INTERACTIVE SKETCH DESIGN RECOGNITION SYSTEM

USING EVOLUTIONARY TECHNIQUES

P. Y. Mok*, X. X. Wang, J. T. Fan, Y. L. Kwok and John H. Xin

Institute of Textiles and Clothing, The Hong Kong Polytechnic University, Hunghom, Hong Kong

Keywords: Sketch design recognition, Interactive genetic algorithms.

Abstract: In this study, a methodology that hybridizes a sketch design recognition approach with an interactive

genetic algorithm is proposed to help laypersons get clothes reflecting their preference. The sketch design

recognition approach consists of a composite description model, a sketch recognition method and a database.

First, a composite description model based on the knowledge of fashion design is developed to describe the

characteristics of a skirt. Second, a sketch recognition method is used to help laypersons get satisfied clothes.

Third, a database contains general elements about skirts. Moreover, an interactive genetic algorithm (IGA)

is used to accelerate the sketch recognition process. The subjective experiments results demonstrate that the

proposed method outperforms the existing fashion design systems.

1 INTRODUCTION

Holland (1975) first proposed genetic algorithms

(GA) in 1975 as computer programs that mimic

Darwinian’s evolutionary processes in nature.

Genetic algorithms (GAs) are stochastic search

algorithms based on the mechanism of natural

selection and natural genetics (Goldberg, 1989).

Since GAs lack the capability to utilize human

intuition and emotion appropriately, it is hard to

implement them in creative applications such as

architecture, art, music, and design. In addition, it is

difficult to evaluate the fitness due to no clear

measure. In order to overcome its drawback, it is

necessary to introduce an approach called interactive

genetic algorithm (IGA) (Takagi, 2001). A review of

research efforts related to IGA was summarized by

Takagi (2001). It provides two advantages over GA:

the first one is that it performs optimization with

human evaluation, and the second one is that it

reflects personal preference. Therefore, IGA is a

good solution in fashion design.

Although IGA has some advantages,

developing a fashion design system based on the

techniques of IGA is a complex task. As a result,

only a few attempts to develop garment style design

system using IGA were reported in the literature. For

instance, Inui (1996) constructed an apparel design

system, in which combined genetic algorithm with

apparel computer aided design system to produce

apparel design that the system users prefer.

Nakanishi (1996) developed a fashion design aid

system using genetic programming, which evolved

each dress design according to the user’s selection.

However, most of its productions were impractical

designs because they did not consider domain-

specific knowledge. To solve this problem, Kim and

Cho (2000) proposed an effective human-oriented

evolutionary system based on the knowledge of

fashion design to encode genotype with OpenGL

design models in order to produce more realistic and

reasonable design. Unlike traditional approaches that

attempt to model the dress design by several spline

curves, Kim and Cho (2000) provided a new

encoding scheme to describe a dress with three

parts: neck and body, sleeve, and skirt. Cho (2002)

also used a new encoding scheme to fashion design

system. Sugahara, Miki and Hiroyasu

(2008)

proposed a yukata design system that adopts IGA to

create a yukata that accommodates user’s taste.

Ogata and Onisawa (2008) designed a cloth design

support system based on IGA considering the factors

including shape, color and material in order to help

laypersons design clothes reflecting their Kansei and

help them get unexpected design candidates by only

evaluating candidates.

In this study, we develop a new sketch design

recognition system which consists of a sketch

recognition approach and an interactive evolutionary

Y. Mok P., X. Wang X., T. Fan J., L. Kwok Y. and H. Xin J..

INTERACTIVE SKETCH DESIGN RECOGNITION SYSTEM USING EVOLUTIONARY TECHNIQUES.

DOI: 10.5220/0003452900670072

In Proceedings of the 8th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2011), pages 67-72

ISBN: 978-989-8425-74-4

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

strategy. This system needs to be not only easy for a

layperson to use, but also needs to fast converges to

the target. The remainder of the paper is organized

as follows. Firstly, a brief description of the sketch

design recognition problems is given in section 1,

and a new methodology is outlined in section 2.

Secondly, a sketch recognition approach is presented

in detail in section 3. Thirdly, in section 4, an

interactive evolutionary strategy is used to accelerate

the sketch recognition process. Fourthly, the

effectiveness of the proposed methodology is

illustrated in section 5. Finally, conclusions are

summarized in section 6.

2 METHODOLOGY

In this paper, as shown in Figure 1, a general

methodology which combines a sketch design

recognition approach and an interactive genetic

algorithm is proposed to help laypersons design

clothes reflecting their preference. Furthermore, we

use skirts (Independent clothes covering the lower

half of one’s body, or a part of clothes under

waistline) (Kim and Cho, 2000) as an example to

illustrate the proposed sketch design recognition

system. The sketch design recognition approach

consists of a composite description model, a sketch

recognition method and a database. A composite

description model based on the knowledge of

fashion design is developed to describe the

characteristics of a skirt. A sketch recognition

method is used to help laypersons get satisfied

clothes. A database shown in Table 1 contains

general elements about skirts. Moreover, an

interactive genetic algorithm (IGA) is used to

accelerate the sketch recognition process. The

outline of this sketch design recognition system is as

follows:

1) The system generates the skirt candidates by

combining style elements in the database in

accordance with the composite skirt model and

decoding of the skirt style information, and then

the skirt candidates are generated and displayed

on screen for users by using the sketch

recognition method.

2) A user subjectively evaluates the skirt

candidates and chooses the more favourite ones.

3) The system modifies skirts according to users’

evaluations by using genetic operations of

crossover and mutation, and the modified skirt

candidates are displayed on screen for users

again.

4) The system iteratively implements procedures

(2) and (3) to produce skirt candidates that can

satisfy a user. If a user is satisfied with some

designed skirts, the system can terminate the

design process.

Figure 1: System architecture.

Table 1: Skirts classification.

3 SKETCH RECOGNITION

APPROACH

3.1 Composite Description Model

The composite skirt model based on the knowledge

of fashion design is composed of three portions: the

general factors, style features and detail factors. The

general factors provide information about the outline

silhouette, the ratio of the length and width, the

waist level, the symmetry of a skirt, waist band and

hem. The style features include dart, pleat, panel,

and york. The detail factors contain pocket, gather,

and slit. In general, dart feature can be classified into

three classes: straight dart, curve dart and tuck dart.

ICINCO 2011 - 8th International Conference on Informatics in Control, Automation and Robotics

The position of a dart feature on a skirt can be

represented by using the coordinates (d

, d

) and

, d

). There are three panel features: vertical

panel, horizontal panel and sidelong panel. All of

them can be further classified into three sub-classes:

straight, curve and angle. The position of a panel

feature on a skirt can be decided by the coordinates

(pa

, pa

) and (pa

, pa

). For the york feature,

there are three classes: straight, curve and angle,

moreover, the york feature can be placed on a skirt

by using the points y

and y

. For the pleat feature,

there are five classes, and the pleat feature can be

placed on a skirt by using the point p

3.2 Elements Database

A database of general elements about skirts is shown

in Table 1. First, at the first level, the outline

silhouette of skirts is categorized into five classes.

Next, at the second level, waist band, hem, four style

features and three detail factors are further classified

into classes, respectively.

3.3 Sketch Recognition

The procedure of the sketch recognition method is

considered as a two-stage process. In the first stage,

general factors are generated. Then in the second

stage, one style element is selected.

Stage 1: Generate General Factors

1) Assume that the origin of the created skirt style

is represented by (ox, oy).

2) Obtain the skirt style information about outline

silhouette of the created skirt style in

accordance with decoding of the skirt style

information,

3) Calculate the enclosing rectangle (i.e., m) of

outline silhouette of the created skirt style,

4) The origin (ox, oy) of the created skirt style

becomes:

ox = m_left+(m_right-m_left)*0.5

oy = m_top

where m_left, m_right, and m_top are the left

coordinate, the right coordinate and the top

coordinate, for the enclosing rectangle of the

created skirt style, respectively.

5) Obtain the length ratio of the skirt style in

accordance with decoding of the skirt style

information,

6) Decide the waist level of the skirt style in

accordance with decoding of the skirt style

information and designers’ experiences , then

oy = oy-Y_waist if above waistline

oy = oy+Y_waist if below waistline

where Y_waist is the distance between the

waistline and the top of the waist band

7) Decide the style type of the waist band feature

in accordance with decoding of the skirt style

information, then place the waist band at the

position (i.e., WB_Y) as below:

WB_Y = oy if no waistband

WB_Y = oy-WB_height if waistband

WB_Y = oy-WB_height if waist-yoke

where WB_height is the height of the waistband

feature.

8) Decide the style type of the hem feature in

accordance with decoding of the skirt style

information, then place the hem at the position

(i.e., H_Y) as below:

H_Y = m_bottom if hemline

H_Y = m_bottom+ hem_width if hemband

H_Y = m_bottom if hemflare

where hem_width is the height of the hem

feature, and m_bottom is the bottom coordinate

for the enclosing rectangle of the created skirt

style.

Stage 2: Select Style Element

1) According to decoding of the skirt style

information, choose only one style element

among style features and detail factors.

2) According to decoding of the skirt style

information, if the dart feature is included in the

created skirt style, then place the dart at the

position (i.e., LDart_X, LDart_Y, RDart_X,

RDart_Y) as follows:

LDart_X = ox-0.5*(d

-d

)

LDart_Y = oy

RDart_X = ox+0.5*(d

-d

)

RDart_Y = oy

3) According to decoding of the skirt style

information, if the pleat feature is included in

the created skirt style, then place the pleat at the

position (i.e., Pleat_Y) as Pleat_Y = oy

4) According to decoding of the skirt style

information, if the panel feature is included in

the created skirt style, then place the panel at the

position (i.e., LPanel_X, LPanel_Y, RPanel_X,

RPanel_Y) as follows:

(a) For vertical panel

LPanel_X =ox-0.5*(pa

-pa

)

LPanel_Y =oy

RPanel_X=ox+0.5*(pa

-pa

)

RPanel_Y =oy

(b) For horizontal panel and sidelong panel

INTERACTIVE SKETCH DESIGN RECOGNITION SYSTEM USING EVOLUTIONARY TECHNIQUES

LPanel_X=pa

LPanel_Y=pa

RPanel_X=pa

LPanel_Y=pa

5) According to decoding of the skirt style

information, if the york feature is included in

the created skirt style, then place the yoke at the

position (i.e., Yoke_Y) as Yoke_Y=oy.

6) According to decoding of the skirt style

information, if the gather feature is included in

the created skirt style, then place the gather at

the position (i.e., Gather_Y) as Gather_Y=oy.

7) According to decoding of the skirt style

information, if the pocket feature is included in

the created skirt style, then the pocket position

(i.e., Pocket_X, Pocket_Y) is decided by the

user.

8) According to decoding of the skirt style

information, if the slit feature is included in the

created skirt style, then the slit position (i.e.,

Slit_X, Slit_Y) is decided by the user.

4 INTERACTIVE GENETIC

ALGORITHM

Assuming that the current generation is t and the

current population is represented by X(t). Here, a

small population of 9 individuals in order to ensure a

good compromise between convergence speed and

usability.

Step 1: Set t = 0, generate an initial population of 9

individuals randomly, then present these

individuals to the user.

Step 2: If the user is satisfied with the individuals,

then terminate the search process.

Step 3: Select individuals by the user, and then

perform crossover and mutation

Firstly, if one individual is selected, then

the selected individual will be cloned eight

times. These eight individuals will be

placed into a mating pool where the

mutation operation is performed. Secondly,

if two individuals are selected, then one

individual is randomly selected within the

two selected individuals to generate three

selected individuals, then these three

individuals will be placed into a mating

pool where the genetic operations of

mutation and crossover are performed.

Thirdly, if three individuals from X(t)

which are selected by the user, then these

three selected individuals will be performed

crossover and mutation operations in the

mating pool.

Step 4: Create a new population for the next

generation

All the newly generated individuals are

then collected to form the new population

known as X(t+1), which will replace X(t)

and serve as the population of individuals

for the next generation t+1. Unlike the

elitist strategy (De Jong, 1975) of genetic

algorithms, in which the single best

individual with the highest fitness value in

parent population is reserved and is copied

directly into the new population; in this

study, all the individuals in each generation

are reserved.

Step 5: Check the pre-specified stopping condition.

In this case, the pre-specified stopping

condition is satisfied when the user is

satisfied with the created individuals. If it is

satisfied, terminate the search process, and

return to the best solution as the final

solution. Otherwise, increase t by 1 and go

to Step 3.

4.1 Structure of the Individuals

Although there are many different representations to

implement interactive genetic algorithm, the most

natural representation for the skirts design problem

is the value encoding representation. (Chen and Hou,

2006) In this study, the information contained in an

individual is used to construct a feasible solution

which corresponds to a unique skirt. In this research,

each chromosome as shown in Figure 2 consists of

three portions: a set of bits in the first portion of the

string that is a set of real numbers to indicate the

general factors of skirts. A set of bits in the second

portion of the string comprises four substrings: dart

substring, pleat substring, panel substring, and yoke

substring, which is a set of real numbers to represent

the style features such as dart, pleat, panel, and yoke.

And a set of bits in the third portion of the string that

is a set of integer numbers contains the information

about pocket, gather, and slit.

Figure 2: Chromosome encoding.

4.2 Crossover Operation

In this study, for the first portion of the chromosome,

ICINCO 2011 - 8th International Conference on Informatics in Control, Automation and Robotics

no crossover operator is employed before the “S” bit

(i.e., symmetry of a skirt). For the bits of the

chromosome after the “S” bit, one-point crossover

operator is applied to recombine the individuals. The

procedure of the crossover operator is presented

below:

Step 1. Select two parents randomly from the

mating pool.

Step 2. Select a crossing-site along the parent

individuals. This crossing-site is used as a

cutting point to swap the bits among the

strings. There are no crossing-sites within

each substring.

Step 3. Generate the offspring individuals by

cutting the parent individuals at the

crossing-site and swapping the bits after

the cut.

4.3 Mutation Operation

When the crossover process is completed, the

mutation operator will be used to guarantee

population diversity. The mutation rate should not

be very high; otherwise, the individual will be

disrupted, and the genetic search bears no difference

from a random search. In this research, for the bits

before the “S” bit (i.e., symmetry of a skirt) in the

first portion of the chromosome, the mutation

operation is not performed; for the remaining

portions of the chromosome, three kinds of mutation

operators are employed.

Firstly, for the four bits such as d

, p

, pa

, y

in the second portion of the chromosome, traditional

gene-alter mutation operator (Goldberg, 1989) is

used. For instance, if an offspring individual is

encoded by using the value encoding representation,

(0 1 1 0), then four random numbers ranging from

0.00 to 1.00 are drawn: (0.653, 0.231, 0.007, 0.014).

If the mutation rate is 0.01, one random number in

the above array has its value smaller than the

mutation rate. This number will trigger the mutation

operation to take place in the third bit of the string.

The mutation operator will cause the bits to change

from 1 to 0 or from 0 to 1 whenever the mutation

operations are triggered. The resulting individual

will become (0 1 0 0).

Secondly, for the four bits (i.e., d

, p

, pa

, y

)

in the second portion of the chromosome to

represent the subclass marks of the features for the

created style, the following mutation operation is

adopted. If the feature mark (i.e., d

, p

, pa

, y

) of

the created style is equal to one and the mutation

operation is implemented, then generate a random

integer ω within a range of [1, l] (l is dependent on

the created style features) to determine the subclass

mark of the feature for the created skirt style;

otherwise, ω=0

Thirdly, for the five bits such as WB and H in

the first portion of the chromosome, P, G, SL in the

third portion of the chromosome, the following

mutation operation is adopted. If the mutation

operation is implemented, then generate a random

integer θ within a range of [0, w] (w is dependent on

the created style features) to determine the feature of

the created skirt style; otherwise, θ=0

Finally, for the remaining bits in the second

portion of the chromosome, let T denote a random

integer number within a range of [-1, 1], and

denote a uniform random number in the range [0, 1].

denotes the value to be mutated, and the notation

’

denotes the value after mutation, which is given

as follows:

’

= D

+(D

kmax

- D

)(1-τ

) if T=1,

’

= D

-(D

- D

kmin

)(1-τ

) if T=-1,

where D

kmax

and D

kmin

are the maximum and

minimum values of D

, respectively, and c is a

constant.

5 SUBJECTIVE EXPERIMENTAL

RESULTS

In this section, subjective experiments are performed

to verify whether the proposed methodology can

help laypersons design clothes reflecting laypersons’

preference or not. In this study, ten subjects are

requested to find good-looking design by using this

system, In all the experiments, the genetic

parameters adopted for the interactive genetic

algorithm after testing are Population size = 9,

Crossover rate = 0.7, Mutation rate = 0.01,

Maximum number of generations = 10. At each run,

subjects are asked to design skirts and evaluate the

skirts with 5-point scale (i.e., -2 to +2). Figure 3

presents some examples of generated skirts after ten

generations. It reveals that various skirts are

obtained after ten generations even if the same initial

individuals at the beginning. In addition, in Figure 4,

the average of satisfaction degrees among subjects is

plotted against the generation number. It indicates

that the satisfaction degree becomes high as

generation progress. As the previous presentation, it

can be noted that the proposed methodology is

useful for non-professional users without knowledge

on clothes design to design and obtain skirts

reflecting their preference.

INTERACTIVE SKETCH DESIGN RECOGNITION SYSTEM USING EVOLUTIONARY TECHNIQUES

Figure 3: Some examples of skirts.

Figure 4: Average of satisfaction degrees.

6 CONCLUSIONS

In this study, a sketch recognition approach and an

interactive evolutionary strategy has been proposed

to establish an effective methodology for laypersons

to design clothes reflecting their preference. This

system possesses several advantages over the

existing fashion design systems. Firstly, a composite

description model based on the knowledge of

fashion design is developed to create various skirts

and to overcome the impractical designs. Secondly,

with interactive genetic algorithm, it can reflect

personal preference in fashion design directly rather

than setting a standard of “goodness of design”.

Finally, a two-stage sketch recognition method is

developed to help laypersons achieve satisfied

clothes without compromising the computational

effort and expense. The subjective experiments

results have shown that the proposed methodology

has provided an effective means to help laypersons

get satisfied clothes according to the laypersons’

preference.

ACKNOWLEDGEMENTS

The work described in this paper was partially

supported by a grant from the Research Grants

Council of the Hong Kong Special Administrative

Region, China (Project No. PolyU5254/08E). The

project was also financially supported by another

project (4-ZZ60) of The HK Polytechnic University.

REFERENCES

Cho, S. B., 2002. Towards creative evolutionary systems

with interactive genetic algorithm, Applied

Intelligence, 16, 129– 138.

Chen H., Xu Z. J., Liu Z. Q., Zhou S. C., 2006. Composite

Templates for Cloth Modeling and Sketching,

Proceedings of IEEE Conference on Computer Vision

and pattern Recognition, 1, 943-950.

Chen J. S. Hou J. L., 2006. A Combination Genetic

Algorithm with Applications on Portfolio

Optimization, Advances in Applied Artificial

Intelligence, Springer-Verlag Berlin Heidelberg,

197–206,

De Jong K. A, 1975. An analysis of the behavior of a class

of genetic adaptive systems, PhD dissertation, Univ.

Michigan, Ann Arbor, MI.

Goldberg, D. E., 1989. Genetic algorithms in search,

optimization and machine learning, Reading MA:

Addison Wesley.

Holland, J. H., 1975. Adaptation in natural and artificial

systems, Ann Arbor: University of Michigan Press.

Kim H. S., Cho S. B., 2000. Application of interactive

genetic algorithm to fashion design, Engineering

Applications of Artificial Intelligence 13(6), 635-644.

Sugahara M., Miki M., Hiroyasu T., 2008. Design of

Japanese Kimono using Interactive Genetic

Algorithm, 2008 IEEE International Conference on

Systems, Man, and Cybernetics, 174 – 178.

Takagi H., 2001. Interactive evolutionary computation:

Fusion of the capabilities of EC optimization and

human evaluation. Proceedings of the IEEE, 89(9),

1275-1296.

Nakanishi Y., 1996. Capturing preference into a function

using interactions with a manual evolutionary design

aid system, Genetic Programming, Late-Breaking

Papers, 133–138.

Ogata Y., Onisawa T., 2008. Interactive clothes design

support system, Neural Information Processing: 14th

International Conference, ICONIP 2007, Part II, 657-

665.

‐3

‐2

‐1

012345678910

ICINCO 2011 - 8th International Conference on Informatics in Control, Automation and Robotics