Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
6, N
o
. 5
,
O
c
tob
e
r
201
6, p
p
. 2
140
~214
9
I
S
SN
: 208
8-8
7
0
8
,
D
O
I
:
10.115
91
/ij
ece.v6
i
5.1
078
5
2
140
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
An Unsu
pervised
Classifi
cation
Technique for Detection of
Flipped Orientations
in Document Images
Vijayashree
C. S
1
, N.
Sh
ob
h
a
Ra
ni
2
, V
a
s
u
dev T
3
1
PES Research
Center
, PES Institute of
Technolog
y
,
Mand
y
a
, Karnataka, India
2
Department of Computer
Scien
ce,
Amrita Vishwa Vidy
ape
e
th
a
m
, M
y
sore
, Am
rita
Universit
y
,
In
dia
3
Maharaj
a
R
e
search Foundat
i
on,
Universit
y
of
My
sore
, Mah
a
raj
a
Institute of
Tech
nolog
y
,
M
y
sore, India
Article Info
A
B
STRAC
T
Article histo
r
y:
Received Apr 10, 2016
Rev
i
sed
Ju
l 4
,
2
016
Accepte
d
J
u
l 20, 2016
Detec
tion of
tex
t
orien
t
at
ion in
docum
ent im
age
s
is of prelim
ina
r
y
conc
er
n
prior to processing of documents b
y
Optica
l
Charac
ter Re
ader
.
The text
direction in do
cument images should exis
t g
e
ner
a
lly
in a specific orientation
,
i.e
.,
tex
t
dir
ect
i
on for an
y au
to
m
a
ted document reading s
y
stem. The flipped
text orientation
leads to an u
n
ambi
guous result in such fully
au
tomated
s
y
stems. In th
is
paper, we fo
cus on de
velopmen
t
of text or
ien
t
a
tio
n direc
tion
detection module which can be incorpor
ated as the perquisite process in
autom
a
ti
c read
i
ng s
y
s
t
em
. Or
ient
ati
on direction detect
ion
of text is
performed thro
ugh emplo
y
ing
direction
a
l gr
adient
features o
f
document
im
age and
adapt
s
an uns
upervis
e
d
learn
i
ng appro
ach for d
e
te
ct
ion
of flipp
e
d
text or
ientation
at which
the document ha
s been
originally
fed
in
to scanning
devic
e
. Th
e unsupervised le
arni
ng is built
on the
direct
ional gr
ad
ient fe
atur
es
of text of
document based o
n
four
possible differen
t
orien
t
ations
. The
algorithm
is exp
e
rimented
on do
cument sa
mples
of printed plain
English tex
t
as well
as filled
in pre-pr
inted
fo
rms
of Telugu s
c
ript. Th
e outco
me attained
b
y
algorithm proves to be
consistent a
nd
adequ
a
te with
an av
era
g
e ac
cur
a
c
y
around 94%
.
Keyword:
Direction
a
l Grad
ien
t
D
o
cu
me
n
t
I
m
a
g
e
Fl
i
ppe
d Te
xt
O
r
i
e
nt
at
i
o
n
Uns
u
per
v
i
s
e
d
Learni
ng
Copyright ©
201
6 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
N
.
Shob
h
a
Rani,
Depa
rtem
ent of Com
puter Sci
e
nce,
Am
rita Vishwa
Vidy
a
p
eetham
,
M
y
sore
,
Am
rita Un
iv
ersity, Ind
i
a.
Em
a
il: n
.
sho
b
a
1
985
@g
m
a
il.c
o
m
1.
INTRODUCTION
Has c
o
m
e
out
cl
earl
y
from
t
h
e st
udy
on a
u
t
o
m
a
ti
c r
ead
ing th
ro
ugh
Optical Charact
er R
ecognition
(OCR) systems th
at th
e OCRs h
a
v
e
li
m
ita
tio
n
s
in
read
i
n
g an
d reco
g
n
i
z
i
ng t
e
xt
d
o
cu
m
e
nt
s whi
c
h a
r
e not
i
n
pr
o
p
er t
e
xt
ual
ori
e
nt
at
i
on i
.
e
.
ot
he
r t
h
a
n
0
0
t
e
xt
o
r
i
e
nt
at
i
o
n.
It
i
s
qui
t
e
pos
si
bl
e t
o
ha
ve
doc
um
ent
i
m
ages
wi
t
h
t
e
xt
ori
e
nt
ed at
di
f
f
e
r
i
n
g
ori
e
nt
at
i
ons
d
u
e
t
o
t
h
e i
m
pro
p
e
r di
r
ect
i
on i
n
feedi
ng t
h
e
do
cum
e
nt
s t
o
scanni
ng
devi
ce
by
t
h
e
ope
rat
o
rs.
The
d
o
cum
e
nt
im
ages
wi
t
h
t
e
xt
ori
e
nt
ed
i
n
ot
her t
h
a
n
o
r
i
e
nt
at
i
on a
r
e
fl
i
p
p
e
d
d
o
c
u
m
en
ts and are
no
t su
itab
l
e to
ex
tract t
h
e in
fo
rm
a
tio
n
by O
CRs.
Su
ch f
lipp
e
d do
cumen
t
i
m
ag
es need
t
o
be
pr
ocesse
d t
o
fi
x t
h
e i
m
age i
n
t
o
p
r
o
p
er/
n
o
r
m
a
l
ori
e
nt
at
i
on.
No
rm
ally
docu
m
ent im
age suffe
rs fr
om
eith
er ske
w
or im
proper
orie
n
t
atio
n. A clear d
i
stin
ctio
n
lies
b
e
tween
sk
ewed
do
cu
m
e
n
t
s an
d
im
p
r
op
er
o
r
ien
t
ed
d
o
c
um
en
ts. Sk
ew i
s
th
e sm
all an
g
u
l
ar ro
tatio
n
/
t
ilt to
th
e
doc
um
ent
i
n
t
h
e
no
rm
al
di
rect
i
on,
w
h
ereas
im
pro
p
er
o
r
i
e
nt
ed
d
o
cum
e
nt
/
f
l
i
ppe
d d
o
c
u
m
e
nt
cor
r
esp
o
nds
t
o
a
t
o
t
a
l
rot
a
t
i
on t
o
a doc
um
ent
i
n
a di
ffe
rent
di
rect
i
o
n ot
he
r
t
h
an t
h
e n
o
rm
al
di
rect
i
on. T
h
e fl
i
p
ped
doc
um
ent
s
co
nf
licts th
e
basic op
er
ating pr
o
c
edu
r
e fo
r v
a
r
i
ou
s
docum
e
nt readi
n
g
syste
m
s
like character re
cognition
sy
st
em
s, pri
n
t
e
rs,
p
hot
o c
o
py
i
n
g
sy
st
em
s and
ot
he
r i
m
agi
ng sy
st
em
s.
W
i
t
h
re
spect
t
o
OC
R
s
, t
h
e i
m
pr
o
p
e
r
tex
t
o
r
ien
t
ation
in
trod
u
c
es am
b
i
g
u
ity in
read
ing
th
e tex
t
and
resu
lts in
erro
n
e
ou
s reco
gn
itio
n ou
t
c
o
m
es
whe
r
eas i
n
ph
ot
oc
opy
i
n
g
p
r
ocess i
n
t
r
od
uc
es am
bi
gui
t
y
i
n
doc
um
ent
fe
edi
n
g
pr
ocess
t
o
t
h
e
de
vi
ces a
n
d
t
h
us
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
An
Un
sup
e
rvised
Cla
s
sifica
tio
n Tech
n
i
q
u
e
fo
r
Detectio
n o
f
Flip
p
e
d
Orienta
tio
n
i
n
.... (Vija
ya
sh
ree C
.
S
.
)
2
141
l
eads t
o
n
o
n
-
uni
fo
rm
t
e
xt
ori
e
nt
at
i
ons i
n
t
h
e sam
e
phot
o c
opi
e
d
d
o
c
um
ent
.
The
m
a
nual
p
r
oce
d
u
r
e
o
f
cor
r
ect
i
n
g
doc
um
ent
ori
e
nt
at
i
o
n
i
n
t
e
r
v
en
es
and
sl
o
w
s
t
h
e
basi
c
ope
rat
i
o
n
of
t
h
e
d
o
c
u
m
e
nt
readi
n
g s
y
st
em
s
and
ot
her
i
m
agi
n
g
de
vi
ces.
Especi
al
l
y
t
h
e
doc
um
ent
read
ing
system
s like
OCR, t
h
e
m
o
st crucial stages
of
d
o
c
u
m
en
t i
m
a
g
e pro
cessing
lik
e seg
m
en
tati
o
n
, feature ex
tractio
n
and
do
cu
m
e
n
t
classific
a
tio
n
are sen
s
itiv
e to
t
h
e fl
i
p
pe
d o
r
i
e
nt
at
i
on
o
f
t
h
e
d
o
cum
e
nt
im
ages
[1]
.
F
lip
ped
o
r
ien
t
ation
in
a do
cu
m
e
n
t
i
m
ag
e is du
e
to
th
e
erro
r in
p
l
acing
/feed
i
ng
th
e
do
cu
m
e
n
t
in
to th
e scann
i
ng
de
vi
ce [
2
]
,
[
3
]
.
D
u
ri
ng
t
h
e
p
r
oce
ss o
f
sc
an
ni
n
g
,
t
h
er
e
is a p
o
ssib
ility to
feed
th
e
do
cu
m
e
n
t
to
scan
n
e
r b
y
p
l
aci
n
g
the do
cu
m
e
n
t
in
wron
g
d
i
rection
s
leadin
g
to
gene
rat
i
o
n o
f
d
o
cum
e
nt
im
ages i
n
co
rres
p
on
di
n
g
fl
i
p
pe
d o
r
i
e
nt
at
i
ons.
Im
pro
p
er
o
r
i
e
nt
ed
t
e
xt
i
s
t
e
xt
whi
c
h i
s
not
at
0
o
ori
e
n
t
at
i
on t
o
t
h
e pa
ge. Im
pro
p
e
r
o
r
i
e
nt
ed
doc
um
ent
can be
ori
e
nt
ed at
9
0o
or
18
0
o
o
r
2
7
0
o
t
o
t
h
e
p
a
g
e
. Th
e
d
i
ff
er
en
t typ
e
s of
tex
t
or
ien
t
ed
sam
p
les ar
e show
n in
Figu
r
e
1.
Fi
gu
re
1.
Di
f
f
e
r
ent
t
e
xt
o
r
i
e
nt
at
i
ons
The i
m
pro
p
erl
y
ori
e
nt
ed
d
o
c
u
m
e
nt
im
ages need
t
o
be
p
r
e
p
r
o
cesse
d
or
c
o
n
v
e
r
t
e
d t
o
00
o
r
i
e
nt
at
i
o
n
for correct rea
d
ing by
OCR.
The m
e
thods
reporte
d
in literature
with
res
p
ect to
Doc
u
m
e
nt Im
age Ana
l
ysis
(DIA) is v
a
rian
t to
ro
tatio
n
an
d
d
e
m
o
n
s
trate lo
wer effi
ci
ency if the docum
ent imag
e su
ffers
fro
m
ro
tatio
n.
Th
is
d
e
m
a
n
d
s
t
h
e
n
e
ed
for t
h
e in
pu
t
do
cu
m
e
n
t
im
ag
es to
b
e
ro
tatio
n free.
In
ad
d
ition
,
OC
Rs also
d
e
m
o
nstrate
a
m
a
rk
ed
red
u
ctio
n
in
recogn
itio
n
rate if th
e inp
u
t
tex
t
images are orie
nted in a
n
y
di
r
ect
i
on ot
her t
h
an 0
0
.
Hence
the
r
e i
s
a
need for
a preproc
e
ssing sta
g
e
wh
i
c
h
det
ect
s a
n
d
cor
r
ect
s t
h
e
fl
i
ppe
d
ori
e
nt
at
i
o
n
o
f
doc
um
ent im
a
g
e
before
it is s
u
bjected for re
ading
by OCR.
We ha
ve f
o
cu
sed i
n
t
h
i
s
res
earch t
o
det
ect
t
h
e fl
i
pped
ori
e
nt
at
i
on o
f
t
h
e doc
um
ent
im
age an
d
tran
sform
it
to
ach
iev
e
o
r
ien
t
atio
n
at to
t
h
e
d
o
c
u
m
en
t
p
a
g
e
to
prep
are th
e
tran
sform
e
d
tex
t
m
o
re su
itab
l
e for
pr
ocessi
ng
by
OC
R
.
2.
LITERATU
R
E
SU
RVE
Y
Co
n
s
i
d
erab
le
nu
m
b
er of research
wo
rk
s are
n
o
ticed during
literatu
re su
rv
ey o
n
do
cu
m
e
n
t
o
r
ien
t
atio
n
d
e
tectio
n
an
d
co
rrectio
n.
W
i
th
th
e av
ailab
i
lity o
f
ap
riori k
nowledg
e abo
u
t
th
e layou
t o
f
th
e
d
o
cu
m
e
n
t
, th
e
page
o
r
i
e
nt
at
i
o
n
det
ect
i
on
w
o
ul
d
be m
a
de si
m
p
li
fi
ed t
o
ce
r
t
ai
n ext
e
nt
.
B
u
t
t
h
i
s
ap
ri
ori
k
n
o
w
l
e
d
g
e
w
o
u
l
d n
o
t
be avai
l
a
bl
e fo
r gene
ri
c cases
and re
qui
res uns
u
p
er
vi
se
d cl
assi
fi
cat
i
on ap
pr
oac
h
es. Th
e gene
ri
c ap
pr
oa
ches
for
d
e
tectio
n of
p
a
g
e
orien
t
atio
n d
i
rectio
n calls for ex
tr
action of
certain gl
obal
feat
ures
from
the docum
e
nt.
In
th
is d
i
rection
research
ers
h
a
ve
m
a
d
e
atte
m
p
ts in
d
e
t
ect
i
ng
t
h
e pa
ge o
r
i
e
nt
at
i
on di
rect
i
o
n
and are
rep
o
r
t
e
d i
n
literatu
re.
Few
related
works referred are
d
i
scu
ssed
in th
is sectio
n
.
C
a
pra
r
i
[4]
ha
s pr
op
ose
d
a
m
e
t
hod f
o
r pa
ge u
p
/
d
o
w
n
or
i
e
nt
at
i
on det
e
c
t
i
on m
odel
usi
ng B
a
y
e
si
a
n
cl
assi
fi
er. T
h
e
al
go
ri
t
h
m
oper
a
t
e
s on a
bi
t
-
m
a
ppe
d t
e
xt
pat
t
e
rn a
rray
t
o
de
t
e
rm
i
n
e t
h
e u
p
/
d
o
w
n o
r
i
e
nt
at
i
on
o
f
t
h
e pa
ge, i
.
e.
w
h
et
he
r t
h
e
page
i
s
up
ri
g
h
t
or i
nve
rt
ed
by
ex
p
l
oi
t
i
ng an
u
p
/
d
ow
n asy
m
m
e
t
r
y
of pa
ssage
s o
f
t
e
xt
co
m
p
o
s
ed
o
f
ro
m
a
n
letters a
n
d
Arab
ic nu
merals. Th
is ap
pr
o
a
ch
is li
m
i
te
d
to
classify on
ly o
r
ien
t
ations in
00
and 1800. Ne
stares [5] ha
s fil
e
d a
patent
for the detec
t
i
o
n of d
o
m
i
nant
or
i
e
nt
at
i
on
est
i
m
at
i
on by
usi
n
g
seve
n
steerab
le filters. Th
e orien
t
atio
n
d
e
tectio
n
is p
e
rform
e
d
u
s
ing
steerab
le filters wh
ich
provid
e
an
en
erg
y
v
e
rsus
ori
e
nt
at
i
on cu
r
v
e of t
h
e im
age dat
a
. A m
a
x
i
m
u
m
of t
h
e ener
gy
cur
v
e m
a
y
i
ndi
cat
e t
h
e am
ount
o
f
an
gul
a
r
rotation that may be corrected by
t
h
e o
r
i
e
nt
at
i
on co
rrect
or
. A m
e
t
h
o
dol
o
g
y
was p
r
o
p
o
s
e
d by
A
d
i
t
y
a et
.al
[6]
to
d
e
tect
o
r
ien
t
atio
n
in non
-tex
tu
al im
ag
es by ad
op
ting
Bayesian
classifier fo
r estim
a
tin
g
th
e orien
t
atio
n and
m
e
t
hod i
s
n
o
t
e
x
t
e
n
d
abl
e
t
o
t
e
xt
i
m
ages.
D.
X. Le
et
. al
[7]
pr
o
p
o
s
ed
an al
g
o
r
i
t
h
m
to
det
ect
pa
ge
ori
e
nt
at
i
on a
n
d s
k
ew
i
n
d
o
c
u
m
e
nt
usi
n
g
p
r
oj
ectio
n pro
f
ile. Th
e lim
ita
t
i
o
n
of th
e m
e
t
h
odo
log
y
is that it can
d
e
tect th
e
o
r
ien
t
atio
n
o
f
th
e
do
cumen
t
i
n
p
o
rtrait an
d lan
d
s
cap
e d
i
rectio
n bu
t
d
o
not d
i
stin
gu
is
h
b
e
tween
no
rmal
and
flip
p
e
d
with
in
po
rt
rait
or
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 5
,
O
c
tob
e
r
20
16
:
214
0
–
21
49
2
142
l
a
ndsca
pe
di
re
ct
i
on.
A
v
i
l
a
an
d Li
ns
[8]
s
u
g
g
est
e
d
a fa
st
m
e
t
hod t
o
det
e
ct
ske
w
a
n
d
o
r
i
e
nt
at
i
on i
n
c
o
m
p
l
e
x
m
onochrom
atic doc
um
ent im
ages. T
h
e m
e
thod is lim
ited only
to detect skew bet
w
een t
o
to
norm
al position
but
d
o
es
not
a
d
d
r
ess t
h
e
det
ect
i
on
of
fl
i
p
p
e
d
ori
e
nt
at
i
o
ns
. Vas
u
dev
et
al
[9]
p
r
op
ose
d
a s
k
e
w
co
rr
ect
i
o
n
fol
l
o
we
d by
or
i
e
nt
at
i
on
det
ect
i
on a
nd c
o
rrec
t
i
on a
p
p
r
oa
ch
.
A non
-ro
t
ational tran
sform
a
t
i
o
n
m
o
d
e
l is app
lied
i
n
t
w
o
st
ages.
Thi
s
ap
p
r
oac
h
w
o
rks i
n
de
t
ect
i
on o
f
fl
i
p
ped
o
r
i
e
nt
at
i
o
n i
n
m
ono
ch
r
o
m
e
t
e
xt
im
ages wi
t
h
ori
e
nt
at
i
ons
.
M
u
ral
i
et
. al
[
10]
hav
e
p
r
op
ose
d
t
h
e
ske
w
co
rrection
in
th
e first stag
e
wh
ich
is
b
a
sed on
lin
e
t
r
ans
f
o
r
m
a
ti
on
m
odel
.
I
n
t
h
e
seco
nd st
a
g
e a
sim
p
l
e
x-cut
a
nd y
-
cut
t
ech
ni
que t
o
det
e
rm
ine t
h
e
ori
e
nt
at
i
on
o
f
t
h
e doc
um
ent
.
A fi
ner
deci
si
on
on
ori
e
nt
at
i
on i
s
m
a
de based
on t
h
e
dom
ai
n kno
wl
ed
ge o
f
t
h
e
pi
xel
di
st
ri
b
u
t
i
on i
n
t
h
e doc
um
ent
im
age. T Asan
o et
al
[11]
p
r
esen
ted
an
algo
rith
m
fo
r ro
tatin
g
a su
b
im
a
g
e in
place without
using any ext
r
a worki
ng a
r
ray. They over
write pixel
values with in
te
rpolated values
.
Only
linear inte
rpol
ation is c
o
nsidered and t
h
e c
o
rrectness
fo
r lar
g
e w
i
ndo
w sizes
is
no
t guar
a
n
t
eed
.
Yo
u G
u
ang
C
h
en
et
al
[
1
2
]
pr
op
ose
d
a
m
e
t
hod
f
o
r
d
o
c
um
ent
ori
e
nt
a
t
i
on
det
ect
i
o
n
and
cl
assi
fi
cat
i
o
n
by
usi
n
g
S
u
p
p
o
rt
Vector Machi
n
e (SVM) a
n
d t
h
en the
or
ien
t
atio
n
o
f
unk
now
n do
cu
m
e
n
t
i
m
ag
es
is classified. Vee
n
a et
al [13]
pr
o
pose
d
a ske
w
co
rrect
i
o
n f
o
l
l
o
we
d
by
ori
e
nt
at
i
on
det
ect
i
on a
nd c
o
r
r
ec
t
i
on o
f
ve
hi
cl
e num
ber pl
at
e.
The
pr
o
pose
d
hy
bri
d
m
odel
wo
rk
i
n
t
w
o st
a
g
es.
It
fi
rst
det
ect
s t
h
e ske
w
usi
n
g
R
a
do
n t
r
a
n
sf
or
m
a
t
i
on and t
h
e
n
t
h
e
doc
um
ent
ori
e
nt
at
i
on i
s
det
ect
ed usi
ng a
u
t
o
co
rrel
a
t
i
o
n.
Thi
s
ap
p
r
oac
h
wo
rk
s o
n
s
k
ew an
d
ori
e
nt
at
i
on
det
ect
i
on i
n
ve
hi
cl
e num
ber pl
at
e im
ages
wi
t
h
ori
e
nt
at
i
ons
, t
h
e
m
e
t
hod i
s
i
d
eal
fo
r
t
h
e im
ages wit
h
very
sm
al
l
num
ber
of
cha
r
act
ers
a
n
d
can
n
o
t
be
e
x
t
e
n
d
ed
f
o
r
i
m
ages
wi
t
h
l
a
r
g
e
am
ount
o
f
t
e
xt
.
It is ob
serv
ed
th
at few of th
e
ap
pro
ach
es repo
rted
in
t
h
e literatu
re are
d
e
vised
p
a
rtly b
y
e
m
p
l
o
y
in
g
kn
o
w
l
e
d
g
e
bas
e
an
d
rem
a
i
n
i
ng a
r
e m
i
cro l
e
v
e
l
feat
u
r
e
base
d a
p
pr
oache
s
.
The
use
o
f
d
o
m
ai
n k
n
o
w
l
e
d
g
e
bas
e
or e
x
t
r
act
i
o
n
o
f
m
i
cro l
e
vel
feat
ures
usi
n
g
poi
nt
p
r
oces
si
ng
o
r
bl
ock
p
r
oces
si
n
g
m
a
y sl
ow t
h
i
s
ve
r
y
basi
c
ope
rat
i
o
n i
.
e.,
det
ect
i
on
of t
e
xt
ori
e
nt
at
i
on a
nd c
o
r
r
ect
i
n
g
it. Th
is m
o
tiv
ate
s
u
s
to
d
e
v
i
se a
m
acro
lev
e
l featu
r
e
ap
pro
ach
witho
u
t
em
p
l
o
y
in
g an
y p
r
e-b
u
ilt k
nowledg
e b
a
se for d
e
tection o
f
fli
p
p
e
d
docu
m
en
ts. Sectio
n
3
descri
bes t
h
e
m
e
t
hod
ol
o
g
y
pr
o
pose
d
f
o
r t
h
e det
ect
i
o
n a
nd c
o
r
r
ect
i
o
n of
fl
i
ppe
d o
r
i
e
nt
at
i
on i
n
do
c
u
m
e
nt
im
ages.
3.
PROP
OSE
D
METHO
D
The
det
ect
i
o
n
of t
e
xt
i
m
age ori
e
nt
at
i
on i
n
t
h
e
pr
op
ose
d
w
o
r
k
ass
u
m
e
s t
h
e i
n
p
u
t
as a
pl
ai
n E
ngl
i
s
h
tex
t
do
cu
m
e
n
t
. In
itially th
e in
pu
t im
ag
e is su
bj
ect t
o
p
r
e-p
r
o
cessi
n
g
proced
ures
fo
llo
wed
b
y
th
e d
i
rectio
n
a
l
gra
d
i
e
nt
feat
u
r
e com
put
at
i
o
n
.
Fi
na
lly the
features c
o
m
p
uted are
subj
ec
ted to a
m
u
lti-level
uns
upervised
classifier th
at
p
r
ed
icts th
e orien
t
atio
n
o
f
t
e
xt
i
n
t
h
e
i
m
age. The
bl
oc
k
di
ag
ram
of t
h
e
t
e
xt
im
age o
r
i
e
nt
at
i
on i
s
sho
w
n i
n
t
h
e
F
i
gu
re
2.
Fi
gu
re
2.
B
l
oc
k
di
ag
ram
for t
e
xt
o
r
i
e
nt
at
i
o
n
det
ect
i
o
n
The i
n
vocat
i
o
n
of t
e
xt
ori
e
nt
at
i
on det
ect
i
o
n
al
gori
t
h
m
beg
i
ns by
acqui
ri
ng t
e
xt
i
m
age
as an i
n
p
u
t
.
The i
n
p
u
t
im
age can
be rea
d
i
n
ei
t
h
er 0
0
,
9
0
0
,
1
8
00
or
2
7
00
ori
e
nt
at
i
on
of t
e
xt
. The c
o
m
put
at
i
on of
f
eat
ures
an
d m
u
lti-lev
e
l
classificatio
n
o
f
tex
t
o
r
ien
t
at
io
n
is elu
c
id
ated
in th
e
sub
sectio
n
s
A and
B
.
3.
1.
Feat
ure An
al
y
s
i
s
The gray scale
im
age is proce
ssed
to
ob
tain
th
e g
r
ad
ien
t
of th
e i
m
ag
e [1
4
]
.
Th
e grad
ien
t
in
fo
rm
atio
n
of
an
i
m
age wi
t
h
re
spect
t
o
a
t
e
xt
ori
e
nt
at
i
o
n
poi
nt
s t
o
larg
est
po
ssib
l
e in
tensity
in
crease in th
at tex
t
d
i
rect
io
n.
Each
pi
xel
of a
gra
d
i
e
nt
i
m
age m
easures t
h
e
cha
nge i
n
i
n
t
e
nsi
t
y
of t
h
e sa
m
e
poi
nt
i
n
t
h
e
ori
g
i
n
al
i
m
age wi
t
h
reg
a
rd to
t
h
e
orien
t
atio
n of tex
t
. Th
e m
a
g
n
itu
d
e
of th
e grad
ien
t
represen
ts ho
w rap
i
d
l
y t
h
e in
ten
s
ity chan
g
e
s
fr
om
one
poi
nt
t
o
a
not
her
p
o
i
n
t
i
n
t
h
e c
o
r
r
es
po
n
d
i
n
g
di
rect
i
o
n
.
T
h
e
gra
d
i
e
nt
o
f
t
h
e i
m
age i
s
gi
ven
by
(1
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
An
Un
sup
e
rvised
Cla
s
sifica
tio
n Tech
n
i
q
u
e
fo
r
Detectio
n o
f
Flip
p
e
d
Orienta
tio
n
i
n
.... (Vija
ya
sh
ree C
.
S
.
)
2
143
x
y
f
g
x
f
f
g
y
(1
)
whe
r
e
f
x
is th
e g
r
ad
ien
t
in
th
e x
-
d
i
rection
and
f
y
is th
e g
r
ad
ien
t
in
th
e y-d
i
rectio
n
.
Th
e g
r
ad
ien
t
d
i
rection
i
s
com
put
ed
by
(
2
)
1
ta
n
y
x
g
g
(2
)
Th
e
g
r
ad
ien
t
i
n
form
at
io
n
o
f
a
typ
i
cal tex
t
i
m
ag
e
with
resp
ect to
0
0
, 90
0
, 180
0
a
nd 2
7
0
0
orien
t
atio
n
s
is
rep
r
ese
n
t
e
d
i
n
Fi
gu
re 3.
Fi
gu
re 3.
Te
xt
gra
d
i
e
nt
at
0
0
, 1
8
0
0
, 90
0
an
d 27
0
0
Th
e gr
ad
ien
t
featu
r
es
o
f
input tex
t
i
m
ag
es a
r
e co
m
p
u
t
ed
with
r
e
sp
ect to
fo
ur
po
ssi
b
l
e orien
t
atio
n
s
.
The
gra
d
ient f
eatures a
r
e int
e
rp
reted
fu
rthe
r f
o
r the
se
lection
of a
g
gre
g
a
t
e features
. The feature
selection is
accom
p
lished
via summ
arizi
ng t
h
e
gra
d
ient features t
o
statistical quantities like kurt
osis and sum
.
The
statistical
q
u
a
ntities
rep
r
esen
t
th
e ag
g
r
eg
ated
g
r
ad
ien
t
f
eat
u
r
es
o
f
th
e im
a
g
e. Th
e
featu
r
es o
f
in
pu
t i
m
a
g
e are
com
put
ed f
o
r
al
l
possi
bl
e f
o
ur
ori
e
nt
at
i
o
ns
. Furt
her t
h
e
s
e
lected feature
s
are forw
a
r
de
d f
o
r cl
assi
fi
ca
t
i
on t
o
d
e
tect th
e
o
r
ien
t
atio
n
of tex
t
i
n
th
e inp
u
t
im
ag
e.
In
itially, th
e alg
o
rith
m
ass
u
m
e
s th
e in
pu
t do
cu
m
e
n
t
o
r
ien
t
atio
n
as
1
R
irresp
ectiv
e
o
f
its tex
t
ori
e
nt
at
i
on an
d o
b
t
a
i
n
t
h
e r
o
t
a
t
e
d ve
rsi
o
n
s
of i
n
put
d
o
c
um
ent
i
n
ot
her t
h
ree o
r
i
e
nt
at
i
ons
23
,
RR
and
4
R
resp
ectiv
ely. Fig
u
re
4
sh
ows th
e inpu
t im
ag
e and
its
v
a
riou
s ro
tated
v
e
rsi
o
n
s
for co
m
p
rehen
s
ion
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 5
,
O
c
tob
e
r
20
16
:
214
0
–
21
49
2
144
Fi
gu
re
4.
I
n
p
u
t
im
age an
d i
t
s
rot
a
t
e
d
ve
rsi
o
n
s
From
Fi
gure
4
,
i
t
i
s
evi
d
ent
t
h
at
pr
ocessi
ng
of an
input image conside
r
s the im
age at all its rotated
ori
e
nt
at
i
ons a
s
i
ndi
cat
ed
. T
h
e
feat
u
r
es
of
t
h
e rot
a
t
e
d v
e
rsi
ons
are
em
pl
oy
ed f
o
r
det
ect
i
o
n
o
f
t
e
xt
o
r
i
e
nt
at
i
o
n
d
i
rection
in th
e in
pu
t im
ag
e
1
R
.
Let
I
is a gray scale im
age a
nd
12
3
,,
RR
R
and
4
R
repres
ents the rotated im
ages of
I
in
four
di
ffe
re
nt
ori
e
n
t
at
i
ons res
p
ect
i
v
el
y
.
The
12
3
,,
GG
G
and
4
G
represents the
gra
d
ient features com
puted
from
12
3
,,
RR
R
and
4
R
and
ii
SK
is the features sel
ected from
the
com
puted
gradie
nt feature
s
for t
h
e
th
i
ori
e
nt
at
i
on, wh
ere
1,
2
,
3
,
4
i
. The propose
d
m
odel for feature analys
is and selection proces
s of a
n
im
age
i
s
sh
ow
n i
n
Fi
g
u
re
5
.
Fi
gu
re
5.
Pr
o
p
o
se
d m
odel
f
o
r
feat
u
r
e a
n
al
y
s
i
s
an
d sel
ect
i
o
n
The pr
o
p
o
s
ed
al
go
ri
t
h
m
dem
ons
t
r
ates the
process
of feat
ur
e com
putation
and selection.
Al
g
o
ri
thm
_
Fe
atu
re_
An
al
ysi
s
1.
Read a
n
im
age
I
2.
Ob
tain th
e
ro
tated
im
ag
es
12
3
,,
RR
R
and
4
R
of
I
.
3.
C
o
m
put
e gra
d
i
e
nt
feat
ure
vec
t
ors
i
G
wh
er
e
1,
2
,
3
,
4
i
o
f
ro
tated
im
ag
es.
4.
C
o
m
put
e
i
S
, t
h
e
sum
of
gra
d
i
e
n
t
s an
d
i
K
th
e
ku
rto
s
is of
g
r
ad
ien
t
s wh
ere
1,
2
,
3
,
4
i
.
whe
r
e
1
n
ij
j
i
SG
a
n
d
4
1
[]
1
()
n
ji
i
j
i
G
K
nG
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
An
Un
sup
e
rvised
Cla
s
sifica
tio
n Tech
n
i
q
u
e
fo
r
Detectio
n o
f
Flip
p
e
d
Orienta
tio
n
i
n
.... (Vija
ya
sh
ree C
.
S
.
)
2
145
‘
n
’is th
e leng
th
o
f
grad
ien
t
and
[]
j
i
G
is a feat
ure
val
u
e at inde
x
’j’
of gradie
nt feat
ure
vect
or
i
G
wh
er
e,
1,
2
,
3
,
4
i
.
‘
’ an
d
()
i
G
is th
e mean
an
d stand
a
rd d
e
v
i
atio
n of
th
e grad
ien
t
i
G
with
1,
2
,
3
,
4
i
res
p
ectivel
y.
5.
C
o
m
put
e t
h
e
O
r
i
e
nt
ation coe
f
ficient
i
=
ii
SK
i
G
wh
er
e
1,
2
,
3
,
4
i
.
6.
Proceed
i
i
G
for
de
cision a
n
alysis, where
1,
2
,
3
,
4
i
.
The orientation coe
fficient
i
is t
h
e pr
o
duct
o
f
sum
and k
u
rt
osi
s
feat
u
r
es
()
ii
SK
for each
i
G
and
i
s
fu
rt
her
di
rec
t
ed f
o
r
det
ect
i
o
n o
f
t
e
xt
ori
e
nt
at
i
on. T
h
e
deci
si
on a
n
al
y
s
i
s
f
o
r
det
ect
i
on
o
f
t
e
xt
ori
e
nt
at
i
o
n
i
s
as
di
scuss
e
d
i
n
t
h
e su
bse
que
nt
s
ect
i
on.
3.
2.
Deci
si
on
a
n
al
ysi
s
f
o
r
detec
t
i
o
n
of
te
xt
ori
e
nt
ati
o
n
Decision anal
ysis is one of the crucial proce
d
ures in
any im
age processing syste
m
and also
considere
d
as t
h
e final stage t
h
at decides t
h
e
efficiency
of
th
e system
. I
n
t
h
e pr
opo
sed
syste
m
th
e o
r
ien
t
atio
n
detection
of text in a
n
im
age is done t
h
rough c
o
nstr
ucting a
decision tree, each le
vel
fulfills the c
r
iteria for
one
o
f
t
h
e t
e
xt
ori
e
nt
at
i
on.
T
h
e deci
si
o
n
rul
e
s are
de
vi
sed
b
y
i
d
ent
i
f
y
i
n
g
a
set
o
f
i
n
fere
nc
es f
r
om
t
h
e fea
t
ure
s
selected in the feature a
n
alysis stage. In
re
ga
rd wi
t
h
t
h
e
ori
e
nt
at
i
on coe
ffi
ci
ent
i
d
e
term
in
ed
at o
r
ien
t
ation
s
1,
2
,
3
,
4
i
t
h
e i
n
fere
nces
1,
2,
3
an
d
4 a
r
e de
ri
ve
d f
o
r
d
e
t
ect
i
on
of t
e
xt
o
r
i
e
nt
at
i
o
n
.
Infere
nce1:
13
&&
24
Infere
nce 2:
13
2
4
&&
Infere
nce 3:
13
2
4
&&
Infere
nce 4:
13
2
4
&&
The Fi
g
u
re
6 depi
ct
s t
h
e de
ci
si
on t
r
ee f
o
r
t
e
xt
ori
e
nt
at
i
o
n det
ect
i
on
by
em
pl
oy
i
ng t
h
e i
n
fere
nce
s
obt
ai
ne
d
fr
om
ori
e
nt
at
i
on c
o
e
ffi
ci
ent
i
at
1,
2
,
3
,
4
i
.
Fi
gu
re
6.
Pr
o
p
o
se
d m
odel
f
o
r
t
e
xt
o
r
i
e
nt
at
i
o
n
det
ect
i
o
n
At each le
vel of
decision, t
h
e analysis is perform
e
d with
the observe
d
i
n
fe
rence
s
to
determine the
o
r
ien
t
atio
n
o
f
th
e tex
t
in
th
e i
m
ag
e. Th
e ob
serv
atio
ns drawn
from
the
selected
are e
m
ployed as decision
m
a
kers
fo
r t
h
e det
ect
i
o
n
of
t
e
xt
ori
e
nt
at
i
o
n
.
T
h
e
ex
per
i
m
e
nt
al
i
n
fere
nces
d
r
aw
n
w
i
t
h
res
p
ect
t
o
pr
o
duct
feature
s
i
at
1,
2
,
3
,
4
i
and
i
n
fere
nce
1 i
s
p
l
ot
t
e
d as
Fi
g
u
r
e
7 a
n
d Fi
gu
re
8.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 5
,
O
c
tob
e
r
20
16
:
214
0
–
21
49
2
146
Figure
7.
Orie
ntation c
o
efficients-
00
01
8
0
Fi
gu
re
8.
O
r
i
e
nt
a
t
i
on c
o
ef
fi
ci
ent
s
-
00
90
2
7
0
The i
n
fe
rence
s
o
f
e
xpe
ri
m
e
ntat
i
on
fo
r i
n
fer
e
nce
2 a
r
e
pl
ot
t
e
d as
Fi
g
u
re
9
a
n
d
Fi
g
u
r
e
10
.
Figure
9.
Orie
ntation coefficients-
13
Figure
10.
Orie
ntation c
o
efficients-
24
The i
n
fe
rence
s
o
f
e
xpe
ri
m
e
ntat
i
on
fo
r i
n
fer
e
nce
3 a
r
e
pl
ot
t
e
d as
Fi
g
u
re
1
1
and
Fi
g
u
r
e
12
.
Figure
11.
Orie
ntation
coefficients-
13
Fi
gure
12.
O
r
ientation coe
f
f
i
cients-
24
The i
n
fere
nces
of
ex
peri
m
e
nt
at
i
on f
o
r i
n
fere
nce
4 are
pl
ot
t
e
d as Fi
g
u
re
1
3
a
nd
Fi
g
u
re
1
4
. F
r
om
t
h
e
i
n
t
e
rp
ret
a
t
i
on
of ex
pe
ri
m
e
ntal
i
n
ference
s
dra
w
n fr
om
Fi
gu
re 6 t
h
r
o
ug
h Fi
g
u
re 1
3
, i
t
i
s
evi
d
ent
t
h
at
t
e
xt
ori
e
nt
at
i
on i
s
d
i
scrim
i
nant
f
r
o
m
one cl
ass t
o
anot
her
cl
ass
wi
t
h
respect
t
o
t
h
e deci
si
on
r
u
l
e
s i
d
ent
i
f
i
e
d.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
An
Un
sup
e
rvised
Cla
s
sifica
tio
n Tech
n
i
q
u
e
fo
r
Detectio
n o
f
Flip
p
e
d
Orienta
tio
n
i
n
.... (Vija
ya
sh
ree C
.
S
.
)
2
147
Figure 13. Orie
ntation coefficients-
13
Figure
14.
Orie
ntation c
o
efficients-
24
4.
EX
PER
I
M
E
NTA
L
ANA
LYSIS
The p
r
o
p
o
se
d m
e
t
hod
ol
o
g
y
f
o
r t
e
xt
o
r
i
e
nt
at
i
on det
ect
i
o
n, t
h
e ex
peri
m
e
nt
at
i
on i
s
per
f
o
r
m
e
d usi
ng t
h
e
d
a
tasets th
at in
clud
es 80
prin
ted
Eng
lish
d
o
c
u
m
en
ts an
d
50
Telug
u
a
p
p
lica
tion
form
doc
u
m
e
nts
.
Each
d
o
c
u
m
en
t is tested
wit
h
all th
e
o
r
ien
t
atio
n
s
00
0
0
,
90
,
180
and
0
270
. The
expe
ri
m
e
nt
al
anal
y
s
i
s
of
t
h
e
pr
o
pose
d
al
go
r
i
t
h
m
i
s
di
scuss
e
d as
f
o
l
l
o
w
s
.
If
ea
D
represen
ts th
e to
tal nu
m
b
er of
d
o
c
u
m
e
n
ts em
p
l
o
y
ed
for exp
e
rim
e
n
t
atio
n
in
wh
ich
‘
e
’
i
ndi
cat
es
pri
n
t
e
d E
n
gl
i
s
h
d
o
c
u
m
e
nt
s and
‘
a
’i
ndi
cat
es t
h
e a
p
pl
i
cat
i
on
fo
rm
doc
um
ent
s
i
n
ea
D
. The num
b
er
of rec
o
gni
ze
d
pri
n
t
e
d En
gl
i
s
h doc
um
ent
s
i
s
gi
ven by
e
N
and rec
o
gni
ze
d
ap
pl
i
c
at
i
on f
o
rm
doc
um
ent
s
rep
r
ese
n
ts
a
N
, t
h
e
n
t
h
e
t
e
xt
ori
e
n
t
at
i
on det
ect
i
o
n
rat
e
i
s
gi
ve
n
by
eq
uat
i
o
n
(1
)
.
Text
ori
e
nt
at
i
o
n rat
e
,
ea
or
ie
nt
ea
NN
T
D
(1
)
If th
e false po
sitiv
e rate o
f
d
o
cu
m
e
n
t
reco
gnized
is g
i
v
e
n
by
()
ea
FN
N
, th
en
th
e true p
o
s
itiv
e
rat
e
o
f
ori
e
nt
at
i
on
det
ect
i
o
n i
s
gi
ve
n
by
()
ea
TN
N
an
d is d
e
p
i
cted
i
n
eq
u
a
tion
(2
). Tru
e
p
o
sitiv
e rate,
[(
)
(
)
]
()
ea
ea
ea
ea
NN
F
N
N
TN
N
D
(2)
The e
x
peri
m
e
nt
al
st
at
i
s
t
i
c
s of
t
h
e p
r
op
ose
d
m
e
t
hod
ol
o
g
y
i
s
t
a
b
u
l
a
t
e
d i
n
Tabl
e
1.
Tabl
e
1. E
x
per
i
m
e
nt
al
anal
y
s
is o
f
fl
i
ppe
d
ori
e
nt
at
i
on
det
ect
i
o
n
Or
ientation
N
u
mb
e
r
o
f
S
a
mp
l
e
s
Or
ientation
Detection Rate
T
or
ient
True Positive
Rate
T(N
e
+N
a
)
E
nglish
Docu
m
e
nt
T
e
lugu
Docu
m
e
nt
0
0
80x4
50x4
96.
15%
93.
84%
90
0
80x4
50x4
90.
76%
87.
69%
180
0
80x4
50x4
96.
92%
94.
61%
270
0
80x4
50x4
92.
30%
88.
46%
So
m
e
o
f
th
e
in
pu
t do
cu
m
e
n
t
i
m
ag
es considere
d
and t
h
e orienta
tion direction det
ected by the
pr
o
pose
d
m
e
t
hod
ol
o
g
y
i
s
l
i
s
t
e
d i
n
t
h
e
Tabl
e
2.
Th
e m
e
th
o
d
o
l
og
y fails to
wo
rk
fo
r so
m
e
cases lik
e do
cu
m
e
n
t
im
ag
es with
lo
w
reso
lu
tion. It fu
rt
h
e
r
fails wh
ere th
e do
cu
m
e
n
t
i
m
a
g
es
h
a
ve m
o
re nu
m
b
er of
p
i
ctu
r
es/em
b
le
m
s
an
d fo
r do
cu
men
t
s wit
h
v
e
ry
litt
le
t
e
xt
. If t
h
e d
o
c
u
m
e
nt
im
age has vary
i
n
g fo
nt
s, t
h
i
s
wo
ul
d al
so res
u
l
t
i
n
wr
on
g o
r
i
e
nt
at
i
o
n
di
rect
i
on det
e
ct
i
on.
Fig
u
re
15
illu
strates so
m
e
o
f
in
stan
ces of i
n
pu
t im
ag
es wh
ere th
e
d
e
tectio
n
o
f
orien
t
ation
fails.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 5
,
O
c
tob
e
r
20
16
:
214
0
–
21
49
2
148
Tabl
e 2. Sam
p
les
o
f
doc
um
ent
im
ag
es with detected
orien
t
atio
n
Fi
gu
re 1
5
. Sam
p
l
e
s of d
o
cum
e
nt
w
h
i
c
h
res
u
l
t
e
d
i
n
w
r
on
g ori
e
nt
at
i
on det
ect
i
o
n
5.
CO
NCL
USI
O
NS
The w
o
r
k
p
r
es
ent
e
d det
ect
s t
h
e fl
i
p
ped
ori
e
nt
at
i
on di
rect
i
on o
f
d
o
cum
e
nt
im
age. The
m
e
t
hod i
s
foc
u
se
d f
o
r
det
ect
i
ng d
o
c
u
m
e
nt
ori
e
nt
at
i
on
f
o
r t
e
xt
do
cum
e
nt
s. T
h
e m
e
t
hodol
ogy
has be
e
n
t
e
st
ed f
o
r En
gl
i
s
h
text doc
um
ents and
Telugu te
xt docum
e
nts.
Based on th
e s
a
m
p
les tested,
this m
e
thod is
success
f
ul
on
m
o
st of
t
h
e t
e
xt
d
o
cu
m
e
nt
im
ages i
n
cl
u
d
i
n
g t
hose
w
h
i
c
h c
o
nt
ai
n sm
all e
m
ble
m
s. Doc
u
m
e
nt images in t
h
e
form
of
appl
i
cat
i
o
n fo
rm
s have al
so bee
n
succe
ssful
l
y
t
e
st
ed.
Thi
s
m
e
t
hod
of ori
e
nt
at
i
o
n det
ect
i
on
p
e
rf
orm
s
effectively wit
h
an overall efficien
cy o
f
94
%.Th
e
m
e
th
od
do
es no
t
r
e
qu
ire a
n
y speci
al hardwa
re se
tup t
o
acqui
re t
h
e
do
cum
e
nt
im
age, t
h
e im
age o
b
t
a
i
n
ed
by
fl
at
-
b
ed sca
nne
r i
s
s
u
f
f
i
c
i
e
nt
. T
h
i
s
m
e
t
hod ca
n
ve
ry
wel
l
be e
x
t
e
n
d
ed
f
o
r
ot
her
l
a
n
gua
g
e
t
e
xt
d
o
c
u
m
e
nt
im
ages.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
An
Un
sup
e
rvised
Cla
s
sifica
tio
n Tech
n
i
q
u
e
fo
r
Detectio
n o
f
Flip
p
e
d
Orienta
tio
n
i
n
.... (Vija
ya
sh
ree C
.
S
.
)
2
149
REFERE
NC
ES
[1]
Srihari S. N., “Handwritten
ad
dress interpr
e
tat
i
on:
A
task of
m
a
n
y
p
a
t
t
ern r
ecogni
tion prob
l
e
m
,
”
Journal of
Pattern
Recognition and
Artificia
l Intelligence
, vo
l/issue: 14
(5), pp
. 663-674
, 2000
.
[2]
R. C. Gonzales
and R.
E. Woods, “D
igital Image
Processing,” 2nd
Edition,
Pearso
n Education
Publication
,
2002
.
[3]
L.
O’Gorman, R.
Kasturi,
“
E
x
e
c
u
tive
Brief
i
ng:
Docum
e
nt Im
ag
e Anal
ys
is
,
”
I
E
E
E
Computer
So
c
i
et
y Pr
es
s
, 1998.
[4]
Caprari R. S., “Algorithm for
Text Page
Up/Down Orientation Determinatio
n,”
Pattern Recognition Letters
,
vol/issue: 21(4), pp.
311-317
,
20
00.
[5]
O. Nes
t
ares
, “
A
utom
atic Dom
i
n
a
nt Orien
t
at
ion
Es
tim
ation in
Te
xt Im
ages
Bas
e
d
on S
t
eerab
le F
i
l
t
ers
,
” IP
C8 Cl
as
s
:
AG06K932FI, USPC Class: 38229, Agents: C
OOL PATENT, P.C
.
;c/o CPA Global, 2009
.
[6]
A.
Vailay
a
,
et a
l
.
, “
A
utom
atic I
m
age Orienta
tio
n Detec
tion,
”
IEEE transactions
on Image Processing,
vol/issue
:
11(7), 2002
.
[7]
D.
X.
Le,
et al.
, “Automated Page Orien
t
ati
on and Skew Angle
Detection fo
r
Bi
nar
y
Docum
e
nt
Im
ages
,”
Patt
ern
Recogn
ition
, vo
l/issue: 27(10)
, p
p
. 1325-1344
, 1
994.
[8]
B. T. Avila
an
d R. D. Lins, “A
fast orientation and skew d
e
tecti
on algorith
m for
monochromatic document
im
ages
,”
Proceedings of
the AC
M sympos
ium on Document engineering,
pp. 118
-126, 2005
.
[9]
Vasudev T. an
d P. Nagabhus
han, “Detection
and Correctio
n of Page
Orientation
in Monochrome Textu
a
l
Docum
e
nt Im
ag
e,”
Journal of C
o
mputer Science
, vol/issue: 2(1),
pp. 465-473
.
[10]
Murali S.,
et al.
, “Language Ind
e
penden
t
Skew De
tection and C
o
rrection of Printed Tex
t
Docu
ment Images: A
Non-rotation
a
l Approach,
”
Int
e
rnational journa
l of
Arti
ficial
Int
e
llig
ence
, vol/issue: 16(2)
, pp
. 08
-15, 2006
.
[11]
T. Asano
,
et a
l
.
, “In-place
algor
ithm
for image
rotation
,
”
Algorith
m
s and Computation
, pp. 704-71
5, 2007
.
[12]
Y.
G.
Chen,
et al.
, “
A
m
e
tho
d
for de
tec
ting
docum
ent or
ie
ntation
b
y
usin
g SVM classifi
er,”
Int
e
rnation
a
l
conferen
ce
on M
u
ltimedia
Techn
o
logy (
I
CMT)
, pp. 47-50
, 2011
.
[13]
Veena M
.
N.
,
et al.
, “Detection of oriented and skewed nu
mber
plate in vehicle
using
h
y
brid approach,”
International Jo
urnal of multim
edia and
its applications (
I
JMA)
, vol/issue: 7(4/5),
2015.
[14]
C. L.
Liu, “Normalization-coop
erated gr
adien
t
feat
ur
e
extractio
n for handwritt
en ch
ara
c
ter
re
c
ognition,
”
IE
EE
transactions on
pattern ana
lysis
and machine in
t
e
llig
ence
, vol/issue: 29(8)
, 2007
.
BIOGRAP
HI
ES OF
AUTH
ORS
Vijay
a
shree C. S, obtained her
B.E. Degr
ee in
Computer Scien
ce from B.I.T,
Bangalor
e
and
M
.
E. Degr
ee
in
Com
puter S
c
ien
ce from
U.V.C
.
E, Bang
alor
e. S
h
e is
purs
u
ing r
e
s
earch
towards
her P
h
. D. Deg
r
ee in
Com
puter
S
c
ienc
e of Univ
ers
i
t
y
of M
y
s
o
re
, M
y
s
o
re
,
at P
.
E
.
S
.
Coll
ege o
f
Engineering, Mand
y
a
.
N. Shobha Rani is currently
wo
rking as Assistant professor in
Amrita
Vishwa Vid
y
ap
eetham
University
, M
y
sore an also pursuing her Ph
.D degree
in Maharaja R
e
sear
ch Foundation,
Univers
i
t
y
of
M
y
s
o
re, M
y
s
o
r
e
. Her ar
eas
of
interes
t
includ
e Docum
e
nt im
age proc
es
s
i
ng,
Optica
l
ch
ara
c
t
e
r recogn
ition
an
d Com
puter visi
on.
Dr. Vasudev
T i
s
Professor, in t
h
e Depar
t
m
e
nt o
f
Com
puter App
lic
ations,
Mahar
a
ja
Institu
te o
f
Techno
log
y
, My
sore. He obtained his Bachelor
of Science and post graduate diploma in
computer programming with two Masters Degrees
one in Computer Appli
cations
and oth
e
r one
is Computer Science and Tech
nolog
y
.
He wa
s awarded Ph.D. in Computer Science from
University
of My
sore. He is having 30
y
e
ars of
e
xperien
ce in
ac
a
d
em
ics
and his
a
r
ea of res
e
arch
is Digital Image Processing sp
ecif
i
cally
docu
m
ent image processing and authored about 50
res
earch
pap
e
rs
.
Evaluation Warning : The document was created with Spire.PDF for Python.