-
英文原文及中文翻译
（一）英文原文
One: A
Method of Slant Correction of Vehicle License
Plate
Based on Watershed Algorithm
In
a vehicle license plate recognition system, slant
vehicle license plate has a bad
effect
on the character segmentation and recognition. A
method of slant correction of
vehicle
license
plate
is
proposed
in
this
paper.
The
method
consists
of
five
main
stages: (1) the
extraction of the boundaries of characters using
watershed algorithm;
(2) dividing the
boundaries of vehicle license plate into small
segments using vertical
differential
method;
(3)
connection
of
the
fracture
characters
using
expansion
and
corrosion; (4) computing
centroids of the left and the right part in the
vehicle license
plate respectively; (5)
finding the slant angle by means of two centroids.
Experimental
results show that the
error rate of using the method is 6.13%, which is
lower than that
of
the
principal
component
analysis.
The
running
time
of
using
this
method
is
less
than that
of Hough transform. The method improves accuracy
of the slant correction.
With the rapid development of highways
and the wide use of vehicles, people have
started to
pay
more and more
attention on
vehicle license plate recognition system.
Vehicle license positioning, extraction
and character segmentation are one of the most
difficult
topics
in
the
vehicle
license plate recognition
system.
Slant vehicle license
plate
has a bad effect on the character segmentation and
recognition. In the last few
years some
achievements in vehicle license positioning and
slant correction have been
obtained.
These
achievements
have
distinguished
effects
in
special
conditions.
However,
under
a
complex
background,
the
effect
of
slant
correction
needs
to
be
enhanced further. Many problems such
as: small contrast, non-uniform illumination,
image
distortion
as
well
as
the
contaminate
dlicense
plate
and
so
on
may
bring
difficulty in slant
correction of vehicle license plate.
This
article presents a method
(called SCWA method) of slant
correction of vehicle license plate based on
watershed
algorithm. As documented in
the experiments of 460 vehicle license plates, the
error
rate of using the SCWA method is
6.13%, which is lower than that of the principal
component
analysis.
The
running
time
of
using
SCWA
method
is
less
than
that
of
Hough transform. Good slant correction
is achieved with SCWA method. The paper is
outlined as follows: section I presents
the introduction, section II describes the SCWA
method and section III presents a
conclusion of the experiments of 460 vehicle
license
images.
II. SCWA METHOD
A.
Extraction of the Boundaries of Characters
Using
Watershed
Algorithm
There
are
many
boundaries
of
characters
in
the
vehicle
1
license plate. These
characters are very important to slant
correction of vehicle license
plate.
The steps
of extraction of the boundaries of
characters
are as
follow:
1) Produce
gradient
image
The
watershed
algorithm
is
sensitive
to
noise
and
has
excessive
segmentation. In order to avoid these
problems,
we apply prewitt
operator to produce
gradient image
of
vehicle
license.
The prewitt
operator is as follow:
where H1 is x direction border, H2 is y
direction border, gradient magnitude is:
Watershed segmentation on
gradient image
The
gradient
magnitude
of
the
gradient
image
of
the
vehicle
license
plate
is
considered
as
a
topographic
surface
for
the
watershed
transformation.
The
idea
of
watershed
segmentation can be viewed as a
landscape immersed in a
lake; catchment
basins will
be filled up with water starting
at each local minimum. Dams must be
built
in order to avoid the merging of
catchment basins. The water shed lines are defined
by
the catchment basins divided by the
dam at the highest level. As a result, watershed
lines
can
separate
individual
catchment
basins
in
the
landscape.
The
result
of
watershed
segmentation
is
shown
in
Figure
1.
The
watershed
segmentation
is
as
follow: Assume that G is
a simple connected graph, the distance between
pixel x and
pixel y in G graph is the
minimal route from pixel x to pixel y, min ( )
h I
refers to
minimal gradient magnitude in license
image I when the altitude is h, hmin and hmax
denote
minimum
and
maximum
in
gradient
magnitude
domain
DI
respectively,
h
value changes from hmin to hmax.
2
Watershed
segmentation orders gradient magnitudes according
to increase and then
scans from hmin to
hmax according to width preferential algorithm.
Step 1. These pixels whose gradient
magnitude is h are marked with a flag sign. The
pixels which are marked with a flag
sign are put into first-in-first-out queue.
Step 2. A pixel
P is got
from the queue. Assume that P’
around
pixel P is the same
flag region as P.
P’ and P are
merged if the
distance between P’ and P is smaller than
the current distance.
Step
3. P' is put into first-in-first-out queue if the
distance between P' and the marked
regions is not computed. P' distance is
that the current distance adds 1.
Step
4.
The
current
distance
adds
1
when
the
computation
of
current
distance
has
finished.
Step 5. Go to step 2 if the queue is
not empty.
Step 6. Sign a new mark for
these pixels which are not handled from step 2 to
step 4
and which are min ( )
h I
.
B.
Dividing
the
Boundaries
of
Vehicle
License
Plate
into
Small
Segments
Using
Vertical
Differential
Method
Respecting
the
more
intensive
density
of
the
vertical
3
edge than the
level edge of vehicle license plate region and the
regular characteristics
of
characters
spacing
of
vehicle
license
plate,
we
divide
the
boundaries
of
vehicle
license
plate into small segments using
vertical differential method
(shown in Fig.2).
where I(i,j) is a matrix of the vehicle
license plate image, G is a border matrix.
C. Connection of the Fracture
Characters Using Expansion and Corrosion Operation
The boundaries of vehicle license plate
are divided into small segments using the
vertical differential method(shown in
Fig. 2). The white area of less than 10 points is
set to background-color in order to
eliminate the boundaries of vehicle license plate.
The fracture characters are connected
by using expansion and corrosion operation.
The erosion operation is defined as:
The expansion operation is defined as:
where I is a
matrix of the vehicle license plate image, B is
structuring element set.
D. Computing
Ccentroids of the Left and the Right Partin the
Vehicle License Plate
Respectively
Assume
that
I
is
an
image
of
vehicle
license
plate
which
contains
m×
n
pixels,
Sum_x1 and Sum_y1 is the sum of X
coordinate value and Y coordinate value of the
white pixel of left part in the image I
respectively, Sum_x2 and Sum_y2 is the sum of
X
coordinate
value
and
Y
coordinate
value
of
the
white
pixel
of
right
part
in
the
image I respectively.
4
Assume that num1 and num2 is the number
of pixels ofthe left and right part in the
image I respectively, (centX1,centY1)
and (centX2,centY2) is the centroids of the left
part and the right part in the image I
respectively.
E. Finding the Slant Angle by Means of
Two Centroids
The connection of two
centroids constitutes a main axes of the license
plate. The
angle between the main axes
and the horizontal is θ(shown
in Fig. 3).
The angle of θ
of counterclockwise rotation
is
：
The transformation matrix of counter-
clockwise rotation is
：
5
The angle of θ of clockwise rotation
is
：
The result of slant correction of
vehicle license plate is shown in Figure 4.
Figure 3. The angle between
the main axis of License plates and horizontal
line. (a)
6
angle of θ of counterclockwise
rotation;(b) the angle of θ of clockwise
rotation.
Figure
4. Slant correction of vehicle license plate
III. CONCLUSIONS
For testing
the MWF algorithm, the experiment of 460vehicle
license plate images
is
carried
on.
The
error
rate
of
slant
correction
of
vehicle
license
plate
using
the
different methods is 6.13% (SCWA
method) and 10.25% (PCA method). Comparison
of the results of SCWA method and PCA
method is shown in Figure conclusion
is that the SCWA method is more
effective than the PCA method. The running time
using this method is less than that one
of Hough transform. Our future work will be to
test rigorously the SCWA method over a
wide variety of images and improve further
accuracy of the slant correction of
vehicle license.
7
Figure5.
Comparison
of
the
results
of
SCWA
method
and
PCA
method.
(a)
the
original Slant Vehicle License Plate;
(b) slant correction of vehicle license plate
using
PCA method. (c) slant correction
of vehicle license plate using SCWA method.
Two
：
A Method of
Slant Correction of Vehicle License Plate
Based on Hough Transform and
Mathematics Morphology
In a real
Vehicle License Plate Recognition System, the
license images obtained by
vidicon
are
usually
slantwise.
The
slant
of
vehicle
licenses
will
do
harm
to
the
Character
Segment
and
Recognition.
The
paper
advances
a
new
method
combining
Hough
Transform
and
Mathematics
Morphology
by
the
analysis
of
the
vehicle
licenses’
slant
pattern
and
the
interference
characteristics.
Compared
with
the
conventional methods, it overcomes the
perplexity that too many disturbed lines and
imperfect
detection
criterions.
The
experimental
results
show
that
the
proposed
method can improve
the accuracy of the slant correction. It is
confirmed that the noise
immunity of
the method is
excellent,
and the performance is
robust. The
correction
8
rate of the
newly developed algorithm has reached over 95%.
The
typical
steps
involved
in
a
video-based
Vehicle
License
Plate
Recognition
System
are
Obtaining
Image,
Plate
Location,
Character
Segment
and
Character
Recognition.
The
obtained
license
image
is
usually
slantwise
and
not
a
normal
rectangle
because
of
the
CCD
vidicon’s
perspective
warps.
The
slant
of
Vehicle
Licenses will do harm to the Character
Segment and Recognition, and it will affect
the accuracy and reliability of the
whole system. Therefore, it is necessary to do
slant
correction
before
character
recognition.
According
to
the
analysis,
there
are
several
characteristics of the slant license
image. The information comprised in the image is
complex,
and
quite
a
number
of
information
is
the
interference.
The
slant
of
the
license mainly reflects
on the horizontal warp. At present, the existing
researches in
Slant
Correction
have
been
developed
on
the
basis
of
Hough
Transform.
Hough
Transform
can
detect
the
plate’s
frame
lines,
obtain
the
incline
information
and
realize the correction.
(1) Combining with Edge Detection, viz. doing edge
detection
firstly before Hough
Transform processing. This method is liable to
infection by the
non-frame lines, and
the veracity is not good. (2) The Longest Line
Detection method
(Yen, 1995). Its idea
in nature is detecting the slant angle of the
longest straight line to
correct the
plate. This method demands a high integrality of
the frame lines. However,
the
plates
in
real
can
hardly
satisfy
the
demands
on
account
of
the
external
disturbance,
and
the
effect
is
also
not
good.
This
paper
proposes
a
new
approach
combining
Hough
Transform
and
Mathematics
Morphology.
The
steps
for
slant
correction
can
be
summed
up
as
the
following:
At
first,
binarize
the
image
of
the
vehicle license, than
using Mathematics Morphology methods to exact the
framework
of
it;
Then,
do
erosion
operation
to
filter
the
portrait
lines
which
interfere
with
the
slant
correction; At last, use Hough Transform and
knowledge reasoning to detect the
transverse parallel lines, reckon the
slant angle of the vehicle license, and design the
rotation
algorithm
adapted
for
the
situation
that
the
rotated
information
region
will
become larger.
Available Lines Picking-up based on
Mathematics Morphology
The
straight
line
detection
using
the
method
of
Hough
Transform
is
subject
to
interference from non-straight line
information. Therefore, Mathematics Morphology
is employed to pick up the available
lines in advance.
Image
Thinning
Generally
speaking,
image
thinning
is
getting
rid
of
some
points
in
the
original
image but holding the former shape of
the objective region. Thinning is the variant of
the
erosion
manipulation
in
nature.
The
course
of
t
hinning
is
to
decide
a
point’s
remove-
or-reserve according to its 8 neighborhood points
continually.
Image Erosion
Because the longitudinal lines in the
thinned image will interfere to the extraction
of
the
available
slantwise
information,
the
erosion
manipulation
is
applied
and
the
structure element
G
=[0]1
×
n
=[
g
1,
g
2,
……
,
g
n]
gi
=0,
i
=1,
……
, n
9