![](bg2.png)
(2003) proposed a method of matching between the
input image and the stored template without
resorting to exhaustive search using both the wavelet
statistical features and wavelet co-occurrence
features. However, it uses all the pixels in the
wavelet sub-band image for the computation of the
statistical features, and it is much time-consuming.
Tico et al. (2001) suggested another matching
algorithm using wavelet domain features. They used
a feature vector of length 12 to represent a
fingerprint image. The feature vector represents an
approximation of the image energy distribution over
different scales and orientations. Fung et al. (2004)
proposed an improved approach of ref (Tico, et al.
2001). In their work, critical wavelet coefficients
were selected to form a feature vector of a
fingerprint image. However, the vector with 12
features is not sufficient to use all the information of
a fingerprint image so that the recognition rate may
not be appropriate for some applications. Lee W.K.
et al. (1997) proposed an algorithm that extracts the
dominant local orientation features in the wavelet
transform domain. The performance of the
algorithm is directly related to the accuracy of the
detection of the local directions. Mokju et al. (2004)
proposed an algorithm based on directional image
constructed using the expanded Haar Wavelet
Transform. In the work, they first obtain a
directional image, and then quantize the directional
image into a few grey-level values that represent a
range of angle orientations. In this method, the
quantizing process may be error-prone in computing
the directional information.
To overcome the drawbacks of these methods, a
new matching method is proposed in this work. We
use a sophisticated FingerCode method in the
wavelet transform domain for fingerprint
recognition. In the work, FingerCode are extracted
in the decomposed wavelet sub-band images instead
of the original fingerprint image. There are two
advantages to extract features from the wavelet sub-
band images. Since the wavelet transform is a multi-
resolution tool in signal processing, it can easily
remove the high-frequency noise, usually contained
in HH sub-band image. With this approach one can
eliminate some pre-processing steps such as noise
removing, binarization, thinning and restoration. In
addition, the size of decomposed sub-band images is
half of the original image, so that a matching method
using features extracted from sub-band images can
speed up the whole matching process comparing to
other approaches.
The paper is organized as follows: In Section 2
The theory of FingerCode is briefly reviewed. The
proposed recognition method is explained in Section
3 and its experimental results are shown in Section 4.
The conclusion remarks are given in Section 5.
2 FINGERCODE
The FingerCode, introduced in ref (Jain et al., 2000),
is a fixed length representation that can effectively
capture both the local and global details in a
fingerprint, with a bank of Gabor filters. The typical
FingerCode generation process can be summarized
in the following steps:
1. Locate the reference point and determine the
region of interest for a fingerprint image.
2. Tessellate the region of interest, centered at the
reference point, into a series of B (=5) concentric
bands and divide each band into k (=16) sectors.
3. Normalize each sector to a predetermined
constant mean M
0
(=100) and variance V
0
(=100).
4. Filter the region of interest in eight different
directions using a bank of Gabor filters.
5. Computer the average absolute deviation from
the mean (AAD) of grey level values in each of the
80 sectors for every filtered image. The collection of
all the AAD features in each filtered image is
defined as FingerCode.
6. Rotate the features in the FingerCode cyclically
to generate five templates corresponding to five
rotations (±45
0
, ±22.5
0
, 0
0
) of the original fingerprint
image, thus to approximate the rotation-invariance;
7. Rotate the original fingerprint image by an
angle of 11.25
0
and generate its FingerCode.
Another five templates corresponding to five
rotations are generated in the same way as step 6.
8. Match the FingerCode of the input fingerprint
with each of the ten templates stored in the database
to obtain ten matching scores. The final matching
score is the minimum of the ten matching scores,
which corresponds to the best matching of the two
fingerprints.
In this paper, we use the reference point location
method developed in ref (Sha, et al., 2003) for the
original fingerprint image, which is known to be
robust and rotation-invariance. The average
orientation of each sector is also computed for the
reference point.
3 THE PROPOSED ALGORITHM
The proposed algorithm for the fingerprint
recognition consists of three main steps:
VISAPP 2006 - IMAGE UNDERSTANDING
162