A novel approach in predicting virtual garment fitting sizes with psychographic characteristics and 3D body measurements using artificial neural network and visualizing fitted bodies using generative adversarial network

Abstract

Advances in technology have brought accessibility to garment product fitting procedures with a virtual fitting environment and, in due course, improved the supply chain socially, economically, and environmentally. 3D body measurements, garment sizes, and ease allowance are the necessary factors to ensure end-user satisfaction in the apparel industry. However, designers find it challenging to recognize customers’ motivations and emotions towards their preferred fit and define ease allowances in the virtual environment. This study investigates the variations of ease preferences for apparel sizes with body dimensions and psychological orientations by developing a virtual garment fitting prediction model. An artificial neural network (ANN) was employed to develop the model. The ANN model was proved to be effective in predicting ease preferences from two major components. A non-linear relationship was modeled among pattern parameters, body dimensions, and psychographic characteristics. Also, to visualize the fitted bodies, a generative adversarial network (GAN) was applied to generate 3D samples with the predicted pattern parameters from the ANN model. This project promotes mass customization using psychographic orientations and provides the perfect fit to the end users. New size-fitting data is generated for improved ease preference charts, and it enhances end-user satisfaction with garment fit.

Highlights

• •Psychographic orientations and anthropometric data were included to construct size charts.
• •Applying ANN model to predict virtual garment fitting sizes.
• •Applying GAN model to visualize 3D prototypes from predictions.
• •Better-off apparel supply chain socially, economically, and environmentally.
• •Achieved mass customization with best garment fit for end users.

1. Introduction

In recent years, an increasing number of designers and apparel companies have employed 3D virtual garment simulation technologies when they develop a product in its early phases. They simulate 3D prototypes for manufacturers and evaluate the realistic 3D full-view prototypes in pre-production development. Variables that are often modified in designs include patterns, fabrics, textures, and sewing methods, which can be immediately adjusted in the virtual environment. Not only does the 3D simulation software reduce the waste of redundant samples and materials, but it also engages consumer involvement for mass customization [1,2]. In the virtual prototyping process, many physical garment properties are well-tailored, such as the fabrics, texture effects, stress and stretch features, and sewing patterns. A benefit of using 3D virtual garment simulation software for designers and manufacturers is flexibility. The software can simulate practical processes in a virtual environment with functionalities for 3D virtual garment simulation, 3D stitching, 2D pattern modification, and 2D pattern grading [3,4]. Virtual 3D prototypes can be visualized for evaluation, and it is much easier for designers and consumers to discuss the details before confirming their orders [3,5]. Hwang Shin and Lee [6] reviewed the virtual fashion literature and found that up to a 50% reduction in production lead time and a 70% cost savings on pattern development can be obtained by employing virtual samples. This enhancement in sample review has led to efficient communication, development, and production between technical designers and manufacturers. Furthermore, such advanced apparel design and fitting processes improved the fashion supply chain from garment manufacturing to retail sales in economic, social, and environmental facets. Knowing that garment fitting can bring efficient performance and enhance consumer interaction, professionals in the fashion industry are increasingly passionate about employing 3D virtual simulation software [7].

Although 3D virtual simulation software can aid designers and manufacturers in developing their products in the early stages and has huge impacts on improving the whole supply chain's efficiency, there are still some constraints. In practice, there are no universal standards for garment grading due to demographic differences, and thus apparel manufacturers will have their own rules or principles for their products. One critical factor for them to define the size charts is ease allowance, where designers would add extra lengths in the sketches to provide sufficient room for body movement when it comes to garment fit assessments [8]. Chen et al. [9], have demonstrated that the changes in ease allowance will affect garment fit and comfort, and many studies have been conducted to find the best ease allowances for different garment types, genders, demographics, and cloth textures in the real and virtual worlds (e.g., [[10], [11], [12], [13], [14]]). Nevertheless, their methods may not be flexible for different consumer groups and require additional fitting experts' judgment to assess the final fit evaluation. Those studies also did not look at the influence of consumers' psychographic characteristics on garment fit and satisfaction. It means the consumers' fashion sense and their lifestyle orientation, i.e., activities, interests, and opinions (AIO). AIO stands for the level of consumer willingness to invest time and money in lifestyle [[15], [16], [17], [18]]. Unquestionably, physical attributes, such as body measurement and garment pattern measurements, are important for evaluating garment fit. However, notably, consumers' preference for style is a valuable factor to be considered. As a result, this study evaluates both the physical fit and psychographic characteristics of consumers to recommend ease allowance preference charts in terms of their garment fit satisfaction.

This study will investigate the effectiveness of using an artificial neural network (ANN) to predict virtual garment fitting sizes with 3D body measurements and psychographic characteristics of wearers and present a framework modeled by a generative adversarial network (GAN) to visualize a fitted 3D prototype immediately after the predicted pattern parameters are retrieved from the ANN model. Section 2 of this paper reviews previous works, and Section 3 will introduce the methods. The results are discussed in Section 4, and the study is wrapped up in Section 5.

2. Literature review

2.1. Factors for using 3D simulation software in apparel design and production

Generally, the 3D simulation software requires designers to create or import 2D pattern sketches, allowing modifications to the garment's physical properties in real time. The design collections can then be fitted and shaped onto the appropriate 3D male or female mannequins for prototype demonstration and fit evaluation by looking at the drape features, stress, and stretch effects. A previous study has raised three important elements for creating these fitted 3D prototypes, which involve: i) dimensions and measurements of the clothing; ii) ease allowances between each clothing size; and iii) 3D avatar size choices in software [19]. The authors also concluded that an accurate and efficient sizing system should have a comprehensive apprehension of body proportions, so that it can aid in enhancing garment fit and have favourable implications on the apparel industry's pursuit of sustainable growth. With defined lengths in CAD systems, obtaining garment dimensions, measures, and ease allowances is simple. Nevertheless, choosing the size of 3D avatars requires real data from consumers. Body scanning technology plays a major role in acquiring the data, and it helps define sizes in the ready-to-wear clothing market [1,14,20]. A lot of fashion businesses use 3D body scanners to determine the body shapes and dimensions of their target audiences, and thereby, produce custom-fitted clothing and update their sizing charts simultaneously [[21], [22], [23], [24]]. With real data from consumers, designers can generate 3D mannequins that are normally used for fitting and construct size charts according to the consumer's body data. Before using simulations with virtual avatars and apparel, it is essential to understand the relationship between body measurements and outfit sizes.

2.2. Defining size charts and constraints

Since differences in measurements for each body part, including the chest, waist, hips, shoulder length, body length, and more, are used to determine how clothing sizes vary, there is no international standard for grading garments due to demographical differences, and different companies will have their own rules or principles for their products. When companies define their size charts, one critical factor is ease allowance, to which designers would add extra lengths in the sketches to provide sufficient room for body movement when it comes to garment fit assessment [8]. Variations in ease allowance will affect garment fit and comfort, and many previous studies have been conducted to find the best ease allowances for different garment types, genders, demographics, and cloth textures [[25], [26], [27], [28]]. Recent studies have also evaluated the ease allowance using virtual garment fitting technologies [29,30]. However, these studies did not investigate the effects of consumers' psychographic characteristics on garment fit and satisfaction. It is essential to know consumers’ preferences for style when fit is evaluated, since some may prefer wearing oversized garments while others prefer tight clothing such as yoga pants. Therefore, to assess the overall comfort and satisfaction of virtual clothing, both physical and psychological appraisals should be taken into consideration.

2.3. Mechanism in selecting examined garments

In this study, simple and unisex garment types were chosen as the control prototypes. It aims at reducing the variability in the data and obtaining a clearer understanding of the garment parameters. By reducing the complexity of the garments being tested, the study can focus on the judgment of the visualized 3D avatars wearing the common garment types with the subjects and improve the reproducibility and validity of the study.

The four chosen garment types are blazers, long-sleeved shirts, short-sleeved shirts, and trousers. They took up relatively large portions of the men's and women's apparel market revenues in 2022 [31,32] (see Appendix 1). With the objective of providing control prototypes in both genders for further complex garment designs, such as multiple layering, the choice of employing the selected garment types in this study is significant as a baseline.

2.4. Preliminary work for the study

Our preliminary work (see Chan et al. [19]) developed a conceptual framework for the study. The study constructed an ANN model to predict the virtual garment fitting sizes for one type of garment, short-sleeved shirts, using two components: 3D body measurements and psychographic characteristics of the recruited subjects. The subjects were asked to conduct body scans to retrieve their anthropometric data using body scanners and complete a questionnaire survey to consolidate their psychological orientations toward fashion. They were also involved in the co-designing process of the virtual garment simulation to retrieve their preferences on ease allowances for the assessed garment type. The results were satisfactory with high accuracy, and we observed a non-linear relationship among garment pattern parameters, body dimensions, and psychographic characteristics. The work had some limitations; only one garment type was studied, and the number of subjects may not be sufficient to indicate any bias.

2.5. Using artificial neural networks in apparel industry

In the current decade, ANN has been gradually employed in scientific research and industrial applications due to its advantages in non-linear estimation capabilities and self-arranging methodology. One of the most well-employed machine learning methods ANN models have been developed for prediction in designing clothing patterns. Zheng Liu et al. [33] presented their non-linear model using feature parameters to forecast the specific body sizes; however, the approach was too complicated for practical applications to be calculated by designers and manufacturers. Another research by Liu et al. [34] offered their prediction technique to forecast the body measurements for clothing pattern manufacturing based on anthropometric data. Their approach was significant, with higher prediction accuracy and stability than linear regression models. Nevertheless, their prediction model only projected the lower body measurements for pants pattern design. Wang et al. [35] improved the methods by applying the radial basis function ANN and engaging with the difficulties of the designers and manufacturers with limited expertise and experience; nonetheless, only pants pattern design was considered in the application. Despite the lack of consideration for full-body apparel pattern design, the mentioned articles did not address the influence of lifestyle orientations towards clothing selection and purchase intention in the development of the prediction model. Islam et al. [36] identified several factors that would affect customers in Bangladesh purchasing garments, ranked from quality, colors and designs, output and comfort ability, price, variations in styles, wearing purpose, and status symbol. These factors were valuable to be included in the prediction since the psychographic orientation of a customer can also affect the garment fit; for example, some customers may be concerned about the fabric texture, and some may prefer wearing tight garments. Therefore, the purpose of this study is to present a new ANN model that incorporates both anthropometric data and the psychographic orientation of a customer wearing a garment. Full-body garment pattern parameters for design and making can be delivered. Additionally, the proposed model would include real-time 3D simulations for customers to co-design the virtual garments with designers and obtain their first-hand experience in evaluating the garments.

3. Methods

To improve size predictions of 3D clothes, this study examines apparel sizes for virtual fitting using two main components: body measurements and psychographic traits of subjects and their ease preferences. The examined garment types are blazers, long-sleeved shirts, short-sleeved shirts, and trousers. It is anticipated to generate a wide-ranging ease allowance preference chart for achieving the ideal fit for consumers. An ANN was engaged to develop a prediction model so that the pattern parameters for generating the ideal fit garment sketches could be obtained from the body dimensions and psychographic orientations of the consumers. GAN was also deployed to visualize 3D avatars wearing the garment immediately after the predicted results from the ANN model were received. In this study, 120 subjects aged 18 years or older were recruited to perform 3D body scans and a survey questionnaire for defining their clusters of physical and psychological data correspondingly. They were also requested to evaluate fitting preferences by adjusting some parameters in the co-designing process using commercial 3D garment simulation software called Optitex [37], for example, sleeve length and waist. Ease allowance preferences from the respondents were personalized against the predetermined values on the software, where the criteria were mainly based on the participants’ responses. A theoretical framework in Fig. 1 indicates the workflow of the study in two stages. The study complies with all regulations and confirms that the experiments with human subjects were approved by the Human Subjects Ethics Sub-committee (HSESC) of Technological and Higher Education Institute of Hong Kong (HSESC No.: HE2021-07).

Fig. 1.

Theoretical framework for the research.

3.1. Body measurement collection and pre-processing procedures

Anthropometric data, such as the body's length, waist, hips, and chest, were collected from the participants by 3D body scanning. A suite of 3D body scanner device and software called Styku was employed to retrieve the body dimensions of participants. It can execute full-body scanning, generate 360-degree 3D models, and extract the measurements of scanned body shapes. The innovative recognition technology in the scanner can automatically distinguish body parts from the scanned 3D models and produce accurate body circumferences, volumes, surface areas, and lengths for garment fit processes (see Fig. 2). Classification of body shapes and sizes was executed to allocate subject data according to the four foremost garment sizes, which were Small, Medium, Large, and Extra Large.

Fig. 2.

Co-designing an experiment on virtual fitting with participants (3D point clouds imported into Optitex).

3.2. Instruments for psychographics categorization

To distinguish the psychological preferences on fashion and ease allowance preferences of the recruited subjects, they were involved in an interview to carry out a structured questionnaire survey based on their lifestyle orientation, i.e., AIO. The survey includes two sessions: i) demographic information (which is age, occupation, and income level); and ii) 15 fashion lifestyle statements [38,39] and activity-related AIO items. Those AIO items are categorized by four factors, which are brand consciousness, sensational, practical, informational. Knowing the AIO items is crucial to gauge lifestyles and consumption and determine the favorite activities that respondents devote the majority of their time and effort to Refs. [[15], [16], [17], [18]].

3.3. Mechanisms for 3D garment fit simulation

After clustering the respondents by their body dimensions and psychographic characteristics, their unique preferences in garment sizes were observed during the co-designing session. Participants would evaluate the garment fit for each 3D model using the virtual simulation software, Optitex. The 3D mannequins were scanned models from the participants’ body scans, which were fitted with four different garment types, blazers, long-sleeved shirts, short-sleeved shirts, and trousers. In the draping simulation, respondents were given the option to change the physical characteristics of the clothing to get the fit they desired, where the mannequins and garments were visualized in real-time for evaluation [40,41].

3.4. The co-designing process of virtual fit evaluation with subjects

As a virtual body was created after each body scan with the Styku body scanning device, each subject had their equivalent avatar per their body scan and point cloud data captured from the scan. For each subject, his or her virtual avatar was imported into Optitex to digitalize the drape performance when the virtual garment was fitted onto the virtual body. By adjusting the amount of ease allowance on various body parts, the participant was instructed to co-design the fit for four garment types (a blazer, a long-sleeved shirt, a short-sleeved shirt, and pants). Speedy, reliable, and real-time fit simulation by Optitex permits users to examine the garment sizes and forms from a series of studied items and compare them for justification. Once the subject has made the adjustments to their most desired version, a 2D pattern sketch will be created from the 3D apparel, where the pattern parameters can be retrieved for creating sets of ideal ease charts for each subject. A significant component of this study is the co-design process with the subjects. During the process, the subjects’ immediate responses to the garment size, shape, and visualization were obtained to refine the virtual garments according to their preferences. Customer engagement and their psychological orientations can be acquired through interaction with 3D pattern design software and real-time realistic visualization.

3.5. Hyperparameters for artificial neural network to develop a prediction model

The output of this study is to produce consistent ease allowance preference charts based on two factors: body measurements and psychographic characteristics. A virtual garment fitting prediction model was developed with the application of an ANN for predicting the critical factors to create a custom 2D pattern sketch. In our model, the input units are P (Perception of Fitting Factors by Consumers’ Psychological Orientation) and B (Body Measurements), in which these two units are the necessary components for drafting sketch patterns, and the output unit is R (Pattern Parameters for Sketch Drafting). The hyperparameters for developing the ANN prediction model are reviewed in Table 1, and the network architecture is in Fig. 3. The activation function at the hidden and output layers is the leaky rectified linear activation function (ReLU), and it was selected since the data was in a continuous form and had a higher computational efficiency with small ranges of input values in [0, 1], where the normalized input values can have fewer fluctuations in the training process.

Table 1.

Hyperparameters for artificial neural network prediction model.

Hyperparameters	Values
Activation function at hidden layer	ReLU $f\left(x\right)=\mathit{max}\left(0,x\right)$
Activation function at output layer	ReLU $f\left(x\right)=\mathit{max}\left(0,x\right)$
Input parameters	$\mathit{P}$ (Perception of Fitting Factors by Consumers' Psychological Orientation) $\mathit{B}$ (Body Measurements)
Output parameters	$\mathit{R}$ (Pattern Parameters for Sketch Drafting)
Number of hidden layers	6
Number of hidden nodes in each layer	Layers 1 & 2: 70 Layers 3 $ 4: 140 Layers 5 & 6: 50
Number of training rows	70
Number of testing rows	50
Learning rate	0.001
Loss function	Mean Squared Error
Optimizer	Adaptive Moment Estimation (Adam)
Training Time for the Model	1–5 min

Fig. 3.

Network architecture of the artificial neural network model.

The Python language was employed to write a program for the ANN training and application model. There were four garment types: i) blazer; ii) long-sleeved shirt; iii) short-sleeved shirt; and iv) trousers. The data for each garment type was stored in an Excel file for program training and evaluation. The division for training and data testing is as follows: In the ANN training-testing procedure, 70 randomly selected respondents' data would serve as the training data and 50 respondents' data would serve as the testing data. The data was normalized into the format [0, 1], and the learning rate was set to the comparative effective value of 0.001 [42,43]. Two features, the optimizer and loss function, were defined before model training. The optimizer in the model used adaptive moment estimation (Adam) for stochastic optimization for simple datasets [44]. To compute the average of the squared deviations between the predicted data and the actual data, the loss function was applied using the Mean Squared Error (MSE), which is the standard loss function for many regression situations. Training the dataset reduces the losses between the predicted and actual values and produces high accuracy in the evaluation of the squared correlation coefficient (R²) for the trained model. The application would access the Excel files containing the input data and employ the activation function for the training process. Another Excel file would be used to record the projected pattern parameters for each type of garment.

3.6. Generative adversarial network model for creating prediction visualization

With the body measurements and the predicted pattern parameter values from the ANN model, 3D sample prototypes can be established with the use of a GAN method and the application of an open-source algorithm called PIFuHD (Pixel-aligned Implicit Function in High Resolution) as proposed by Saito et al. [45]. The retrieved body scan data from the 120 subjects would be the reference data for the GAN model to produce relatively realistic samples according to the user's body dimensions. The system would retrieve a similar dataset to the user data and generate a 3D model of the garment (in “.obj” format) for visualization. Table 2 shows the hyperparameters for controlling the sample image generation from the thematic data. Data augmentation was performed to produce more training data [46]. Each size of the garment would have 5000 to 15,000 training images. The GAN model follows the structure proposed by Goodfellow et al. [47] and Saxena and Cao [48], which is a Deep Convolutional Generative Adversarial Network (DCGAN) model. Fig. 4 is the GAN network architecture for this study. It consists of a generator and discriminator to make and define false images correspondingly, until the produced images are high in accuracy and have realistic appearance that the discriminator cannot identify whether the image is real or false. At the end, numerous 2D front-view images of the avatar wearing the garments are generated for different clusters of body measurements and psychological orientations.

Table 2.

Hyperparameters for generative adversarial network generation model.

Hyperparameters	Values
Training Data	5000-15000 front-view images of each garment type (applied with data augmentation)
Batch Size and Image Size	128
Channel	3 for colour images
Number of training epochs	50
Generator Layers	15
Discriminator Layers	17
Learning rate for optimizer	0.0002
Optimizer	Adaptive Moment Estimation (Adam)
Loss function	Binary Cross Entropy Loss
Training and Generation Time	30–60 min

Fig. 4.

Network architecture for GAN model (with Generator and Discriminator).

After acquiring sufficient 2D front-view images from the GAN model, the images would be inserted into a modified version of PIFuHD to convert the 2D images into 3D human models. PIFuHD is a single-view human image reconstruction algorithm projected by Saito et al. [45], where its outcomes are of high fidelity when compared to other cutting-edge 3D human reconstruction techniques like DeepHuman and Tex2Shape. Through those pioneering methods, it can present a whole human model from the image. Saito et al. [45], compared the qualitative results of the 3D avatars from their substitutes, and demonstrated that their PIFuHD model has more visualizing details. As a result, PIFuHD was employed in this study to establish the visualization of 3D fitted samples with the garment.

Open-source code for PIFuHD is available online [49]. However, the current source codes only allow one image reconstruction per run. To generate thousands of 3D models in one run, we employed modified versions that input all images for each cluster and return the 3D models in .obj format.

Despite the 3D scanning data can also develop 3D models of the participant, yet the participants often do not (not able to) wear tight garment for the scanning (as instructed), leading to inaccurate body measurement and visualization. The aim of using the DCGAN model in the study is that it can generate 3D samples of more accurate and realistic body proportions of the subjects and the appropriate garment dimensions. The subjects' body shapes may be slightly different during the day, for example, the waistline may increase after the subject consumed meal, or wearing loosely fitting clothes, and thereby variations on subjects’ perceptions are expected in this case. Thus, our approach is providing more tolerance in simulating a more realistic and practical wearing environment. Subjects can evaluate each edition and provide comprehensive perceptions of the four garment types in the co-design process. Designers can also have unlimited 3D avatars in similar sizes and shapes to test drape features and fit evaluation, and they base their decisions on the results to define sizes with high accuracy and pragmatic simulations. The proposed model, if works, could also be useful for mass customization of tailor-made garment productions.

4. Results and discussion

4.1. Factual data for psychographic characteristics

A K-means clustering analysis was performed to divide 120 participants into five groups based on their psychographic characteristics. Table 4 summarizes the data for each division. The group with the highest AIO involvement in spending their spare time on lifestyle is P5, while the group with the lowest AIO involvement is P1. From the table, the mean score of psychographic characteristics for P5 is 4.57 (out of 7), while for P1, it is 3.37.

Table 4.

Summary of psychographic characteristics for each K-means clustering division.

	P1	P2	P3	P4	P5
Mean	3.37	3.72	4.01	3.94	4.57
Min	2.33	2.87	3.33	3.07	3.67
Max	4.07	4.47	5.2	4.8	6.13
SD	0.3883	0.4314	0.5037	0.4661	0.5811

4.2. Comparative analysis of artificial neural network with alternative mechanistic models

To understand the superiority of ANN in developing the prediction model, a comparative analysis with several mechanistic models was done to assess how well the models worked with the dataset. Support vector regression (SVR) in a multi-output regressor, k-nearest neighbors regressor (KNN), and linear regression in a multi-output regressor (MLR) were developed to compare with the ANN model in one hidden layer by evaluating the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). According to the results presented by Moges et al. [50], it is expected that ANN can have the best performance among all the chosen regression-based models. Table 3 provides a performance evaluation matrix for the training and validation of four different garment kinds. It explains that the ANN model had the best performance and stability when the pattern parameters were predicted for the four garment types. All models performed well during validation using the train dataset, with RSME ranging from 0.0025 to 0.0203 and MAE ranging from 0.0001 to 0.0151. Negative R² were only identified in SVR, showing that SVR did not follow the trend of the train dataset and was not competitive in either case, while the remaining models had comparable higher R² in the descending order of MLR, ANN, and KNN. MLR had the lowest RMSE, followed by ANN, whereas the lowest MAE varied depending on the kind of clothing. Although MLR seemingly had the best performance in the training dataset for the four garment types, the testing data did not fit the model. During validation using the test dataset, negative R² was recognized for SVR, KNN, and MLR, except for ANN with improved R². Overall, ANN performance was stable in both train and test datasets for the four garment types with positive R², while the ANN structure is much more flexible in adjusting hidden units for improving performance. As a result, using ANN to predict the pattern parameters for the four garment types is significant and valid.

Table 3.

Performance matrix for training and evaluation (4 garment types).

	Models	Training			Testing
		RMSE	MAE	R2	RMSE	MAE	R2
Blazer	ANN (1HL)	0.0031	0.0023	0.8770	0.0012	0.0009	0.8886
	SVR	0.0121	0.0099	−0.9427	0.0117	0.0142	−16.3981
	KNN	0.0061	0.0046	0.5508	0.0085	0.0202	−11.2558
	MLR	0.0025	0.0020	0.9208	0.0374	0.0511	−87.4852
Long-sleeved shirt	ANN (1HL)	0.0070	0.0039	0.8866	0.0022	0.0012	0.9627
	SVR	0.0192	0.0007	−0.6219	0.03915	0.0012	−78.2865
	KNN	0.0102	0.0003	0.5495	0.0450	0.0010	−87.9119
	MLR	0.0040	0.0001	0.9308	0.1157	0.0022	−228.4546
Short-sleeved shirt	ANN (1HL)	0.0081	0.0042	0.8812	0.0038	0.0023	0.9787
	SVR	0.0203	0.0029	−0.5203	0.0638	0.0034	−14.7377
	KNN	0.0107	0.0015	0.5784	0.0615	0.0030	−14.0414
	MLR	0.0055	0.0008	0.8898	0.0811	0.0080	−15.8907
Trousers	ANN (1HL)	0.0032	0.0022	0.8992	0.0008	0.0005	0.9683
	SVR	0.0164	0.0151	−2.1174	0.0162	0.0158	−28.0462
	KNN	0.0057	0.0032	0.5957	0.0097	0.0283	−22.3827
	MLR	0.0025	0.0017	0.9263	0.0302	0.0328	−62.4016

Note: HL means hidden layer.

4.3. Artificial neural network prediction model for pattern parameters of four garment types

In the ANN prediction model, pattern parameters for four garment types would be projected for sketch drafting on behalf of the body measurements and psychological orientations of each subject. Two major components, B and P, were put into the model for evaluating the output elements, R, and the model was established with a non-linear relationship between the inputs and outputs. Predictions for pattern measurements of the four garment types (blazer, long-sleeved shirt, short-sleeved shirt, and trousers) were attained with the adjusted pattern parameters from the 120 subjects after their fit evaluation procedures. Table 5 summarizes the R² scores for the predictions of all subject data and the segmented data by psychological orientations, respectively, for each garment type. Changes in R² scores for each garment type are presented in Fig. 5, Fig. 6, Fig. 7, Fig. 8. In general, most of the training and testing data fit the 45° trend line for regression. The close-to-1 R² scores mean the predictions were nearly close to the actual values. The segmentation of participants by psychological orientation raised the R² scores in all garment types compared with the results reported from the overall data. The ANN prediction model is proven to be an effective way for apparel pattern sketch drafting when the designers have body dimensions and psychographic characteristics of their target market with the enhanced R² scores from the 5 different psychographic clusters. A customized individual fit is achieved, and the predicted data is practical for virtual fitting.

Table 5.

R² score comparison for overall and segmented groups of psychographic characteristics by garment types.

	Training Data (70 rows)	Testing Data (50 rows)
Blazer
Overall	0.94481	0.94862
Group 1 (Train: 15 rows; Test: 9 rows)	0.98580	0.99037
Group 2 (Train: 12 rows; Test: 9 rows)	0.99844	0.95525
Group 3 (Train: 11 rows; Test: 8 rows)	0.99058	0.98725
Group 4 (Train: 17 rows; Test: 16 rows)	0.99955	0.98965
Group 5 (Train: 15 rows; Test: 8 rows)	0.99724	0.99843
Long-sleeved shirt
Overall	0.99149	0.98970
Group 1 (Train: 15 rows; Test: 9 rows)	0.99755	0.99883
Group 2 (Train: 12 rows; Test: 9 rows)	0.99882	0.99272
Group 3 (Train: 11 rows; Test: 8 rows)	0.99933	0.99031
Group 4 (Train: 17 rows; Test: 16 rows)	0.99824	0.99513
Group 5 (Train: 15 rows; Test: 8 rows)	0.99899	0.99958
Short-sleeved shirt
Overall	0.99897	0.99045
Group 1 (Train: 15 rows; Test: 9 rows)	0.99975	0.99826
Group 2 (Train: 12 rows; Test: 9 rows)	0.99983	0.99970
Group 3 (Train: 11 rows; Test: 8 rows)	0.99865	0.99984
Group 4 (Train: 17 rows; Test: 16 rows)	0.99992	0.99901
Group 5 (Train: 15 rows; Test: 8 rows)	0.98149	0.99973
Trousers
Overall	0.99104	0.99515
Group 1 (Train: 15 rows; Test: 9 rows)	0.99815	0.99711
Group 2 (Train: 12 rows; Test: 9 rows)	0.99976	0.99573
Group 3 (Train: 11 rows; Test: 8 rows)	0.99780	0.99987
Group 4 (Train: 17 rows; Test: 16 rows)	0.99919	0.99799
Group 5 (Train: 15 rows; Test: 8 rows)	0.99957	0.99978

Fig. 5.

Changes in coefficients of determination for the garment type blazer.

Fig. 6.

Changes in coefficients of determination for the garment type long sleeves.

Fig. 7.

Changes in coefficients of determination for the garment type short sleeves.

Fig. 8.

Changes in coefficients of determination for the garment type trousers.

4.4. Ease allowance preferences for four garment types using ANN

Ease allowance preference charts (see Appendix 2) for each garment type are formed with the identified clusters of psychographic characteristics from the 120 subjects and the trained ANN virtual fitting preference prediction model. In most cases, P2 has a relatively greater ease allowance for the pattern parameters in each garment type, while P3 would rather have smaller ease allowances for the described pattern parameters. It may indicate that people with higher AIO involvement in spending their spare time on lifestyle would much prefer wearing just-fit or relatively tight garments. Different ease allowances were found in each garment type, though the correlated relationships among the AIO groups were positive (>0.99) for all garment types.

4.5. Discussion on the findings

The individuals were classified into five groups using k-means clustering in accordance with their psychographic traits, as shown in Table 4. The subjects in P5 were most concerned about the AIO items, which means they were much more willing to put time and effort into clothing selection. In contrast, the subjects in P1 were the least active on those AIO items. From the generated ease allowance preference charts in Appendix 2, P5 preferred larger ease allowances in blazers and long-sleeved shirts than P1, while in contrast, P5 chose smaller ease allowances in short-sleeved shirts. Comparatively, P5 would favour larger ease allowances in their garments, except for short-sleeved shirts in general. It is notable that among the groups, P2 has the largest ease allowances for the four garment types; nonetheless, the subjects in P2 were moderately inactive in the AIO items. Subjects’ willingness to devote time and effort to lifestyle and clothing selection may be affected by their salary, income, occupation, peer influence, and more. These factors are still unobservable and will be evaluated in further studies. Also, the 3D samples generated by the DCGAN model in Table 6, Table 7 were in good shape, so designers and customers can evaluate various editions of fitted bodies in slightly different body shapes. Overall, the ANN prediction model was significant for forecasting the pattern parameters for the four garment types, especially a 5% increase for blazer type after adding psychographic orientation in consideration. Eventually, with the psychographic orientations and anthropometric data of customers, designers and manufacturers in the apparel industry can easily identify the matching fit with the predicted pattern parameters. They can also interact with customers using real-time, realistic 3D visualization and allow customers to adjust the garment accordingly to satisfy their preferences.

Table 6.

3D visualization for small, medium, large sizes (female).

Table 7.

3D visualization for small, medium, large sizes (male).

4.6. Limitations

There are a few limitations to the study. The chosen outputs for each garment type, which are the pattern parameters, were defined by the commercial software, Optitex. Other software may use different pattern parameters to generate the 2D pattern sketches for these garment types, and the data from Optitex may not be compatible with other virtual garment simulation software. Also, the retrieved body dimensions from the subjects may include small variations since some subjects wore their clothes during the body scanning sessions.

5. Conclusion

This study proposed a systematic approach to recognizing ease allowance preferences with the assessments of 3D body measurements and psychological orientations of the consumers for four garment types: blazers, long-sleeved shirts, short-sleeved shirts, and trousers. The results demonstrated that the ease allowance preferences of the subjects were notably different from the prearranged values in the 3D virtual simulation software, and the Artificial Neural Network (ANN) is significant in modelling the nonlinear interaction between pattern parameters, psychographic characteristics, and body measurements. The quantitative correlation among 3D body dimensions, psychographic characteristics, and sketch pattern measurements achieves mass customization for individual fit in the apparel industry. It is easier to retrieve the best fit from consumers, and designers can create massive, personalized designs in a short time and with minimal effort. More fine-tuned ease allowance preference charts for the four assessed garment types were obtained. Despite the ANN model's performance, the stakeholders would benefit most from seeing 3D-fitted samples to interpret how the clothes fit and how they drape. The DCGAN model envisages immediate prototypes with the options of garment sizes, ease preferences, genders, and garment types. High-quality features, such as draping and basic accessories, were displayed in the 3D samples. In summary, the study presents a way to enhance communication in the supply chain, where stakeholders can have an advanced instrument to obtain comfortable garment fittings for their customers using their body dimensions and psychographic characteristics by generating new size-fitting data. Designers can also easily evaluate their designs with 3D-developed samples. A new level of end-user satisfaction from mass customization and a more efficient design process with a sense of customer orientation can be obtained.

Ethics approval statement

The study complies with all regulations and confirms that the experiments with human subjects were approved by the Human Subjects Ethics Sub-committee (HSESC) of Technological and Higher Education Institute of Hong Kong (HSESC No.: HE2021-07). Informed consent was obtained from each subject during the experiment.

Author contribution statement

Nga Yin Dik: Conceived and designed the experiments; Performed the experiments; Analyzed and interpreted the data; Contributed reagents, materials, analysis tools or data; Wrote the paper.

Paul Wai Kei Tsang; Ah Pun Chan: Conceived and designed the experiments; Performed the experiments; Contributed reagents, materials, analysis tools or data; Wrote the paper.

Chris K. Y. Lo: Conceived and designed the experiments; Analyzed and interpreted the data; Contributed reagents, materials, analysis tools or data; Wrote the paper.

Wai Ching Chu: Conceived and designed the experiments; Performed the experiments.

Data availability statement

Data will be made available on request.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix

Appendix 1.

Trend Graphs of Men and Women Apparel Revenue from 2014 to 2026.

Appendix 2.

Ease Allowance Preference Charts for Each Garment Type.

T	Pattern Parameters	P1	P2	P3	P4	P5
Blazer
1	Front Length	0.56	0.74	0.42	0.60	0.70
2	Back Length	0.45	0.59	0.33	0.48	0.56
3	Pocket Position	0.04	0.07	0.03	0.05	0.06
4	Front Waist	0.56	0.74	0.42	0.60	0.70
5	Side Waist	0.89	1.18	0.67	0.96	1.12
6	Back Waist	0.56	0.74	0.42	0.60	0.70
7	Front Shoulder Width	0.78	1.04	0.58	0.84	0.98
8	Front Chest Girth	0.67	0.89	0.50	0.72	0.84
9	Side Panel Underarm Width	1.12	1.48	0.83	1.20	1.40
10	Back Chest Girth	0.45	0.59	0.33	0.48	0.56
11	Back Width	0.64	0.84	0.48	0.69	0.80
12	Back Neck Width	0.22	0.30	0.17	0.24	0.28
13	Back Neck Height	0.11	0.15	0.08	0.12	0.14
14	Back Shoulder Length	0.56	0.74	0.42	0.60	0.70
15	Back Scye Depth	0.56	0.74	0.42	0.60	0.70
16	Front Scye Depth	0.56	0.74	0.42	0.60	0.70
17	Side Slim Dart	0.56	0.74	0.42	0.60	0.70
18	Centre Back Slim Dart	0.39	0.52	0.29	0.42	0.49
19	Upper Sleeve Height	0.44	0.58	0.33	0.47	0.54
20	Upper Sleeve Head Upper Width	0.58	0.77	0.43	0.62	0.73
21	Upper Sleeves Width	0.89	1.18	0.67	0.96	1.12
22	Lower Sleeve Width	1.11	1.47	0.83	1.19	1.38
23	Upper Sleeve Length	0.47	0.62	0.35	0.50	0.59
24	Lower Sleeve Length	0.30	0.40	0.23	0.32	0.38
25	Wrist Girth	0.30	0.77	0.20	0.48	0.64
Long sleeved shirt
1	Neck Drop- Back	0.08	0.09	0.06	0.08	0.08
2	Neck Drop- Front	0.09	0.10	0.07	0.09	0.09
3	Neck Width	0.61	0.75	0.47	0.64	0.62
4	Shoulder Length	0.30	0.36	0.23	0.31	0.30
5	Across Shoulder	1.22	1.49	0.93	1.28	1.24
6	Front Length from HPS	0.70	0.85	0.53	0.74	0.71
7	Back Length from CB	0.68	0.84	0.52	0.72	0.69
8	Across Chest (1″ Below AH)	4.87	5.97	3.72	5.13	4.96
9	Across Waist	4.87	5.97	3.72	5.13	4.96
10	Sweep	4.87	5.97	3.72	5.13	4.96
11	Sleeves Length	0.77	1.02	0.59	0.83	0.80
Short sleeved shirt
1	Neck Drop- Back	0.71	0.70	0.69	0.72	0.57
2	Neck Drop- Front	0.70	0.75	0.72	0.98	0.74
3	Neck- Width	2.00	2.06	2.02	1.97	1.65
4	Shoulder- Length	1.15	1.21	1.18	1.14	0.96
5	Across Shoulder	4.22	4.40	4.31	4.18	3.51
6	Front Length from HPS	6.61	6.34	6.66	6.41	5.54
7	Back length from CB	5.33	5.27	5.30	5.32	4.77
8	HPS to Underarm	0.80	1.42	0.86	1.35	1.31
9	Across Chest	4.08	7.16	4.39	6.50	6.59
10	Across Waist	4.06	7.13	4.37	6.32	6.56
11	Side- Length	0.30	0.54	0.33	0.52	0.49
12	Bottom Edge Opening (Sweep)	10.64	10.85	10.41	10.82	8.80
13	Sleeve Length from CB (3 Point Measure)	12.15	14.31	13.39	13.25	11.50
Trousers
1	Waist to Floor	0.38	0.78	0.36	0.56	0.64
2	Body Rise	0.46	0.91	0.43	0.65	0.75
3	Waist to Hip	0.38	0.78	0.36	0.56	0.64
4	Knee Position	0.07	0.12	0.07	0.10	0.11
5	Half Front Hip Measurement	0.80	1.46	0.75	1.08	1.24
6	Hip at Rise Level Girth	0.75	1.38	0.70	1.02	1.17
7	Half Front Waist Girth	0.60	0.87	0.55	0.70	0.79
8	Front Bottom Width	0.13	0.17	0.12	0.14	0.16
9	Front Knee Width	0.34	0.63	0.32	0.46	0.53
10	Half Back Waist	0.66	1.23	0.61	0.90	1.04
11	Hip Girth at Rise Level	0.85	1.62	0.80	1.18	1.35
12	Half Back Hip	0.70	1.31	0.65	0.95	1.10
13	Back Bottom Width	0.10	0.13	0.09	0.12	0.13
14	Back Knee Width	0.34	0.63	0.31	0.46	0.53
15	Back Pants Length	0.38	0.77	0.36	0.55	0.64

Appendix A. Supplementary data

The following are the Supplementary data to this article:

Multimedia component 1

Pre-Survey Questionnaire.

Multimedia component 2

Ethical Clearance for Research Project involving Human Subjects.