Journal of Physics:

Conference Series

PAPER • OPEN ACCESS You may also like

- A semi-supervised fault diagnosis method
Intelligent Clothing Interaction Design and for axial piston pump bearings based on
DCGAN
Evaluation System Based on DCGAN Image You He, Hesheng Tang, Yan Ren et al.

- Fault identification method based on

Generation Module and Kansei Engineering generative adversarial network in
distributed acoustic sensing
Ying Shang, Jiawen Wang, Sheng Huang
To cite this article: Lu Chen et al 2021 J. Phys.: Conf. Ser. 1790 012025 et al.

- Optimizing the classification ability of CNN

for SAR fully polarized radar data based
on DCGAN
ZeCong Bu, Yue Zhang and AnMai Zheng
View the article online for updates and enhancements.

This content was downloaded from IP address 154.64.227.31 on 10/11/2024 at 04:59

SIFTIC 2020 IOP Publishing
Journal of Physics: Conference Series 1790 (2021) 012025 doi:10.1088/1742-6596/1790/1/012025

Intelligent Clothing Interaction Design and Evaluation System

Based on DCGAN Image Generation Module and Kansei
Engineering

Lu Chen1,2, Qi Fang1,2 and Yu Chen1,2

*

1
Clothing Institute, Shanghai University of Engineering Science, Shanghai, 201620,
China

Abstract. The deep convolutional generative adversarial network (DCGAN) model can be
applied to personalized fashion design. Determining the best design from fashion sketches that
is of interest to customers is an important research field. In this study, we propose a rapid
evaluation method based on Kansei engineering to evaluate personalized fashion images
generated by DCGAN. The degree of customer interest in the generated fashion images is
determined by tracking the eye movements and facial expressions of customers. Images with
high number of fixations are added to the training data for iterative training; this aims to
improving customer satisfaction. The method was verified in an interaction design of
personalized black dresses. After several iterations, the customers’ satisfaction rating of the
generated images by DCGAN improved. This method can be applied to other styles of fashion
design.

1. Introduction
At present, research on the image generation model is closely related to the fashion image generation
technology. The image generation technology can be generally divided into two categories: generative
parametric approaches and adversarial approaches[1]. Various studies have shown that adversarial
approaches have clear advantages in image generation, especially the generative adversarial network
(GAN). The generation of deep convolutional GAN (DCGAN) algorithm[2] is a milestone in the
development of the GAN model, which determines a stable training structure to facilitate engineering
implementation.
In recent years, research on GANs has increased, and related research has made some progress. For
example, Phillip et al.[3] used the conditional GAN algorithm as an image conversion solution that
effectively implements the conversion of linear drawings to rendered images. Jun-Yan Zhu et al.[4] also
carried out a research on image style migration through the CycleGAN algorithm. Yoo D et al.[1] set a
converter in the generative model to generate an image of a person wearing clothes. Makkapati V et
al.[5] put forward related theories on the symmetry of generated fashion images. Although the
aforementioned research has made contributions to the application of deep learning in fashion design,
these studies have not considered users’ personalized needs. Therefore, generated fashion images
should be evaluated and selected from the perspective of customers, and explorations should be made
in terms of improving customer satisfaction.
Currently, in addition to traditional methods, Kansei engineering has also been widely used[6]. For
example, eye tracking[7] can reflect customers’ attraction on a certain product. The longer the fixation

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
SIFTIC 2020 IOP Publishing
Journal of Physics: Conference Series 1790 (2021) 012025 doi:10.1088/1742-6596/1790/1/012025

time, the more interested the customer is in the design[8]. In addition, facial expressions are the most
important, natural, and direct channels for people to express emotions[9] .
This study creatively applies deep learning algorithms to automatically generate fashion designs
and Kansei engineering for the evaluation of costumers’ preferences, achieving a fusion of Kansei
engineering and fashion design. This paper is organized as follows: In Section 2, the framework of the
system proposed in this study is presented. In Section 3, the experimental procedure and results are
presented to show the effectiveness of our method. In Sections 4 and 5, a discussion and future
outlook are given, respectively.

2. Framework of interactive design system

Figure 1 presents a schematic of the proposed interactive design system.

Figure 1. Proposed system.

The dashed box represents the DCGAN network that automatically generates design modules.
Outside the dashed box, the customer quickly selects the evaluation module. The specific steps are
described as follows:

2.1. Initialization training

We selected a certain number of fashion images from the browsing records of customers on shopping
websites or related web pages. After the data augmentation, the images were used to train the DCGAN
network as real training data.

2.2. Selection
After the training of the DCGAN network was completed, we used the DCGAN module to generate
the fashion design images. When the fashion designs were automatically and continuously played in
the form of a slideshow, the user showed her response to the related designs and screened out the
clothing designs that are positively attracted.

2.3. Replacement
The fashion design images that the customer selected through the interactive evaluation module were
used as new training data to replace part of the original training data after passing through the data
augmentation module, and the design images not selected were discarded.

2.4. Iterative training

After the training data were updated, DCGAN underwent an iterative training. After the training was
completed, a new generation of DCGAN generated new design images again. The customer’s
preferred images were selected by the interactive evaluation module while viewing slides. Then, the
selected images replaced some of the training date again for iterative training.

3. Experiment

2
SIFTIC 2020 IOP Publishing
Journal of Physics: Conference Series 1790 (2021) 012025 doi:10.1088/1742-6596/1790/1/012025

This study used little black dress images as a training data set. The specific processes of the
experiment are presented as follows:

3.1. Procedure

3.1.1. Construction of the DCGAN model to generate little black dresses DCGAN already has a ready-
made code in Tensorflow:carpedm20/DCGAN-tensorflow; we used this code directly. We found a few
lines of code to read data, as shown in Figure 2. Then, we created a new anime folder in the data folder,
placed little black dress images into this folder, and specified the dataset anime at runtime. The
training environment is win10+cuda9.2+Tensorflow.

Figure 2. Data-read code.

3.1.2. Generate personalized little black dress designs We selected two female college students, aged
23 and 25 years old, as the experimental subjects, and named them S1 and S2 hereinafter, respectively.
For each customer, we obtained 20 little black dress images from their recent online browsing records
and added the images to the training data. Then, we let them choose their top 20 favorite little black
dresses to be added into the training data. The last 60 little black dress images, which were randomly
selected from the database, were added to the training data. The 100 little black dresses were
augmented to 1,000 images and normalized to a 28 × 28 × 1 size. Figure 3 presents the selection of the
initialization training data set.

Figure 3. Initialized training data source distribution.

In this manner, each of the two customer has their own training data set. Accordingly, we named
two groups of modules, i.e., DCGAN-S1 and DCGAN-S2, as shown in Figure 4. The training
environment is win10 + cuda9.2 + Tensorflow.

Figure 4. Modules of each customer..

3
SIFTIC 2020 IOP Publishing
Journal of Physics: Conference Series 1790 (2021) 012025 doi:10.1088/1742-6596/1790/1/012025

After the training, we randomly selected 25 images through the DCGAN network, removed five
pictures with inferior pictures, and made the remaining 20 pictures into continuous slides. The Tobii
eye tracker was used to collect the eye tracking data of customer S1, whereas the FaceReader software
(launched by VicarVision and Noldus Information Technology) was used to measure and analyze the
facial expressions of customer S2. Before the experiment, the customers were informed that for the
research study, expression video materials need to be collected; they were ensured of data
confidentiality. We recorded the number of designs that the customers tended to positively select (i.e.,
customers’ favorite designs) in each round, as shown in Figures 5 and 6.The selected design images
replaced some of the training images.

Figure 5. Customer S1’s preferred design

Figure 6. Customer S2’s preferred design images.
images.

3.2. Results
Table 1 presents the number of images selected by each experimental subject through the interactive
evaluation module in the first four rounds. The table shows that after three rounds of iterative training
of the DCGAN, the proportion of selected generated images increased.

Table 1. Selected images and percentage of selection (PoS) after four rounds of
iterative training of the two customers.

S1 Pos S2 Pos

Round 0 3/20 15% 2/20 10%

Round 1 5/20 25% 3/20 15%
Round 2 6/20 30% 3/20 15%
Round 3 4/20 20% 5/20 25%

4. Discussion and summary

This study proposes a intelligent clothing interaction design and evaluation system based on the
DCGAN image generation module and Kansei engineering, which creatively allows customers to
participate in the design process.We performed experiments to verify the effectiveness of the method.
From the results shown in Table 1, after three rounds of iterative training, the number of small black
dress designs selected by customer S1 increased from 15% to 20%, whereas that selected by customer
S2 increased from 10% to 25%. However, due to the limitation of equipment and technology, only
four rounds of experiments were performed in the present study; only two customers participated; and
no large-scale experiments were implemented. It is expected that with multiple rounds of experiments
and higher number of participants, the customers’ satisfaction ratio will increase, and the performance
of the entire system will greatly improve.

5. Future outlook
This research creatively applies deep learning algorithms to automatically generate clothing designs,
achieving a fusion between Kansei engineering and fashion design. In the future, more experiments
should be performed to verify the effectiveness of this system. In addition, the following
improvements will be made:the specific effects of eye tracking and facial expression analysis on

4
SIFTIC 2020 IOP Publishing
Journal of Physics: Conference Series 1790 (2021) 012025 doi:10.1088/1742-6596/1790/1/012025

customers’ fashion design preferences should further examined;training data should be updated in a
timely manner;customers’ selected design images should be used to make real clothing; fashion
designers will benefit from this type of system.

Acknowledgments
2
This research was funded by the Program for Professor of Special Appointment (Eastern Scholar) at
Shanghai Institutions of Higher Learning (Grant No. TP2017074).

Two students joined in the experiment, S1 Qi FANG, S2 Chongchong LEI

References
[1] Yoo D, Kim N, Park S, et al. Pixel-level domain transfer. European Conf. on Computer
Vision. 2016 (Springer, Cham)
[2] Radford A, Metz L and Chintala S. Unsupervised Representation Learning with Deep
Convolutional Generative Adversarial Networks 2015 J. Computer Science
[3] Isola P, Zhu J Y, Zhou T, et al. Image-to-Image Translation with Conditional Adversarial
Networks IEEE Conf. on computer vision and pattern recognition. 2017
[4] Zhu J Y, Park T, Isola P, et al. Unpaired Image-to-Image Translation using Cycle-Consistent
Adversarial Networks IEEE Conf. on computer vision and pattern recognition. 2017
[5] Makkapati V and Patro A. Enhancing Symmetry in GAN Generated Fashion Images
International Conf. on Innovative Techniques and Applications of Artificial Intelligence.
2017 (Springer, Cham)
[6] Li F, Liu X G and Gu W. Evaluation process of costume design based on evaluation subject
perspective 2017 J. Journal of Textile
[7] Li Zh, Gou B C, Chu J J, et al. Way of getting user requirements based on eye tracking
technology. 2015 J. Computer Engineering and Applications 51 9 233–7
[8] Jiang Q, Peng L, Gao Y K, Yang X Y and Huang W Extraction and research oncultural
characteristics of Longhorn Miao costume and adornment 2019 J. Modern Electronics
Technique 42 14 135–9
[9] Keltner D and Ekman P. 2010 Facial Expression of Emotion Handbook of Emotions 3rd.
(New York: The Guilford Press) pp 173–83