0% found this document useful (0 votes)
57 views6 pages

Wick Rama Rath Ne 2019

Uploaded by

Pratham Patel
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
57 views6 pages

Wick Rama Rath Ne 2019

Uploaded by

Pratham Patel
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

2019 International Seminar on Research of Information Technology and Intelligent Systems (ISRITI)

TrendiTex: An Intelligent Fashion Designer


Minoli De Silva Chathurangi Weerasinghe
Poorni Wickramarathne
Dept.of Information Technology Dept.of Information Technology
Dept.of Information Technology
Sri Lanka Institute of Information Sri Lanka Institute of Information
Sri Lanka Institute of Information
Technology Technology
Technology
Colombo, Sri Lanka Colombo, Sri Lanka
Colombo, Sri Lanka
minoligeethma@gmail.com hansiweerasinghe95@gmail.com
poorniwickramarathne@gmail.com
Dr. Pradeep Abeygunawardhana Ms. Suranjini Silva
Heshani Nanayakkara
Dept.of Information Systems Dept.of Information Systems
Dept.of Information Technology
Engineering Engineering
Sri Lanka Institute of Information
Sri Lanka Institute of Information Sri Lanka Institute of Information
Technology
Technology Technology
Colombo, Sri Lanka
Colombo, Sri Lanka Colombo, Sri Lanka
heshaninanayakkara@gmail.com
pradeep.a@sliit.lk suranjini.s@sliit.lk

Abstract— In a highly changing technical era, Intelligent In the proposed system “TrendiTex”, our main objective
Fashion Designing systems play a key role to bridge the gap is to develop an advanced intelligent fashion designer to
between fashion designers and the customers. Most of the design a unique cloth according to the customer specific
people specially females, are fond of fashion. Currently, expectations and provide the user with the facility to make
fashion has become a way of defining a person’s preferences the best use of the fashion designing system. This paper
and personality. Analyzing through a large number of fashion focuses on the proposed system’s implementation which will
trends and selecting a one among them will be a highly time- get the user’s preferences along with measurement details,
consuming task. Even though most of the people are keen on event time, event type, color preference, undertone and skin
fashion, with their busy schedules, spending time on selecting a
complexion. It also extracts user preferences and predicting
cloth for an occasion among numerous numbers of designs
available is a hard task. Therefore, it would be exhausting to
personal body shape, building the unique model according to
select a proper design for an occasion for them. Prevailing the those preference details and giving user the augmented
difficulty in finding the clothes up to the user’s expectation, we reality facility to try on the newly generated fashion design.
propose a user-friendly fashion designing mobile application When getting user preference details, user will be asked to
and a web application called “TrendiTex”. Extracting user select any image which will be prompted randomly to get the
preference details, user’s body shape predicting and user’s preferred color by color optimization. Those
recommending trending fashion designs according to their preference details will be extracted for predicting the user’s
shape, generating the unique 2D new fashionable design for a personal body shape and relevant to that body shape, few
specific event and the augmented fit-on facility are fashion designs will be prompted for the user to select the
implemented in TrendiTex. This system represents an efficient most preferred design among them. Considering all those
approach to design new unique products according to user’s user’s preference details, the unique fashion design will be
preferences and gives augmented fit-on facility. generated and displayed for the user. User can wear the
generated fashion design model using Augmented Reality
Keywords—Machine Learning, Image Processing, trial room implemented in this system.
Classification, Augmented Reality (AR)
The remainder of this paper is structured as follow. In
I. INTRODUCTION section II literature review, we explain the previous
approaches and existing systems in the domain of intelligent
Artificial Intelligence (AI) based systems are becoming fashion designer. In section III we discussed the
more efficient for the fashion designers at present, since methodology which will focus on the implementation of
these systems can identify people’s preferences and handle proposed system and the substantial components. Section IV
more attributes and process large volumes of data faster than explained the results and discussion. Finally, in section V,
a human stylist and it has the ability to learn a person’s style we conclude the paper and discuss our future works.
more accurately. These systems can memorize a person’s
preferences and accept feedbacks and process accordingly.
II. LITERATURE REVIEW
At present, hiring a fashion design expert is considerably cost-
intensive. Also, it’s the greatest challenging task to apply AI Currently, multiple systems on AI fashion domain have
in a field which uses human creativity very much and to been created. Fashion Design applications are used to bridge
create a design which will be accepted by the user because a the gap between human and the machine. IBM’s AI image
person’s choice depends on many variables such as search engine [1] is a large database with a large number of
individual uniqueness, cultural dependence, individual images of various fashion designs created so far. Users can
distinctiveness and adoption preference. search a dress using photograph of a dress to check whether
it already exists in the database or not. This system can only
In existing intelligent fashion designing applications, be used to check whether a particular model was previously
matching color selection based on the user’s skin complexion produced or not.
and undertone (The undertone implies the color which is
underneath the skin of a person which affects to the overall Amazon has also created their own AI fashion tool [2] to
hue), clothing recommendation for a personal body shape, help them in making new garments for their retails,
generating unique completed wearable designs and developing its own clothing brands, giving its customers the
augmented fit on facility are not available.
try-on facility and also an Echo Look app which provides

978-1-7281-4520-4/19/$31.00 ©2019 IEEE 505


users feedbacks on their outfits. Furthermore, they have
developed a machine learning algorithm to check whether a
designed product is considered to be stylish or not by
analyzing a few labels attached to images. That is an
innovative approach since most of the time the images
attached to social media will be only having a one label even
though the computers are used to work with extensive
labeling. The software would provide recommendations for
adjustments as well. Further, they developed an algorithm
which learns about a style from any image and then generate
new designs similar to that particular style. This process
involves deep neural networks, Generative Adversarial
Network (GAN). The design innovativeness lacking is the
major problem they face.

StitchFix [3] is another fashion design system which Fig. 1. High level architecture of the system
generates the outline of the garments. The produced
garments are to be filled with colors and patterns. System is A. User Preference Extraction and Personalized Color
prompted to take the direct user input regarding what colors Palette Generation
and patterns and other features they want in their dress.
For identifying the related body shape of a specific
Google’s Project Muze [4] is another project has been
user, height (in feet) and UK standard body size for
done in this particular domain. In this project, a design is
dresses [6] were used. Waist, bust, hip details were
created taking minimal inputs from a user. Parameters like
derived accordingly from the standard UK size chart
age, mood, tastes, gender are being used for the creation of
and used in determining the user’s body shape.
new design. These designs are created using Generative
Adversarial Networks but the designs lack originality. Some
TABLE I. WOMEN’S BODY MEASUREMENTS TABLE
of the generated pieces are not suitable for actual wearing FOR DRESSES [6-8]
and therefore, it needs to be improved.
UK 6 8 10 12 14 16 18 20
Facebook’s Artificially Intelligent Fashion Designer [5] size
is another project done in this domain. There, some of the
Bust 32 33 34 36 38 40 41 44
created designs are disjoint from sleeves and some are created
with large holes in them. That is hardly to be considered as a Waist 25 26 28 30 31 33 34 37

successful outcome. Facebook’s effort on making a fashion Hips 34 36 37 39 41 43 44 46

designer tool is considered to be not much successful


because of the incompleteness of its output designs. User’s undertone and skin complexion are major factors
related to matching the suitable dress color for a specific
III. METHODOLOGY occasion. One of the three undertones are cool, warm and
neutral are identified [2]. It is important to know the person’s
The proposed system “TrendiTex” consists of four main undertone because the system takes that as a user input to
components. The four components of the proposed system output the suitable colors from a wide range of colors which
are categorized depending on the undertone.
depend on each other. In brief, the User Preferences
Extracting module takes the user inputs including user’s Matching undertone with the occasion type and time
measurements (UK size and height), event type, event time, process is done next. Different set of colors are related with
undertone, skin complexion and user’s preferred color. different occasions. First, the suitable colors for that specific
occasion should be identified. Fashion designers and fashion
Among those inputs height and UK size are the inputs to the
experts have strong knowledge in this when it comes to
Body Shape Predicting component. The extracted user’s occasion specific color matching. From the ideas collected
preferred color and the user’s selected trending fashion from them, the most suitable range of colors are identified
design are the inputs to the 2D Model Building component. occasion wise.
The color converted uniquely generated fashion design is the
Once the undertone is specified by the user, then suitable
input to the Augmented Fit-on component and user’s fit-on dress colors for that undertone are identified. Likewise,
process is the final output of the system. System is suitable dress colors for the occasion are also identified.
representing as a web application. Although suitable dress colors for these two factors are
identified separately, to find a set of colors which fulfills
The following figure 1, presents an overall system both of these criteria, a color comparison has to be done
architecture for a better explanation of the processing of the between these two set of colors which have been selected by
system. the system. That is done by using the CIELAB [9] color
model. Lab color space is used to visualize colors in a 3

506
Dimensional form where L for lightness and A and B for
color dimensions. First, the input RGB (Red, Green, Blue)
[10] colors are mapped to Lab color space and then color
difference in visual perception is calculated using deltaE2000
metric [11].

Where a hue rotation term (RT), to deal with the


problematic blue region (hue angles in the neighborhood of
275°)
L- Lightness, C-Chroma, H- Hue Compensation for
neutral colors (the primed values in the L*C*h differences) -
Compensation for Lightness (SL), Compensation for
Chroma (SC), Compensation for Hue (SH).
After calculating the difference values for colors taking
two at a time, colors with the lower difference values are
identified as more similar colors. Those colors are later used Fig. 2. Genetic Algorithm for selecting contrasting colors
for arranging the color palette for the suggestions.
Extracting user’s preference colors process involves B. Clothing Recommendation for Personal Body Shape
identifying the prominent colors in a user selected picture
and determining user preference colors in an indirect manner.
Body shape is an important thing to consider
Identifying user’s preferred color is important but that before choosing the perfect clothing style. Therefore, first
alone is not enough for suggesting the most suitable colors identify the user’s personal body shape using the users
for the user. Even though a user prefers some color, that body measurements which are bust, waist and hip.
specific color itself might not be matching for the user to Basically, there are five basic body types. [12] .
wear to an occasion. Some other variation of that color might i. Rectangle - Straight up & down proportions with
go well with the user’s skin complexion. Therefore, by very little waist definitions. Bust, waist and
finding those colors which are close to user preferred colors hip measurements are within close range.
and also matches to the skin complexion as well, a more ii. Triangle - Broader hips than shoulders. Hip
personalized color suggestion can be given to user. For that, measurement is more than 5% bigger than
a genetic algorithm is used to generate new colors from those bust measurement.
prominent colors in the picture. iii. Hourglass - full bust & full hips with waist
The best contrasting colors for the user is selected definition. Waist is at least 25% smaller than
considering the skin complexion and user preference colors hip and bust measurement.
extracted previously from image. These are the inputs to the iv. Inverted triangle - broad shoulders & small hips.
genetic algorithm which are: four prominent colors extracted Bust measurement is more than 5% bigger
from the image along with the user skin complexion color. than hip measurement.
v. Round - full bust & full hips without waist
User skin complexion is taken as a direct input to the
system where user has to select the most similar color to their definitions. Hip measurement is more than
complexion from the given skin colors by the system. The 5% smaller than bust measurement.
genetic algorithm takes all colors in RGB format. Red,
Green, Blue components in those colors are taken as the After identifying user’s personal body shape,
DNAs in a chromosome. Fitness function is calculated by trending clothing images are recommended for the user
taking the difference between RGB components for according to their body shape. For that system have to be
respective skin color and black color which is (0,0,0). updated on latest fashion trends. Using social media like
Likewise, from the generated colors, colors with the highest Twitter, latest trending clothing items are extracted. Web
fitness values are selected and used to create the personalized scraping is done using python to extract fashion trending
color palette. images from social media in Real-time. To prepare the
dataset, Web Scraping is used to get fashion images from
the websites since they have most popular female
celebrities’ records including their body shapes and
fashion designs which they prefer.
For dress identification- a data sets which includes
1000 dress images, for event classification – data set that
include 2000 images which fall into Casual, Office, Party,
Wedding dress categories and for body shape
classification – 2000 images belong to five basic body
shapes. As fashion trending images are extracted from

507
social media, there can be images with some similar
features of clothes. For that, duplicate removing is
performed.
There are three classification models which used to
train the neural networks. One for identifying the dress
images from the extracted fashion trending images using
social media and to classify dress images according to the
event types (Casual,Party,Wedding,Office) and another Fig. 5. Sigmoid function [15]
one for classifying dress images according to the matching
body shapes. Once identified dress images from the extracted social
media images, they are classified according to event types
The CNN model is built with multiple dense layers. such as wedding dresses, party dresses, office wear, casual
There are three Conv2D and two MaxPooling2Dlayers. wear. Then those dress images are classified further
And also the ReLU is used as an additional operation. according to the suitable body shapes. The classified dress
images are stored separately in the database. Then the
fashion images are retrieved from the Database which are
relevant to the user input (user’s body shape and event
type).

C. 2D Model Building Component


This module is responsible for generating the 2D
image of the unique fashion design model for the user
according to their preferences. Reimagine Retail was a
Fig. 3. ReLu Function [13]
powerful example of what happens when fashion partners
Neural network is built using, convolution, pooling with a global tech leader to advance challenging
and flattening. innovations – Michael Ferraro, director of the FIT/Infor
Convolution is used to extract features from an image. DTech Lab [16]. In brief Model Building module takes the
It maintains the relationship between pixels by learning particular user’s most desired fashion image selected by
image features. For that it uses small squares of input data. the user from the Body Shape Predicting module and the
In Pooling layer reduces the dimensionality of each best fit color from Extracting User Preferences module as
feature map but retains the most important information. In the main inputs and then process them to design the unique
Max Pooling, it takes the largest element from the feature design by preserving all the user’s preference details.
map within that window. Therefore, user doesn’t need to face any disappointments
After pooling, flattening is done to convert the data with the result as it is up to the relevant user’s expectation.
into a 1-dimensional array to input for next layer. The
Image colorizing process involves changing the main
output of the convolutional layers used to create a single
long feature vector. And it is connected to the fully- input image’s color according to the color selected by the user
connected layer which is the final classification model. which is extracted from the User Preferences Extracting
component. Without affecting to the dress pattern, complete
design’s color will be converted to the given input color.
Before coloring the input design, the original image is
converted to grayscale and that converted gray color image is
again the subject to color converting process. That output
image is sent to design the unique fashion design.
Designing the unique design process involves redesigning
the main input image which was selected by the user. In this
Fig. 4. High level architecture of CNN [14] process, several standard sleeve pattern types and neckline
The matrix is regenerated into a linear array therefore, patterns are changed randomly. Sleeve pattern types include
to input it into the nodes of the neural network. to puff sleeve, cap sleeve, flutte, bishop and the bell patterns.
initialize the CNN used below steps. Although neckline pattern types include sweetheart,
i. Convolution convertible and the peter pan patterns. New designs are
ii. Pooling prompting to the user by changing the fashion design
iii. Flattening randomly until user is agreeing with the unique design. When
designing the unique fashion design, fifteen several designs
iv. Full Connection
are being prompted randomly by the system. After the user’s
v. Compile the CNN
agreeing process, that unique design is moved to the next
component for the user by giving augmented fit-on capability.
Sigmoid function is used as an activation function for That output 2D image will be sent to the next component as
the last layer. its main input.

508
This component’s output is presented as a color converted ordinates. The transformation matrix can be also used to get
uniquely designed 2D fashion design which is designed the coordinates of a point after rotation about the center of
according to the user’s specific preferences. the image. The transformation matrix for rotating a point by
θ looks like:
D. Augmented Trial Room
This module augments the garment design that is created [αβ(1−α) ⋅center.x−β⋅center. y+ (new. width/2−centre.x)
through the AI. This module considers the 2D images of the −βαβ⋅center. x+(1−α) ⋅center.y+(new.height/2−centre.y)]
unique garment that was created. The user gets the trial
experience in front of the web cam. The output of this component is the augmentation of the
uniquely designed garment on to the user.
First, detecting the edges of the user’s from image frame
has to be done. In order to extract a user from the video IV. RESULTS AND DISCUSSION
stream provided by an input device (Microsoft Kinect, Web
Camera), different researchers have used different This section describes the outcomes of the developed
approaches. For example, Research [17] that was conducted system.
by F.Isikdogan and G.Kara used a method which, first The prominent colors from a user selected image are
extract and isolate the user from the background to create an extracted as in below (Fig. 6) and considering the specific
augmented reality environment. To segment the foreground, user’s skin complexion, a random number (range of 3-10) of
they have used the depth images and user labels which are best fit colors for that skin complexion are generated (Fig.
provided by the Kinect sensor. Then detect skin color and 7).
bring it to the front layer to allow the user to fold arms or
hold hands in front of the garment model. And they have
threshold the image in YCbCr color space, using the values
that were found to be effective for skin-color segmentation,
as follows,
Y = 0.299R + 0.587G + 0.114B < 70
77 <Cb = 128 − 0.169R − 0.332G + 0.5B < 127
133 <Cr = 128 + 0.5R − 0.419G − 0.081B < 173.
Fig. 7. Generated
Fig. 6. Prominent contrasting colors
YCbCr is a family of color spaces used as a part of the colors generated
color image pipeline in video and digital photography from the image
systems. Y represents luma component and CB and CR
represent the blue-difference and red-difference Chroma Then, the system recommends few clothes for the user
components [9] respectively. Other than that, in order to that match well with her specific body shape. From that the
prevent any background pixel from being labelled as skin, user can select the most fashionable dress type that she
they have applied the threshold only on the segmented needs to redesign for her occasion.
foreground.

Superimposing cloths on the user’s body is done next.


VFR applications allow users to see how the selected cloth
may look like on his/her body without putting it on while
standing in the in front of the system. The simplest way to
superimpose clothes on a user is, first capture the user who
is standing in front of a VFR using the Kinect sensor and get
his/her skeleton data. Finally, overlay the selected cloth on
to the skeleton by using necessary points of the skeleton. For
example, to map a t-shirt, it is required to get shoulder and Fig. 8. Dress identification and classification
hip points, to map a handbag, it is required to get wrist left according to event and body shape
and right points. When designing the unique design, user can select the
most suitable dress type which is generating randomly
As the final step, fast rendering is done with the through fifteen design types. Until the user is satisfied with
optimized architecture. The transformation matrix is a the generated design, new designs are prompting to the user.
matrix, which is multiplied by [x y 1] where (x, y) are co-
ordinates of the point. The idea of having a 1 is to facilitate
shearing. Multiplying a 2 x 3 matrix with a 3 x 1 matrix
leaves us with a 2 x 1 matrix containing the new point co-

509
REFERENCES
[1] @EmilyMatchar, F. (2019). Artificial Intelligence Could Help
Generate the Next Big Fashion Trends. [online] Smithsonian.
Available at: https://www.smithsonianmag.com/innovation/artificial-
intelligence-could- help-generate-next-big-fashion-trends-
180968952/ [Accessed 5 Aug. 2019].

[2] W. Knight, "Amazon Has Developed an AI Fashion Designer",


MIT Technology Review, 2017.

[3] Algorithms-tour.stitchfix.com. (2019). Stitch Fix Algorithms


Fig. 9. 2D dress design generation Tour. [online] Available at: https://algorithms-
tour.stitchfix.com/#client- experience-part-1 [Accessed 5 Aug.
2019].
A: Input image which was selected by the user in body
shape predicting module [4] Google. (2019). Project Muze: Fashion inspired by you, designed
B: Color converted uniquely designed new design by code. [online] Available at: https://blog.google/around-the-
which satisfied the user globe/google- europe/project-muze-fashion-inspired-by-you/
[Accessed 5 Aug. 2019].
In augmented reality module, user can try-on their
uniquely designed dress. User only needs to stand in front [5] R. McCall, "Facebook Has Developed An Artificially Intelligent
of their web camera to fit-on the uniquely designed dress Fashion Designer (And It's Bad)", IFLScience, 2018.
which is given by the system. Using edge detection user’s [6] "Clothing sizes", En.wikipedia.org, 2019. [Online]. Available:
body is identified. https://en.wikipedia.org/wiki/Clothing_sizes. [Accessed 5 Aug-
2019].

[7] "Size Chart | Yours Clothing", Yours Clothing, 2019. [Online].


Available: https://www.yoursclothing.co.uk/size-chart. [Accessed 5
Aug. 2019].

[8] "Women's Clothing Size Guide | Ladies' Size Chart | H&M UK",
H&M, 2019. [Online]. Available:
https://www2.hm.com/en_gb/customer-
service/sizeguide/ladies.html. [Accessed 5 Aug. 2019].

[9] “CIELAB color space,” Wikipedia, 03-Aug-2019. [Online].


Available: https://en.wikipedia.org/wiki/CIELAB_color_space.
Fig. 10. Edge detection Fig. 11. Augmenting the dress
[10] "RGB color model", En.wikipedia.org, 2019. [Online].
Considering the development of the artificial intelligence Available: https://en.wikipedia.org/wiki/RGB_color_model.
[Accessed: 05- Aug- 2019].
world, intelligent fashion designing applications are getting
famous day by day. TrendiTex has proven that a user can [11]"Lab Color", Mathworks.com, 2019. [Online]. Available:
select a unique design for their relevant occasion with a short https://www.mathworks.com/discovery/lab-color.html.
time. TrendiTex comes as a web application and a mobile
[12] Agyare, P. (2019). Five Female Body Types. [online] Ghafla!
application for its users. Ghana. Available at: http://www.ghafla.com/gh/five-female-body-
types/ [Accessed 26 Apr. 2019].

V. CONCLUSION AND FUTURE WORK [13] Medium. (2019). ReLU : Not a Differentiable Function: Why
used in Gradient Based Optimization? and Other Generalizations of
The system can be enhanced in a way that users can ReLU..[online]Availableat:https://medium.com/@kanchansarkar/rel
customize the lower part of the dress as well. Texture u-not-a-differentiable-function-why-used-in-gradient-based-
selection could be an added advantage for this kind of optimization-7fef3a4cecec [Accessed 15 Oct. 2019].
system. Since the system is currently focused only about
[14] Medium. (2019). A Comprehensive Guide to Convolutional
female dresses, for more clothing items such as blouses, Neural Networks — the ELI5 way. [online] Available at:
skirts, pants can be considered. In future, the system can be https://towardsdatascience.com/a-comprehensive-guide-to-
extended to reach male audience as well from providing the convolutional-neural-networks-the-eli5-way-3bd2b1164a53
opportunity to generate their personalized unique designs for [Accessed 15 Oct. 2019].
males as well. [15] Medium. (2019). Activation Functions in Neural Networks.
[online] Available at: https://towardsdatascience.com/activation-
ACKNOWLEDGMENT functions-neural-networks-1cbd9f8d91d6 [Accessed 15 Oct. 2019].
The authors would like to convey their gratitude to all the
[16] M. Bagalso and M. Bagalso, "IBM and FIT Announce
individuals who have contributed, either directly or indirectly Collaboration to Help Build the Creative Fashion Workforce of the
to the effective completion of this research paper. The Future — I-COM", I-COM, 2019. [Online]. Available:
original authors own the referenced figures used for the https://www.i-com.org/news-articles/ibm-and-fit-announce-
paper. collaboration-to-help-build-the-creative-fashion-workforce-of-the-
future. [Accessed: 25- Nov- 2019].

[17]I. F and K. G, A Real Time Virtual Dressing Room Application


using Kinect. COMPUTER VISION COURSE PROJECT,
BOGAZICI UNIVERSITY, 2012.

510

You might also like