Graphics Reference
In-Depth Information
which is based on 2D image matrixes rather than 1D vectors, and the computational
complexity is significantly reduced. And M2DPCA was proposed in [9], by
combining the strengths of both modular PCA and 2DPCA, which performances more
efficient and robust.
In recent years, many researchers have taken the advantages of FPGAs to implement
character recognition for practical applications due to its strengths such as low power
consumption, capability to create customizable portable devices, and foremost, high
performance can be achieved by means of applying the embedded memory modules
and DSP units. Depending on the pipeline and parallel processing, real-time license
plate character recognition can be achieved.
This paper focus on achieve an embedded real-time license plate character
recognition architecture. Three kinds of processing element are designed, the
projection element and distance element are explored to operate the data of image
sub-blocks, and the classification element is presented for nearest neighbor
classification. The arithmetic and logic functions are described in Verilog HDL. Since
the processing elements can be operated in pipeline, FPGA implementations for
character recognition based on M2DPCA could achieve a remarkable high speed.
The rest of this paper is organized as follows. In Section 2, the M2DPCA algorithm
for character recognition is discussed. In Section 3, the design of hardware architecture
is illustrated in detail. In Section 4, experimental results are presented. In Section 5,
Conclusions and ideas for future work are given.
2
Review of M2DPCA
In M2DPCA, each sub-block of character image is represented by a matrix in eigen
space, and new character image is classified corresponding to the closest matching
sample in training database.
2.1
Training Phase
Suppose that
p
classes are included in training database, each class contains
q
images,
every image is divided into
s
sub-blocks (
s
=
b
2
), and the size of images is
m
×
n
. Every
image is a training sample, each sub-block is regarded as an
m
/
b
×
n
/
b
image matrix, and
p
×
q
×
s
matrix are included in the sample space
I
. The
j
th sample of
i
th class in
I
is
expressed as:
I
I
I
i
j
,
11
ij
,
12
i
j
,
b
I
I
I
ij
,
21
i
j
,
22
i
j
,
2
b
I
=
(1)
i
j
I
I
I
ij
,
b
1
ij
,
b
2
i
j
,
bb
The average of all image matrixes in training sample space is computed as:
Search WWH ::
Custom Search