantly rising number of people engaging in sports activities has driven the growth in consumer demand for activewear. Due to their versatility for use in a wide range of sports, such as jogging, hiking, and cycling, leggings have gained great popularity among consumers. The shirts and leggings used in the research were provided by the international fashion brand company involved in this program. With these products, business validation in the market will be easier. For the above-mentioned reasons, shirts and leggings were selected as representative of loose- and tight-fitting garments respectively in our study. Furthermore, considering the influence of fabric properties on garment fit, we selected five types of fabric with five sizes of each type of experimental garment. Tables 1 to 3 present the key pattern parameters of the shirts and leggings, and the fabric properties, respectively. The meaning of the measured fabric properties in 3D virtual environment is present in Table 4.

Evaluation value scale and semantic description for fit level

K sensory descriptors were chosen to evaluate the garment fit levels on the k feature positions of the garment in both static and dynamic scenarios. The descriptors were evaluated using a semantic differential scale, including five evaluation scores represented by $\{1, 2, 3, 4, 5\}$ , corresponding to the linguistic values ${“ too tight / short (1) ”, “ tight / short (2) ”, “ perfect (3) ”, “ loose / long (4) ”, “ too loose / long (5) ”} .$

Acquisition of the sensory evaluation data on garment fit level

A real try-on-based sensory experiment was carried out to collect the sensory data on garment fit. To reiterate, the objective of this experiment was to acquire sensory evaluation data on garment fit using real try-on. We began the experiment once all the subjects had agreed to participate. The detailed procedure is given in the following.

Step 1: Ten professionals engaged in designing male garment patterns for over 10 years were surveyed to extract the feature lines on the experimental sample patterns, to determine the set of sensory descriptors $S D (g_{i}^{j}) .$

Step 2: In a laboratory with constant temperature ( $20 \pm 2$ °C) and relative humidity ( $65 \pm 5 %$ ), each of the 17 subjects having different figure types performed his try-on with the experimental garments and perceived the fit levels in static and dynamic scenarios, just as they would in brick-and-mortar stores. Meanwhile, the evaluation value, ${eva}^{k} (g_{i}^{j}, s_{l})$ , constituting the fit level dataset, $EVA (g_{i}^{j}, s_{l})$ , was recorded, which were used to formed the output training dataset to construct the proposed model.

Acquisition of the indicator data for predicting garment fit level

Acquisition of the garment ease allowance in 3D virtual environment

A virtual try-on experiment was realized to procure the feature line measurements of the garments in a 3D virtual environment. Due to its cutting-edge simulation techniques and high accuracy rate (>95%) for real garments and fabrics, the software CLO 3D was employed in this experiment. The exact details of the experiment are described in the following text.

Step 1: Seventeen 3D mannequins were generated based on the collected anthropometric data.

Step 2: We determined the feature lines, $L (g_{i}^{j})$ , of the patterns following the questionnaire survey results from the professional patternmakers (see Figure 4(a) and (c)).

Figure 4.

Ease allowances measurement in the 3D digital environment.

Step 3: After setting the fabric properties based on the approach proposed by Chen et al.,²⁵ we created the digital garment on the mannequins. Afterward, we measured the length of the predetermined feature lines on the digital garments, ${l m}^{k} (g_{i}^{j}, s_{l})$ , during its virtual try-on (see Figure 4(b) and (d)).

Step 4: We computed the ease allowances, $E (g_{i}^{j}, s_{l})$ , in the 3D virtual environment following equation (1).

Acquisition of garment digital pressure data in the 3D virtual environment

Step 1: We determined the key positions, $K P (g_{i}^{j})$ , on the legging patterns after performing several pre-experiments, as illustrated in Figure 5(a).

Figure 5.

The measurement of digital pressure of leggings in the 3D digital environment.

Step 2: The garment digital pressures, $D P (g_{i}^{j}, s_{l})$ , at the key positions, $K P (g_{i}^{j})$ , were measured in the 3D virtual environment, as shown in Figure 5(b).

Finally, the garment ease allowances and the digital pressures measured in the 3D virtual environment were taken as elements of the input learning dataset to establish the proposed model.

Construction of the garment fit level prediction model

Framework of the garment fit level prediction model

The proposed model (see Figure 6) is constituted by k sub-models and one comprehensive model. First, the local fit levels were predicted by the k sub-models. The comprehensive level was subsequently aggregated following equation (2),

pCEVA (g_{i}^{j}, s_{l}) = \sum_{r = 1}^{k} w^{r} {peva}^{r} (g_{i}^{j}, s_{l}), i \in \{1, \dots n,\}, j \in \{1, \dots, m\}, r \in \{1, \dots, k\}, l \in \{1, \dots, q\}

(2)

Figure 6.

The general framework of the garment fit level prediction model.

where

pCEVA (g_{i}^{j}, s_{l})

refers to the comprehensive fit level of the j sized garment with the fabric,

f_{i}

, given by the proposed model;

{peva}^{r} (g_{i}^{j}, s_{l})

is the predictive fit level value of the key position r of the j sized garment with the fabric,

f_{i}

, worn by the subject

s_{l}

; and

w^{r}

is the weight values of the feature positions r of the garments.

The sub-model based on PNN is a four-layer feed-forward network (see Figure 7) that was constructed using MATLAB software. The garment fit level prediction procedures for feature positions using the PNN model are given in the following steps.

Figure 7.

Architecture of the garment fit level prediction sub-model for loose garments.

Step 1: The input vector, $x_{i}$ , is distributed to the input layer of the PNN model. The number of input neurons is the same as the dimensions of the input vector. If we take shirt fit level prediction as an example, after a specific consumer selects a certain shirt, we can obtain the ease allowance data for each key body measurement, as well as the fabric properties. When we want to predict the collar fit level, if the collar ease data is 1.5 cm, and the selected fabric is shirt fabric 1, shown in Table 3, then the input vector, $x_{i}$ , of the PNN model can be expressed by

x_{i} = \{\overset{C}{1.5}, \overset{T H}{0.23}, \overset{S W 1}{55}, \overset{S W 2}{58}, \overset{S H}{14}, \overset{B W 1}{23}, \overset{B W 2}{29}, \overset{B R 1}{90}, \overset{B R 2}{90}, \overset{B S 1}{30}, \overset{B S 2}{39}, \overset{D E}{7}\}

where

C

represents the collar ease data;

T H, S W 1, S W 2, S H, B W 1, B W 2, B R 1, B R 2, B S 1, B S 2,

and

D E,

refer to the fabric properties measured in the 3D virtual environment, whose implication can be found in Table 4.

If we need to predict the fit level of the same garment in other body measurements, we only need to change the collar ease data into the corresponding ease data of the required body measurement. If we need to evaluate the fit levels of the garments in other fabrics, we can alter the fabric parameters to those of the specified fabric.

Step 2: In the pattern layer, the relationship between the input vector, $x_{i}$ , and each pattern of the fit level, ${f l}_{l}^{(j)}$ , is calculated by Euclidean distance following equation (3). The neuron output of the pattern layer is then activated by a radial Gauss function like that of equation (4):

{E d}_{i j} = {(x_{i} - {f l}_{l}^{(j)})}^{T} (x_{i} - {f l}_{l}^{(j)})

(3)

{Output}_{i j}^{pattern} = \exp [- \frac{{(x_{i} - {f l}_{l}^{(j)})}^{T} (x_{i} - {f l}_{l}^{(j)})}{2 σ^{2}}]

(4)

where

x_{i}

is the i input vector;

{f l}_{l}^{(j)}

refers to the l pattern of the fit level

{f l}_{j} (j \in [1, 5])

;

{E d}_{i j}

is defined as the Euclidean distance between the input vector

x_{i}

and

{f l}_{l}^{(j)}

; σ, also known as spread, represents the smoothing parameter.

Step 3: In the summation layer, one summation neuron corresponds to one specific fit level, ${f l}_{j}$ . For each summation neuron, the inputs from the pattern layer belonging to the fit level ${f l}_{j}$ are aggregated following equation (5),

S_{i j} = \sum_{i = 1}^{l} {Output}_{i j}^{pattern} = \sum_{i = 1}^{l} \exp [- \frac{{(x_{i} - {f l}_{l}^{(j)})}^{T} (x_{i} - {f l}_{l}^{(j)})}{2 σ^{2}}]

(5)

where, l is the number of samples belonging to the fit level

{f l}_{j}

Since the probability density function (PDF) of the fit level, ${f l}_{j}$ , is usually unknown, the PNN algorithm introduces the Parzen window method to estimate PDF. Based on this method, the class-conditional probability for the i sample belonging to the fit level ${f l}_{j}$ , is computed following equation (6).

{Output}_{j}^{summation} = P (x_{i}| {f l}_{j}) = \frac{S_{i j}}{l}

(6)

Step 4: Based on Bayes theory, in the output layer, the fit level of the i sample is decided by equation (7). This means that the largest $P (x_{i}| {f l}_{j})$ will be identified and the corresponding j will be declared as the category of the input vector $x_{i}$ .

Categroy (x_{i}) = {Output}_{i}^{model} = argmax ({Output}_{j}^{summation}) = argmax (P (x_{i}| {f l}_{j}))

(7)

Construction of the garment fit level prediction model for loose garments

Construction of the sub-models for predicting the local fit level

Step 1: Determining the indicators of predicting the garment fit level for loose garments

The garment ease allowances were utilized as the indicators for predicting the garment fit level of the loose garments. The collected shirts’ ease allowances are shown in Table 5, in which a negative value means the garment measurement is smaller than the corresponding body measurements.

Table 5.

Shirts’ ease allowances collected in the 3D virtual environment (cm)

SN	Collar ease	Shoulder ease	Bust ease	Waist ease	Hip ease	Garment length ease	Sleeve length ease
1	4.54	3.51	21.94	28.58	13.80	29.02	9.02
2	3.52	2.32	17.86	24.58	10.65	28.78	7.60
3	2.45	1.33	13.95	20.60	7.42	26.60	5.97
4	1.42	0.51	10.00	16.62	4.68	26.15	4.59
5	0.42	−0.67	6.08	12.61	4.20	23.94	3.23
6	2.34	−0.40	6.18	12.58	3.70	23.24	1.66
$⋮$	$⋮$	$⋮$	$⋮$	$⋮$	$⋮$	$⋮$	$⋮$
425	3.48	6.98	21.37	28.64	16.92	27.78	1.95

Note: SN: sample number. Due to the length limitation of the article, Table 5 present parts of the shirts' ease allowances, while others were denoted by the symbol of three dots.

Step 2: Determination of the input learning data and the output learning data

The inputs of the proposed model are constituted by the garment ease allowances combined with the specific fabric mechanical properties measured in the 3D virtual environment (see Table 3), while the outputs are the predicted fit levels (see Figure 7).

In this section, we take the construction of the sub-model for predicting collar fit levels as an example. The general principles of creating the sub-model can be applied to set up other sub-models for predicting local fit levels.

For the collar fit levels prediction model, the input layer was composed of one ease datum integrated with 11 fabric properties data, which were collar ease (see Table 5) and fabric parameters TH, SW1, SW2, SH, BW1, BW2, BR1, BR2, BS1, BS2, and DE (see Table 3). The output layer was the predicted collar fit level.

Step 3: Establishing the PNN model for predicting the garment fit level

To evaluate the proposed model close to reality, avoid the problem of over-fitting, and guarantee the prediction accuracy of the proposed model, we utilized the k-fold (K = 10) cross-validation approach in building the collar fit level prediction model. Specifically, the learning dataset was initially randomly split into $K (= 10)$ smaller sets. Next, for each of the $K$ “folds” we created the collar fit level prediction model using the algorithm given in section “Framework of the garment fit level prediction model.”

Meanwhile, $K - 1 (= 9)$ small sets composed of 382 instances were in turn utilized as the model’s training dataset, using the remaining part of the original dataset for validation. Finally, the collar fit level prediction model was determined after 10-fold cross-validation.

Construction of the model for predicting the comprehensive fit level

After the local fit levels are obtained, the weights for every feature position should be determined. In this study, we performed principal components analysis (PCA) on the ease of allowances to decide the weights using SPSS 25 (IBM Corp, Armonk, NY, USA). The general calculation procedure is introduced as follows:

Step 1: Extracting the principal components

Let $G E$ be a matrix of garment ease data. In this step, $G E = (Collar ease, Shoulder ease, Bust ease, Waist ease, Hip ease, Garment length ease, Sleeve length ease)$ . Let $P C = {({p c}^{1}, {p c}^{2}, \dots, {p c}^{k})}^{T}$ be a vector of principal components. The relationship between $P C$ and $G E$ can be expressed by equation (8)

P C = VGE = λ G E

(8)

where

V

is denoted as the covariance matrix; and

λ = \{{eiv}^{1}, {eiv}^{2}, \dots, {eiv}^{k}\}

represents a set of the eigenvalues of the covariance matrix

V

First, we calculated the standardized matrix, $GES$ , following equation (9)

{GES}_{i j} = G E_{i j} - μ_{j} = G E_{i j} - \frac{1}{n} \sum_{i = 1}^{n} {G E}_{i j}

(9)

where

μ_{j}

is the average of j column in matrix

G E

; and n is the number of rows in matrix

G E

Second, we formed the covariance matrix as $V = {GES}^{T} G E S$ . Third, the eigenvalues of matrix $V$ were found based using SPSS.

Table 6 gives the results of the PCA. The factor loadings were rotated using the Varimax method. In our study, the components whose eigenvalues were larger than 1 were selected. From Table 6, the first three components whose cumulative contribution rate was 87.527% meet the criterion. Thus, they were extracted for further research, and the corresponding eigenvalues ${eiv}^{1}, {eiv}^{2}$ , and ${eiv}^{3}$ were 3.681, 1.270, and 1.175, respectively.

Table 6.

Total variance explained

Component	Rotation sums of squared loadings
Component	Eigenvalues	Percentage of variance	Cumulative percentage
1	3.681	52.591	52.591
2	1.270	18.148	70.739
3	1.175	16.788	87.527
4	0.612	8.744	96.271
5	0.256	3.653	99.925
6	0.005	0.067	99.992
7	0.001	0.008	100

Step 2: Calculating the coefficients in linear combinations

Let ${FLR}_{i} = \{{flr}_{i}^{1}, {flr}_{i}^{2}, \dots, {flr}_{i}^{j}\} (i \in [1, 3], j \in [1, 7])$ be a set of rotated factor loadings of the principal components. Let ${CLC}_{i} = \{{clc}_{i}^{1}, {clc}_{i}^{2}, \dots, {clc}_{i}^{j}\} (i \in [1, 3], j \in [1, 7])$ be a set of the coefficients in linear combinations, which can be calculated following equation (10).

{clc}_{i}^{j} = \frac{{flr}_{i}^{j}}{\sqrt{{eiv}^{i}}}

(10)

Step 3: Calculating the un-normalized weights

Let ${VCR}^{i} = \{{vcr}^{1}, {vcr}^{2}, \dots, {vcr}^{i}\} (i \in [1, 3])$ be a set of values of the rotated principal components. Let $U W = \{{u w}^{1}, {u w}^{2}, \dots, {u w}^{j},\} (j \in [1, 7])$ be a set of un-normalized weights, which can be calculated following equation (11).

{u w}^{j} = \frac{\sum_{i = 1}^{3} ({clc}_{i}^{j} \times {vcr}^{i})}{\sum_{i = 1}^{3} {vcr}^{i}}

(11)

Step 4: Calculating the normalized weights

Let $W = \{w^{1}, w^{2}, \dots, w^{j},\} (j \in [1, 7])$ be a set of normalized weights, which can be calculated following equation (12).

w^{j} = \frac{{u w}^{j}}{\sum_{j = 1}^{7} {u w}^{j}}

(12)

Based on equations (10) to (12), we obtained a set of weights for the shirt's feature positions (see Table 7).

Table 7.

The weights of the feature position of the shirts

Feature positions	Garment length (GL)	Sleeve length (SL)	Collar (C)	Bust (B)	Waist (W)	Hip (H)	Shoulder (S)
Weights	0.134	0.105	0.142	0.163	0.164	0.159	0.133

Therefore, we can now obtain the comprehensive fit level of the shirt using equation (13):

pCEVA (g_{i}^{j}, s_{l}) = 0.134 \times {peva}^{G L} (g_{i}^{j}, s_{l}) + 0.105 \times {peva}^{S L} (g_{i}^{j}, s_{l}) + 0.142 \times {peva}^{C} (g_{i}^{j}, s_{l}) + 0.163 \times {peva}^{B} (g_{i}^{j}, s_{l}) + 0.164 \times {peva}^{W} (g_{i}^{j}, s_{l}) + 0.159 \times {peva}^{H} (g_{i}^{j}, s_{l}) + 0.133 \times {peva}^{S} (g_{i}^{j}, s_{l})

(13)

Construction of the garment fit level prediction model for tight garments

We established two types of fit level prediction model for tight-fitting garments. One model is based on the ease allowances only. The other is based on both the digital pressures and ease allowances.

Construction of the garment fit level prediction model for tight garments based on ease allowances

Step 1: Determining the number of sub-models

Since one sub-model corresponds to one feature position, and there are six feature positions that are highly related to the leggings fit: waist, hip, thigh, knee, calf, and the legging length; the leggings’ fit level prediction model is therefore composed of six sub-models.

Step 2: Creating the local fit level prediction models for the leggings

First, the collected dataset (see Table 8) is split into the learning dataset and the testing dataset at a ratio of approximately 9:1. Next, the ease allowances (see Table 8) and the specific fabric properties (see Table 3) are utilized as the inputs of each sub-model, and the local fit level of the leggings is used as the output. Afterward, we established and trained the PNN-based leggings fit level prediction model using the 10-fold cross-validation approach described in section “Construction of the sub-models for predicting the local fit level.”

Table 8.

Leggings’ ease allowances collected in the 3D virtual environment (cm)

SN	Waist ease	Hip ease	Thigh ease	Knee ease	Calf ease	Leggings’ length ease
1	3.83	6.20	4.05	0.39	1.87	−13.50
2	3.40	5.40	3.42	0.42	1.84	−14.30
3	3.38	3.90	3.39	0.33	1.81	−15.00
4	3.60	3.30	3.29	0.35	1.76	−16.00
5	3.60	3.30	2.90	0.19	1.84	−16.57
6	3.60	3.30	3.71	0.13	1.82	−17.85
$⋮$	$⋮$	$⋮$	$⋮$	$⋮$	$⋮$	$⋮$
425	6.01	4.60	0.84	1.33	1.82	−15.75

Note: SN: sample number.

Step 3: Constructing the comprehensive fit level prediction model for the leggings

The weights of the leggings’ feature positions were computed by PCA, as shown in Table 9.

Table 9.

The weights of feature position of the leggings

Feature positions	Leggings length (LL)	Waist (W)	Hip (H)	Thigh (T)	Knee (K)	Calf (C)
Weights	0.212	0.176	0.176	0.158	0.158	0.119

The comprehensive fit level of the leggings can be acquired following equation (14):

pCEVA (g_{i}^{j}, s_{l}) = 0.212 \times {peva}^{L L} (g_{i}^{j}, s_{l}) + 0.176 \times {peva}^{W} (g_{i}^{j}, s_{l}) + 0.176 \times {peva}^{H} (g_{i}^{j}, s_{l}) + 0.176 \times {peva}^{T} (g_{i}^{j}, s_{l}) + 0.158 \times {peva}^{K} (g_{i}^{j}, s_{l}) + 0.119 \times {peva}^{C} (g_{i}^{j}, s_{l})

(14)

Construction of the garment fit level prediction model for tight garments based on digital pressures and ease allowances

Step 1: Constructing the sub-models for predicting the local fit levels based on the digital pressures and garment eases

The inputs of each sub-model were the digital pressures (see Table 10), the ease allowances of the legging length (see Table 7), and the fabric properties (see Table 3). The outputs also included the local fit of the leggings. We then established and trained a PNN-based model to predict local fit level.

Table 10.

Digital pressures for leggings collected in the 3D virtual environment (kpa)

SN	Waist digital pressure	Hip digital pressure	Thigh digital pressure	Knee digital pressure	Calf digital pressure
1	2.92	3.83	2.16	1.26	4.29
2	3.40	4.46	2.91	1.66	5.38
3	5.60	5.75	4.13	4.03	6.31
4	6.75	6.53	4.80	4.50	7.71
5	7.38	6.30	5.25	4.75	8.03
6	7.76	6.51	5.0	4.86	8.0
$⋮$	$⋮$	$⋮$	$⋮$	$⋮$	$⋮$
425	11.1	9.9	9.5	17.8	28.5

Note: SN: sample number.

Step 2: Constructing the comprehensive fit level prediction model for the leggings based on the garment eases and digital pressures

Based on the weights of the leggings’ feature positions (shown in Table 11), the comprehensive fit level can be aggregated by equation (15):

pCEVA (g_{i}^{j}, s_{l}) = 0.063 \times {peva}^{L L} (g_{i}^{j}, s_{l}) + 0.171 \times {peva}^{W} (g_{i}^{j}, s_{l}) + 0.188 \times {peva}^{H} (g_{i}^{j}, s_{l}) + 0.184 \times {peva}^{T} (g_{i}^{j}, s_{l}) + 0.181 \times {peva}^{K} (g_{i}^{j}, s_{l}) + 0.213 \times {peva}^{C} (g_{i}^{j}, s_{l})

(15)

Table 11.

Weights of the feature positions of the leggings based on garment eases and digital pressures

Feature positions	Leggings length (LL)	Waist (W)	Hip (H)	Thigh (T)	Knee (K)	Calf (C)
Weights	0.063	0.171	0.188	0.184	0.181	0.213

Fit level-driven interactive garment design process in the 3D virtual environment

Having established the garment fit level prediction model, we put forward a new interactive garment design process in the 3D virtual environment based on the proposed garment fit level prediction model (see Figure 8). The overall garment design process involves the indispensable, sequential phases and one optional phase. The essential phases include consumer requirements acquisition, garment recommendation, virtual try-on and display, and garment fit assessment (see the left-hand side of Figure 8). After the apparel fit assessment phase and communication with the consumer, if the consumer is dissatisfied with the fit level at any local position, the garment pattern adjustment phase activates (see the right-hand side of Figure 8). The procedures of garment pattern adjustment are generally as follows. First, according to the outputs of the proposed model, fashion designers will discriminate the target positions of the garments that are unfitted. Next, the most relevant structure lines and points in the original garment patterns will be modified to meet the consumer’s requirements. The fashion designers will then check all the structure lines and points of the patterns to guarantee that only the target structure lines have been adjusted. Finally, a set of new patterns will be generated.

Figure 8.

Flowchart of the fit level prediction-based interactive garment design process in the 3D virtual environment.

For example, if the predictive output of the fit level of collar equals 2, indicating that the collar is a little tight, and the consumer’s target fit level is 3, then the garment pattern adjustment will activate. First, the most relevant structure lines and corresponding control points are identified, that is, the necklines in the front and back panels, the shoulder–neck points, and so forth. Second, the necklines will be lengthened by moving the shoulder–neck points accordingly. All the structure lines in the patterns will subsequently be checked to ensure the accuracy of the adjustment. Finally, a new set of patterns will be generated and recommended and displayed to the consumer for fit evaluation.

Through performing this iterative process of “recommendation–display–assessment–adjustment,” the consumer’s garment product requirements can ultimately be met.

Validation and application of the garment fit level prediction model

Comparison of the performance of the proposed garment fit level prediction model

The performance of the proposed PNN model was compared with other typical classifiers, including the linear regression (LR) model, the Support Vector Machine (SVM) model, the Radial Basis Function Artificial Neural Network (RBF_ANN) model, and the Back Propagation Artificial Neural Network (BP_ANN) model. The comparison experiment was performed as follows. First, the fit level prediction model was created based on the different algorithms, and trained using the 10-fold cross-validation approach. For each fold, all the models were trained using the same training dataset and the validation dataset. For each kind of algorithm, the final model was obtained after 10-fold cross-validation.

Second, the performances of the different models were tested and compared using an identical testing dataset that was independent of the learning dataset. Taking the collar fit level prediction model as an example, performance was compared using the same testing dataset, shown in Table 12.

Table 12.

The testing dataset for the collar fit level prediction model

SN	Input data												Output data
SN	CE	TH	SW1	SW2	SH	BW1	BW2	BR1	BR2	BS1	BS2	DE	Fit level
1	3.70	0.23	55	58	14	23	29	90	90	30	30	7	5
2	2.54	0.56	10	15	8	9	10	0	0	0	0	21	4
3	1.16	0.36	27	27	9	38	38	80	80	80	80	27	3
4	0.56	0.6	43	40	7	32	49	90	90	20	60	17	2
5	0.13	0.14	3	3	0	8	10	0	0	80	80	0	1
6	1.45	0.23	55	58	14	23	29	90	90	30	30	7	3
$⋮$	$⋮$	$⋮$	$⋮$	$⋮$	$⋮$	$⋮$	$⋮$	$⋮$	$⋮$	$⋮$	$⋮$	$⋮$	$⋮$
43	0.26	0.36	27	27	9	38	38	80	80	80	80	27	1

Note: SN: sample number; CE: collar ease.

Tables 13 to 15 show the results of the comparison among the different models. For shirt fit level prediction (see Table 13), in relation to average prediction accuracy, the proposed PNN model had the highest value (93.36%), followed by the BP_ANN (90.37%), the LR (69.77%), the SVM (67.11%), and the RBF_ANN (60.13%), which indicates that the prediction performance of the proposed model was superior to other classifiers. What can also be seen from Table 13 is that the prediction accuracy of each kind of the model fluctuates at different feature positions. For example, the standard deviation (SD) of the proposed PNN model (8.32%) was the lowest among all the classifiers, indicating that the prediction accuracy fluctuations of the PNN-based model were minimal. Therefore, the PNN model had the optimal performance in predicting shirt fit level among the four classifiers. For legging fit level prediction, based on ease allowances only (see Table 14), the BP_ANN model had the highest average prediction accuracy (88.76%) and the lowest SD (9.24%). The prediction accuracy of the PNN model was 72.10%, closed to that of the BP_ANN model. However, the SD of the PNN-based model rose to 21.67% for leggings fit level, which was the maximal SD of the five models. Thus, it can be concluded that the comprehensive performance of the BP_ANN model was optimal for ease-based legging fit level estimation. For legging fit level prediction based on clothing digital pressures combined with ease allowance (see Table 15), the average prediction accuracy and SD of the PNN model were 97.67% and 1.47%, respectively, which demonstrates that the overall performance of the PNN model was outstanding among five predictive models.

Table 13.

Performance comparison for shirt fit level

Sub-model	Prediction accuracy
Sub-model	PNN (%)	SVM (%)	RBF_ANN (%)	BP_ANN (%)	LR (%)
Garment length	86.05	74.42	53.49	67.44	100
Sleeve length	79.07	69.77	65.12	86.05	60.47
Collar	97.67	39.53	41.86	88.37	53.49
Bust	100	83.72	62.79	100	65.12
Waist	100	83.72	62.79	100	48.84
Hip	90.70	18.60	34.88	90.70	90.70
Shoulder	100	100	100	100	39.53
Average	93.36	67.11	60.13	90.37	69.77
Standard deviation	8.32	28.31	20.99	11.74	20.80

Table 14.

Performance comparison of the ease-based model for leggings fit level

Sub-model (ease-based)	Prediction accuracy
Sub-model (ease-based)	PNN (%)	SVM (%)	RBF_ANN (%)	BP_ANN (%)	LR (%)
Leggings length	100	74.42	65.12	76.74	53.49
Waist	86.05	88.37	67.44	93.02	55.81
Hip	79.07	60.47	55.81	95.35	53.49
Thigh	74.42	69.77	79.07	100	67.44
Knee	44.19	62.79	67.44	88.37	67.44
Calf	48.84	48.84	46.51	79.07	69.77
Average	72.10	67.44	63.57	88.76	61.24
Standard deviation	21.67	13.48	11.17	9.24	7.73

Table 15.

Performance comparison of ease and pressure-based models for leggings fit level

Sub-model (ease and pressure-based)	Prediction accuracy
Sub-model (ease and pressure-based)	PNN (%)	SVM (%)	RBF_ANN (%)	BP_ANN (%)	LR (%)
Leggings length	100	74.42	65.12	76.74	53.49
Waist	97.67	95.35	95.35	95.35	100
Hip	95.35	76.74	93.02	93.02	100
Thigh	97.67	88.37	93.02	93.02	90.70
Knee	97.67	81.40	72.09	90.70	100
Calf	97.67	60.47	62.79	76.74	90.70
Average	97.67	79.46	80.23	89.30	89.15
Standard deviation	1.47	12.08	15.20	8.48	18.05

Influence of the indicators on prediction accuracy for predicting the fit level of tight garments

For legging fit level prediction, it can be seen that the average prediction accuracy of the PNN model based on the ease and pressure was over 95.0% (see Table 15), while the average prediction accuracy of the PNN model based on ease was only 72.10% (see Table 14). Moreover, the SD value of the ease and pressure-based model was only 1.47%, which was dramatically lower than that of the ease-based model (21.67%). Therefore, this indicates that multiple indicators contribute to prediction performance for the fit level of tight garments.

Application of the garment fit level prediction model

To validate the proposed garment fit level prediction model, we applied it to the fit level prediction of a real shirt case for a specific male consumer. The application process of this case was as follows:

Step 1: We collected the body measurements of the consumer for the shirt, including neck girth $(N G)$ , shoulder width $(S W)$ , bust girth $(B G)$ , waist girth $(W G)$ , hip girth $(H G)$ , arm length $(A L)$ , and back length $(B L)$ , expressed by a vector $B M = \{\overset{N G}{39.3}, \overset{S W}{42.3}, \overset{B G}{93.3}, \overset{W G}{72.6}, \overset{H G}{88.6}, \overset{A L}{56.8}, \overset{B L}{45.2}\}$ .

Step 2: We obtained three shirts of different sizes and of the same fabric. The ease allowances for the three samples were then collected and given as follows:

{Ease}_{1} = \{\overset{C}{0.3}, \overset{S}{0.8}, \overset{B}{5.8}, \overset{W}{15.3}, \overset{H}{5.6}, \overset{S L}{0.6}, \overset{G L}{23.8}\}

{Ease}_{2} = \{\overset{C}{1.5}, \overset{S}{2.6}, \overset{B}{11.2}, \overset{W}{20.2}, \overset{H}{10.6}, \overset{S L}{2.3}, \overset{G L}{27.6}\}

{Ease}_{3} = \{\overset{C}{2.8}, \overset{S}{4.2}, \overset{B}{17.8}, \overset{W}{26.5}, \overset{H}{16.8}, \overset{S L}{3.8}, \overset{G L}{31.2}\}

where

C

S

B

W

H

S L

, and

G L

represent the ease allowances of the collar, shoulder, bust, waist, hip, sleeve length, and shirt length, respectively.

Step 3: Combined with the fabric parameters, the ease data were imported into the proposed fit level prediction model to obtain the garment fit levels for the samples,

{Fitlevel}_{1} = \{\overset{C F}{1}, \overset{S F}{2}, \overset{B F}{2}, \overset{W F}{2}, \overset{H F}{2}, \overset{SLF}{2}, \overset{GLF}{2}, \overset{F}{1.858}\}

{Fitlevel}_{2} = \{\overset{C F}{3}, \overset{S F}{3}, \overset{B F}{3}, \overset{W F}{3}, \overset{H F}{3}, \overset{SLF}{4}, \overset{GLF}{4}, \overset{F}{3.239}\}

{Fitlevel}_{3} = \{\overset{C F}{4}, \overset{S F}{4}, \overset{B F}{4}, \overset{W F}{4}, \overset{H F}{4}, \overset{SLF}{5}, \overset{GLF}{5}, \overset{F}{4.239}\}

where,

C F

S F

B F

W F

H

SLF

, and

GLF

represent the predicted fit levels of the collar, shoulder, bust, waist, hip, sleeve length, and shirt length, respectively.

F

refers to the predicted comprehensive garment fit level.

From the results, the $F$ value of Sample 2 was 3.239, which is closer to 3 than the other two samples, indicating that Sample 2 was a better fit for the consumer. We were therefore able to recommend Sample 2 to the consumer.

Interactive garment design based on the proposed garment fit level prediction model

A new garment design process for a 3D virtual environment can be developed by integrating the proposed fit level prediction model into commercial 3D clothing design software. The process for a specific consumer is as follows. First, the mannequin of the consumer, whose abdomen was convex (see Figure 9(a)) was created in a 3D design environment. We then designed, recommended, and demonstrated the garment using virtual try-on, as well as the garment fit levels predicted by the proposed model (see Figure 9(b) and (c)). From the results of the fit assessment, we found that the waist elements of the shirt and leggings were tight. To resolve this, we identified and adjusted the most relevant structure lines and the corresponding control points on the real patterns, such as the waist lines in the front and back panels. We subsequently recommended and displayed the new garment to the consumer based on virtual try-on technology. Finally, the optimized garment (see Figure 9(d)) product that met the consumer’s requirements was obtained through performing an interaction between real garment pattern adjustment and the virtual try-on display. This way, we were able to realize the development of individualized garments accurately and rapidly. In addition, we will be able to promote mass customization by comprehensively taking into account the influences of factors like individualization, productivity, and cost.

Figure 9.

Interactive garment design using the proposed garment fit prediction model.

Extension of the proposed garment fit level prediction approach

The general ideas and principles can be extended to various target markets and populations, as well as diverse garment styles. The process is as follows:

According to the characteristics of the target market and population, as well as the specific garment styles, corresponding representative subjects must be recruited and the corresponding learning data collated following the experimental steps described in this paper.

Based on the learning data, the corresponding garment fit level prediction model based on the method proposed in this study must be developed.

Through the above-mentioned steps, the garment fit level prediction model for various target markets, populations, and garment styles can be realized and applied.

Conclusions

In this study, we put forward a new garment fit level prediction approach based on the PNN, as well as a new interactive garment design process in a 3D virtual environment by combining the proposed fit level prediction model and classic 3D clothing design software. From the results, we can conclude that (a) the garment ease allowances in the 3D virtual environment were good indicators for loose garment fit level recognition; (b) for the tight garments, multiple indicators that combined digital pressures with ease allowances were superior to the single indicator of ease allowance in predicting fit level; (c) the PNN model performed better than the LR, the SVM, RBF_ANN, and BP_ANN models in prediction of garment fit level for both loose- and tight-fitting garments. The model proposed in this study had the following advantages: (1) the proposed model can not only predict comprehensive fit levels, but also the local fit levels of the garments rapidly, precisely, and automatically; (2) the model provides a feasible and efficient solution for perceiving garment fit levels without real try-on, which can be utilized in both online and in-store shopping to enhance consumer satisfaction; (3) the performance of the proposed model can be continuously enhanced via learning new datasets. The proposed model can be further improved and extended in the future. A personalized garment design system can be developed to offer consumers accurate garments for each specific fit level. From an industrial application viewpoint , more datasets should be considered and acquired in future research.

Footnotes

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The Fundamental Research Funds for the Central Universities and Graduate Student Innovation Fund of Donghua University (grant no. CUSF-DH-D-2020091); Dong Hua University’s excellent PhD international visit program; the Fundamental Research Funds for the Central Universities (grant no. 2232020G-08); the Open Project Program of Key Laboratory of Silk Culture Heritage and Products Design Digital Technology of Ministry of Culture and Tourism of China (grant no. 2020WLB07); the European Horizon 2020 Research Program (Project: FBD_BModel, No. 761122); the Social Science Planning Project in Anhui (grant no. AHSKQ2019D085); the Scientific Research Project of Anhui Polytechnic University (grant no. Xjky2020055); the Key Research Project of Humanities and Social Sciences in Anhui Province College (grant nos. SK2016A0116 and SK2017A0119); the Open Project Program of Anhui Province College Key Laboratory of Textile Fabrics, Anhui Engineering and Technology Research Center of Textile (grant no. 2018AKLTF15); and the National Key Research and Development Program of China (grant no. 2019YFF0302100).

ORCID iD

Zhujun Wang

References

Z-H

Wei

, et al. Examining collaborative filtering algorithms for clothing recommendation in e-commerce. Text Res J 2018; 89: 2821–2835.

Clement

Worldwide e-commerce share of retail sales 2015–2023, www.statista.com/statistics/534123/e-commerce-share-of-retail-sales-worldwide/ (2020, accessed 26 November 2020).

Pei

Fan

Ashdown

SP.

A novel optimization approach to minimize aggregate-fit-loss for improved breast sizing. Text Res J 2020; 90: 1823–1836.

Hernández

Mattila

Berglin

Can virtually trying on apparel help in selecting the correct size?

Cloth Text Res J 2019; 37: 249–264.

Sohn

J-M

Lee

Kim

D-E.

An exploratory study of fit and size issues with mass customized men’s jackets using 3D body scan and virtual try-on technology. Text Res J 2020; 90: 1906–1930.

Nasibov

Demir

Vahaplar

A fuzzy logic apparel size decision methodology for online marketing. Int J Cloth Sci Technol 2019; 31: 299–315.

timetrade. A look at the apparel industry, www.xxx (2015, accessed xx Month Year).

Hong

Zeng

Bruniaux

, et al. Interactive virtual try-on based three-dimensional garment block design for disabled people of scoliosis type. Text Res J 2017; 87: 1261–1274.

Hong

Bruniaux

Zeng

, et al. Design and evaluation of personalized garment block for atypical morphology using the knowledge-supported virtual simulation method. Text Res J 2018; 88: 1721–1734.

10.

Liu

Zeng

Bruniaux

, et al. 3D interactive garment pattern-making technology. Comput Aided Des 2018; 104: 113–124.

11.

Tao

Chen

Zeng

, et al . A customized garment collaborative design process by using virtual reality and sensory evaluation on garment fit. Comput Ind Eng 2018; 115: 683–695.

12.

Tao

Bruniaux

Toward advanced three‐dimensional modeling of garment prototype from draping technique. Int J Cloth Sci Technol 2013; 25: 266–283.

13.

Thomassey

Bruniaux

A template of ease allowance for garments based on a 3D reverse methodology. Int J Ind Ergon 2013; 43: 406–416.

14.

Song

A novel approach for fit analysis of thermal protective clothing using three-dimensional body scanning. Appl Ergon 2014; 45: 1439–1446.

15.

Liu

Zeng

Bruniaux

, et al. Fit evaluation of virtual garment try-on by learning from digital pressure data. Knowl-Based Syst 2017; 13: 174–182.

16.

Liu

Zou

, et al. Apparel ease distribution analysis using three-dimensional motion capture. Text Res J 2019; 89: 4323–4335.

17.

Thomassey

Zeng

Artificial Intelligence for Fashion Industry in the Big Data Era. Singapore: Springer Singapore, 2018.

18.

Giri

Jain

Zeng

, et al. A detailed review of artificial intelligence applied in the fashion and apparel industry. IEEE Access 2019; PP: 1–1.

19.

Guo

Wong

Leung

SYS

, et al. Applications of artificial intelligence in the apparel industry: a review. Text Res J 2011; 81: 1871–1892.

20.

Hong

Liu

Bruniaux

, et al. Application of 3D-TO-2D garment design for atypical morphology: a design case for physically disabled people with scoliosis. Ind textilă 2018; 69: 1.

21.

Zeng

Koehl

Representation of the subjective evaluation of the fabric hand using fuzzy techniques. Int J Intell Syst 2003; 18: 355–366.

22.

Zeng

Liu

A learning automata based algorithm for optimization of continuous complex functions. Inf Sci 2005; 174: 165–175.

23.

Zeng

Ruan

Koehl

Intelligent sensory evaluation: concepts, implementations, and applications.

Math Comput Simul 2008; 77: 443–452.

24.

Chen

Zeng

Koehl

, et al. Optimization of human perception on virtual garments by modeling the relation between fabric properties and sensory descriptors using intelligent techniques. In: Hüllermeier E, Kruse R and Hoffmann F (eds) Information Processing and Management of Uncertainty in Knowledge-Based Systems. New York: Springer International Publishing, 2014, pp. 606–615.

25.

Chen

Tao

Zeng

, et al. Control and optimization of human perception on virtual garment products by learning from experimental data. Knowl-Based Syst 2015; 87: 92–101.

26.

Choi

T-M

, Hui

C-L

(eds). Intelligent Fashion Forecasting Systems: Models and Applications. Berlin: Springer-Verlag Berlin Heidelberg, 2014.

27.

Wong

Yeung

Predicting clothing sensory comfort with artificial intelligence hybrid models. Text Res J 2004; 74: 13–19.

28.

Wang

Zhou

, et al. An adaptive neural fuzzy network clothing comfort evaluation model and application in digital home. Multimed Tools Appl 2014; 71: 395–410.

29.

Kalkanci

Kurumer

Ozturk

, et al. Artificial neural network system for prediction of dimensional properties of cloth in garment manufacturing: case study on a t-shirt. Fibres Text East Eur 2017; 25: 135–140.

30.

Leung

Tang

, et al. Robust order scheduling in the discrete manufacturing industry: a multi-objective optimization approach. IEEE Trans Industr Inform 2018; 14: 253–264.

31.

Thomassey

Zeng

Optimization of garment sizing and cutting order planning in the context of mass customization.

Int J Adv Manuf Technol 2020; 106: 3485–3503.

32.

Xia

Zhang

Weng

, et al. Fashion retailing forecasting based on extreme learning machine with adaptive metrics of inputs. Knowl-Based Syst 2012; 36: 253–259.

33.

Xia

Wong

WK.

A seasonal discrete grey forecasting model for fashion retailing.

Knowl-Based Syst 2014; 57: 119–126.

34.

Ren

Choi

Liu

Fashion sales forecasting with a panel data-based particle-filter model. IEEE Trans Syst Man Cybern Syst 2015; 45: 411–421.

35.

Hui

PCL

Choi

TM.

Using artificial neural networks to improve decision making in apparel supply chain systems, in Information Systems for the Fashion and Apparel Industry. In: Choi T-M (ed) Information Systems for the Fashion and Apparel Industry. Cambridge, UK: Woodhead Publishing, 2016, pp. 97–107.

36.

Wen

Choi

T-M

Chung

S-H.

Fashion retail supply chain management: a review of operational models. Int J Prod Econ 2019; 207: 34–55.

37.

Specht

DF.

Probabilistic neural networks for classification, mapping, or associative memory. In: IEEE 1988 international conference on neural networks (ed A Editor), San Diego, USA, 24--27 July 1988, paper no. 3327888, pp. 525–532. Piscataway, NJ: IEEE.

38.

Specht

DF.

Probabilistic neural networks. Neural Netw 1990; 3: 109–118.

39.

Mohebali

Tahmassebi

Meyer-Baese

, et al. Chapter 14. Probabilistic neural networks: a brief overview of theory, implementation, and application. In: Samui P, et al. (eds) Handbook of Probabilistic Models. Oxford, UK: Butterworth–Heinemann, 2020, pp. 347–367.

40.

Jing

, et al. Automatic classification of woven fabric structure based on texture feature and PNN. Fibers Polym 2014; 15: 1092–1098.

41.

Mwasiagi

Wang

Huang

XB.

The use of k-means and artificial neural network to classify cotton lint.

Fibers Polym 2009; 10: 379–383.

42.

, et al. A two-scale attention model for intelligent evaluation of yarn surface qualities with computer vision. J Text Inst 2018; 109: 798–812.

43.

Tolba

AS.

Fast defect detection in homogeneous flat surface products.

Expert Syst Appl 2011; 38: 12339–12347.

44.

Hasnat

Ghosh

Khatun

, et al. Pattern classification of fabric defects using a probabilistic neural network and its hardware implementation using the field programmable gate array system. Fibres Text East Eur 2017; 25: 42–48.

45.

Luo

Ghorashi

Duckett

, et al. Color grading of cotton Part II: color grading with an expert system and neural networks. Text Res J 1999; 69: 893–903.

46.

Kuo

C-FJ

Shih

C-Y

Hsu

C-TM.

Pattern-making simulation on embroidery using probabilistic neural network and texture fitting method. Text Res J 2011; 81: 2082–2094.

47.

Mohamed Ahmed

Fault diagnosis of spinning industrial systems by using wavelet analysis and soft computer classifier. Ind Eng Manag 2013; 2: 2.

48.

Jin

Yang

Zou

. Developing an intelligent model for the construction a hip shape recognition system based on 3D body measurement. Fibres Text East Eur 2016; 24: 110–118.

49.

Shahrabi

Hadavandi

Salehi Esfandarani

Developing a hybrid intelligent model for constructing a size recommendation expert system in textile industries. Int J Cloth Sci Technol 2013; 25: 338–349.

50.

Wong

Zeng

XH.

Classification decision model for a Hong Kong clothing manufacturing plant locations using an artificial neural network. Int J Adv Manuf Technol 2006; 28: 428–434.

51.

Standardization Administration of the People's Republic of China. Standard sizing systems for garments—men: GB/T 1335.1-2008. Beijing: Standards Press of China, 2008.