Towards Creation of a Curl Pattern Recognition System

Abstract

This work describes different hair typing systems and proposes a recognition system that can be used to classify human hair curl pattern. The Andre Walker hair typing system is used as a basis for defining curl type. Curl type is a term used to describe how curly, wavy, or kinky a person’s hair is according to 4 categories. This work limits recognition of curl types to categories 4a, 4b, 4c, and 3c. The hair analysis algorithm described in this work was initially developed as a pilot project through studying approximately 30 subjects that provided pictures of their hair. In this work, we describe a classifier that labels hair images and will eventually lead to a system for recognizing an individual’s curl pattern. Results from this work may lead to technologies to suggest hair care products for persons with textured hair.

Full text

Towards Creation of a Curl Pattern Recognition
System
Hill Kymberlee
Howard Univers ity
Wash ington, DC, USA
kymberlee.hill@bison.howard.edu
Ab im b o la Ade s o la
Howard Univeris ity
Wash ington, DC, USA
adesola.abimbola@bison.howard.edu
Abstract—This work describes different hair typing systems
and proposes a recognition system that can be used to classify
human hair curl pattern. The Andre Walker hair typing system
is used as a basis for defining curl type. Curl type is a term used
to describe how curly, wavy, or kinky a person’s hair is
according to 4 categories. This work limits recognition of curl
types to categories 4a, 4b, 4c, and 3c. The hair analysis algorithm
described in this work was initially developed as a pilot project
through studying approximately 30 subjects that provided
pictures of their hair. In this work, we describe a classifier that
labels hair images and will eventually lead to a system for
recognizing an individual’s curl pattern. Results from this work
may lead to technologies to suggest hair care products for
persons with textured hair.
Keywords—hair analysis, convolutional neural networks,
image recognition, curl pattern recognition
I. INTRODUCTION
Th e a ve rage num be r o f ha ir s tra nds on a pe rso n ’s h e a d i s
close to 100,000 [1]. Thus , ha ir seems dense ly packed on the
s ca lp. Ha ir cla ss ifica tion is a com ple x pa tte rn re cogn itio n
pr ob le m be ca us e tra d it ion a l co m pu te r vis io n te ch n iqu es like
object segmentation or image recognition are hard to apply
directly.
The three -dimens ional shape of hair fiber va ries
considerably depending on a person’s e t h n i c it y a n d
geography, with exam ples from very stra ight hair with no
curve, to the tightly sprung coils of the black race [5]. Few
researchers in computer vis ion and pattern recognition have
studied use of human hair shape characteristics [2, 3]. Some
researchers have focused on ha irline and color with a verage
resu lts. Ide ntify ing curl pa ttern ca n give m ore info rma tion
about a person ’s identity, hair texture, and hair porosity.
Many women and men with textured ha ir struggle with
finding hair products that cater to their hair ca re needs.
Because of this, this research was started to study how to
create computer vis ion algorithms that can start to recognize
individual curl patterns common in human hair.
This paper will des cribe the three major ha ir
classification systems, why they e xis t, and how hair features
like curls and hair s trand type can be used as a bas is to classify
hair. Finally, the paper d is cusses an a pproach to class ify
human hair into types suggested by the Andre Walker typing
system. This methodology will be used to create a hair curl
recogn ition system . The propos ed sys tem described in th is
paper ma y be used to produce software that re commends
products to individuals with textured hair.
II. HAIR TYPING SYSTEMS
A. Andre Walker Hair Typing
Th e An d re Wa l ke r h a ir t yp in g s ys te m is the m os t w ide s p re ad
system that has been a ccepted by cons ume rs a nd
cosmetologists. This system divides hair into 4 types –Type
1, Type 2, Type 3 and Type 4.
Type 1
Type 1 hair is straight; however, Andre categorizes this
hair type into three very s pecific segments - 1A, 1B, and 1C.
Type 1A hair is described as fine, very thin and soft with a
noticeable shine. Type 1B hair is medium-textured and has
more body than Type 1A hair. Type 1C hair is the most
res ista nt to curly s tyling and re la tive ly coa rse com pa red to
o the r Ty pe 1 ha ir ty pes . Fig ur e 1 s ho ws th e fol licle sh a pe for
type 1 hair.
Type 2
Type 2 is wavy hair that usually isn’t ove rly oily or very
dry. Andre Walker specifie d that Type 2 hair fa lls in between
Type 1 and Type 3. Type 2A hair is fine and thin. It is
re la tive ly e as y to ha nd le from a styl ing pe rspective be ca use it
can eas ily b e s tra igh tened o r cu rled, a ccord ing to the Curl
Centric website. Type 2B hair characteristically has waves
tha t tend to adhe re to the shape of you r hea d. Type 2C ha ir
will frizz easily and it is fairly coarse. Curly hair textures have
a defin ite “S” sha ped curl pattern. Sin ce the cuticle does n’t lay
Figure 1
Da ng a l P ra jjw ha l
Howard Univers ity
Washington DC USA
pra jjwa l.dan gal@b is on .howa rd.edu
Was hing ton Glo ria
Howard Univers ity
Washington DC USA
gloria.washington@howard.edu
III Burge Legand
Howard Univers ity
Washington DC USA
lburge@howard.edu
30 Int'l Conf. IP, Comp. Vision, and Pattern Reco
g
nition
|
IPCV'18
|
ISBN: 1-60132-485-5, CSREA Press ©

flat, curly ha ir isn’t nearly as shiny as Type 1 (straight ha ir) or
Type 2 (w a vy h a ir) ha ir ty pe s . F igu re 2 s h ows th e fo lli cle
shape for type 2 hair.
Type 3
Type 3A hair is very s hiny and loose. Type 3B ha ir has a
medium amount curls , ranging from bouncy ringlets (spiral
like curls of hair) to tight corkscrews (spiral-shaped corkscrew
curls). Type 3C hair isn’t a part of the Andre Walker Hair
Typing System. Figure 3 shows the follicle shape for type 3
ha ir.
Type 4
Type 4 is “kinky” or more appropriately full of tight coils
(tightly curled ha ir). Typically, Type 4 ha ir is a lso e xtremely
wiry and fragile. Often times, it appears to be coarse, however,
it is rea lly ve ry fine , with seve ra l thin ha ir stra nds dense ly
packed together. Type 4A hair is full of tight coils. It has an
“S” pattern when stretched, much like Type 3 curly hair. Type
4B hair has a less defined pattern of curls and looks more like
a “Z” as the ha ir bends with very sha rp angles . Type 4C ha ir
is n ’t a pa r t o f the Andre Walker Hair Typing System. I think
it’s fa ir to sa y that Andre Walke r’s ha ir typing s ys tem was
meant to s im plify a very complex process. his de liberate lack
of detail in the descriptions of hair types allows them to be
mo re in clus ive than e xclus ive . Figure 4 s hows the follicle
shape for type 4 hair.
B. LOIS Hair Typing
The LOIS Hair Typing system is less widely known than the
Andre Walker Hair Typing System. Created by the former
we bs ite Ou rHa ir.ne t , The LO IS s ys tem is a ime d at de f ining
ha ir a wa y from a numbe r and m ore into a des cription o f the
strand s ha pe, s ize and sheen . It de fines ha ir us ing three
characteristics: pattern, strand size, and texture. The hair
pattern is defined by the letters LOIS. If hair is dominated by
rightangles and substantially bends with nearly no curve , then
it is considered a pattern “L”. If hair strand curls or coils
s ign ificantly and a ppea rs to be sha ped like the letter “O”, the n
it is cons idered a pattern “O”. If ha ir ha s no d is tinct ive curls
or bends and prima rily lies flat a ga ins t your hea d, the n it is
cons idered a pattern “I”. If ha ir strand has “S” sha ped curls or
wa ve s wit h d e fine d h il ls a n d va l le ys , th en it is co ns ide re d a
pattern “S”. (Here et al., 2018) A head of hair may have
se veral different LOIS hair patterns. When this occurs, the
system allows you to combine LOIS letters to determine the
ha ir pa tte rn . Fo r e xam ple , th e h a ir pa tte rn ma y be an “O S”
LOIS hair type. In this example, the hair would contain
p rim a r il y a c om b ina t io n o f “ O” s ha pe d s p ira l c ur ls a n d “ S”
shaped curly waves with defined hills and valleys. Identifying
ha ir strands within the LOIS system sta rts with us ing a strand
of frayed thread. One piece of a frayed thread (like a piece of
se wing threa d) is use d as a pro xy to determ ine the s ize of ha ir
strand. The general thought is that a piece of thread is
approximately the size of a medium sized strand of hair.
Examining a representative strand of hair, will show if it is
thin, a verage or thick hair. It is common to confuse thickness
of ha ir with dens ity. Dens ity is how close ly pla ced toge the r
h a i r s t ra n ds a re a s th e y g r o w o u t o f t h e s ca l p . I f y o u r h a i r l o oks
thinne r than the pie ce of thread, you ha ve thin hair. If it’s
thicke r, ob vious ly you ha ve thick ha ir.
C. FIA Hair Typing
The FIA Ha ir Typing System appears to expand on Andre
Walker’s Hair Typing System and also incorporates
co m pon e nts o f t he L OIS Ha ir T yp ing Sys te m . Th e s ys tem
de fines ha ir us ing three class ifiers: the de finition of your curls
(think Andre Walker), the a ppea rance of most o f your hair
strands , a nd the overall volume of your ha ir. The firs t class ifie r
de fines ha ir type by determ ining the curliness of the ha ir
strands is analogous to Andre Walker sys tem utilizing Type 1
–straight hair, Type 2 –wa vy ha ir, Type 3 –curly ha ir a nd
Type 4–rea lly curly hair. The second classifie r focuses on the
appea ran ce of the ma jor ity of the ha ir s trands . Actua lly, th is
class ifier is very simila r to the ha ir strand sizes use d within the
LO IS s ys te m , b ut the re a re s ome m in or d i ffe re n ces in
term inology. Fine hair is defined as thin ha ir strands tha t feels
Figure 2
Figure 3
Figure 4
Int'l Conf. IP, Comp. Vision, and Pattern Reco
g
nition
|
IPCV'18
|
31
ISBN: 1-60132-485-5, CSREA Press ©

almost like an ultra-fine strand of silk. Medium sized hair
strands, which generally feel like rolling a cotton thread
between your thumb and index fingers, are ones that simply
fall in between the fine and coarse categories. Coarse ha ir are
thick ha ir s trands tha t fee l ha rd and wiry. The th ird
classification of FIA’s Hair Typing System measures the
overa ll volume of the ha ir. Simply pla ce the ha ir in a ponyta il.
As much ha ir as poss ible mus t fit in the ponytail. There is no
need to worry too mu ch about how the ha ir loo ks , because the
goal is to have the majority of your hair included in the
pon yta il. On ce you ha ve the ma jority o f your ha ir in a
pon ytail, the ne xt step is to me asure the circum fe rence of the
ponyta il. A soft tape measure or another method that won’t
dama ge the ha ir to meas ure the circum fe ren ce of the ponyta il.
Simply wrap the soft tape measure around the ponyta il below
the elastic used to hold the ponyta il together. The number
shown on the soft tape measure is used to de termine whe ther
the ha ir fits into the thin, normal or thick according to FIA’s
Hair Typing System. If the circumference of the ponytail is
less than 2 inches (less than 5 centimeters) then it will be
considered thin. If the circumference of the ponytail is
between 2 –4 inches (between 5-10 centimeters) then it will
be considered normal. If the circumference of the ponytail is
greater than 4 in ches (greater than 10 ce ntimeters) then it is
considered thick. The FIA system takes a few more hair
characteristics into a ccount for a be tter o ve ra ll ha ir type
diagnosis.
III. METHODOLOGY
A. Curl Identification Pre Processing
Before beginning the project, a web scraper was built to
ha rves t image s from the web that fit into the 4c, 4b, 4a , 3c, 3b,
and 3a categories. Because each tool needs to train on many
images, upwards of 500 images were collected for each
ca te g or y .
An image file is first converted to greyscale and histogram
analysis is be performed for normalization of the image. The
grayscale images are converted into binary image. Using
active shape models we perform feature extraction and use this
as input into deep learning software to determine curl type.
The binary images are then resized into 30 x 30 pixels as the
starting dimens ion s pecification for a convolutional neural
network. Figure 5 and Figure 6 show binary images and
co n tou rs on b ina ry ima ge s res pe ct ive l y. Th e ima ge s be low a re
for representation purposes only, as the proposed s ystem will
re qu ir e the ima ge s be pre -se gm en te d e lim ina t ing a ny n on -h a ir
part of the image. This will ensure our algorithms only trains
on re le vant ha ir s tructures .
B. Curl Classification
Our te chnique can be described in two steps:
1. Curl Iden tification us ing Active Shape Models a nd
dee p lea rn ing so ftwa re
2. Re commendation of products based on the previous
two steps.
OpenCV Haar Cascade tool and Tensor Flow deep
learning software is used to detect, segment, and class ify the
four categories of hair types explored in this work: 4a, 4b, 4c,
and 3c. These too ls were chos en be ca use they ha ve been used
by other researchers in computer vis ion to build object
de te c to rs re la t ive l y s im pl y a nd the re is a m ple do cu me nta t io n
on how to use them for student research projects . Example
images for each curl type is provided to both the OpenCV
Haar Cascade tool and the TensorFlow Deep learning
software
In order to build a s ystem robust enough to recognize
diffe rent curl types from one individual’s head, it is ne cessary
to bu ild a class ifie r for us ing the Andre Wa lker curl typing
system as the basis for identifying images metatagged with
either 4a, 4b, 4c, or 3c. A web scrapper tool was used to pull
representative images from the Internet for building the
clas s ifier. Ne xt, the image dataset res ulting from the web
scrappe r was ana lyzed to pe rform qua lity assu ran ce of the
images contained in the repository. Note, the webs crapper
tool was only used to find an ample training set for building
the class ifier.
In a ddition to da ta collected from the web, we ga thered
participantdata from 30 subjects. Figure 7shows a breakdown
of the ha ir curl types gathered from the volunteers.
Figure 5
Figure 6
32 Int'l Conf. IP, Comp. Vision, and Pattern Reco
g
nition
|
IPCV'18
|
ISBN: 1-60132-485-5, CSREA Press ©

DISCUS SIO N
This a pproach for deve loping a class ifie r has only been
trained for textured hair (3a and above using the Andre Walker
cla ss ifica tio n s ys tem ). Ho weve r, it co uld be re tra ined a nd
tes ted for non -te xtured or s tra igh t ha ir in the future . This work
wa s d e ve lo pe d fo r te xt ure d ha ir be caus e we be lie ve it is a
more d ifficu lt com puter vis io n prob lem to tackle and results
from this wo rk could pro ve m ore use ful to a wider se gment
of society.
ACKNOWLEDGMENT
This work is supported by a grant from the De partment of
Homeland Security- GENIUS Scho larship Program.
REFERENCES
[1] Gray, J. 2003. The world of hair: Hair facts. P&G Hair Care Research
[2] Center. Retrieved March 2, 2007.
[3] Medland, Sarah E., Gu Zhu, and Nicholas G. Martin. Estimating the
heritability of hair curliness in twins of European ancestry; Twin
Research and Human Genetics 12.5 (2009): 514-518.
[4] Rousset, Cedric, and Pierre-Yves Coulon. ;Frequential and color
analysis for hair mask segmentation. Image Processing, 2008. ICIP
2008. 15th IEEE International Conference on. IEEE,
[5] Yacoob, Yaser, and Larry S. Davis. Detection and analysis of hair.
IEEE transactions on pattern analysis and machine intelligence 28.7
(2006): 1164-1169.
[6] Here, K., Growth, H., Products, H., Anderson, A., Beauty, M.,
Morrison, A., Byrd, J., madison, g. and Byrd, J. (2018). The Only Hair
Typing System Article You'll Ever Need. [online] Curl Centric® |
Rewrite the Rules of Natural Hair Care. Available at:
https://www.curlcentric.com/hair-typing-system/ [Accessed 30 May
2018].
[7] J. Trebilcock, "FIA Hair Typing For Identifying Your Tresses - Black
Hair Spot", Black Hair Spot, 2018. [Online]. Available:
[8] https://blackhairspot.com/blog/hair-talk/fia-hair-typing/. [Accessed:
30- May- 2018]
[9] J. Trebilcock, "L.O.I.S. Hair Typing: Beautiful Black Hair Simplified
- Black Hair Spot", Black Hair Spot, 2018. [Online]. Available:
https://blackhairspot.com/blog/hair-talk/lois-hair-typing/. [Accessed:
30- May- 2018]
[10] "What You Need to Know About The L.O.I.S. Hair Typing System |
NaturallyCurly.com", Naturallycurly.com, 2018. [Online]. Available:
https://www.naturallycurly.com/curlreading/learn/what-you-need-to-
know-about-the-l-o-i-s-hair-typing-system-si. [Accessed: 29- May-
2018]
[11] J. Trebilcock, "Andre Walker Hair Types: Learn about this hair typing
system", Black Hair Spot, 2018. [Online]. Available:
https://blackhairspot.com/blog/hair-talk/andre-walker-hair-types/.
[Accessed: 01- Jun- 2018]
[12] "Reader's Mail: LOIS hair typing system!",
Thenaturalhavenbloom.com, 2018. [Online]. Available:
http://www.thenaturalhavenbloom.com/2009/11/readers-mail-lois-
hair-typing-system.html. [Accessed: 01- Jun- 2018]
[13] Aguh, C. and Okoye, G. (2017). Fundamentals of ethnic hair.
[14] https://www.andrewalkerhair.com/. (2018). [image].
Figure 7
Int'l Conf. IP, Comp. Vision, and Pattern Reco
g
nition
|
IPCV'18
|
33
ISBN: 1-60132-485-5, CSREA Press ©