Professional Documents
Culture Documents
Fingerprint Enhancement Using Iterative Contextual Filtering For Fingerprint Matching PDF
Fingerprint Enhancement Using Iterative Contextual Filtering For Fingerprint Matching PDF
Abstract—Fingerprint matching depends on the quality of the The enhancement methods such as contextual filters are
fingerprint images. When fingerprint image quality is low, it can widely used in the fingerprint enhancement process [1]. This
degrade the performance of fingerprint matching significantly. method aims to improve the clarity of ridges and valleys. In
Fingerprint images are often contaminated by noise. Therefore, contextual filtering, the components used to represent
image quality is crucial for fingerprint matching. In this paper, fingerprint patterns are local orientation, frequency, and
an image enhancement algorithm in which contextual filtering quality. These three things are used to create the filter for each
is applied iteratively to a fingerprint image has been proposed. block on the fingerprint image. In the fingerprint enhancement
The main idea of the algorithm is to iterate the output of the literature, various methods of contextual filters have been
Gabor filter to get better enhancement and matching
proposed. They may have different approaches to the filtering
performance. The result of the algorithm has five filtered images
due to five times iteration. It showed that the proposed method
process, but the processes that occur in them tend to be the
is significantly better based on Equal Error Rate (EER) same. Contextual filters have two main purposes. First, they
compared to the Gabor filter and the modified Gabor filter. The tried to connect the small gaps and fills the impurities due to
proposed method surpassed the Gabor filter by 3.08 % and the noises. Second, they tried to escalate the contrast level
modified Gabor filter by 2.95 %. between the ridges and valleys, also isolate lateral linked
ridges.
Keywords—fingerprint enhancement, feature extraction, The Gabor filter proposed by Hong et al. [2] is contextual
fingerprint matching, iterative contextual filtering, fingerprint
filtering that is often used for fingerprint enhancement. Gabor
recognition
filters make use of the frequency and orientation of the
I. INTRODUCTION fingerprint image patterns in the process. The filters have the
optimal combined resolution in the frequency and spatial
The fingerprint matching process heavily relies on the domains. Local ridge frequency is the basis for determining
quality of the fingerprint image. A better-quality fingerprint the frequency of the filter. Meanwhile, local ridge orientation
image will bring up more fingerprint features that matched. is the basis for determining the orientation of the filter. This
Meanwhile, a bad quality image from a fingerprint that is too approach may create spurious ridges and valleys, especially in
dry or wet may create many spurious features and miss several small fingerprint images. An improvement of the Gabor filter
genuine features. Therefore, creating a fingerprint by Garge et al. [3] proposed a modified Gabor filter's kernel
enhancement algorithm that ables the improvement of the to enhance the image quality and computation. This filter uses
quality of noisy fingerprints is crucial. It will help more the wavelength as the feature to create the Gabor kernel.
fingerprint features in fingerprint images to be detected. In the However, the filter's performance had reduced in small-size
end, fingerprint image enhancement can improve fingerprint fingerprint images. Other filtering methods in the spatial
matching performance to increase its accuracy. domain have been proposed by various authors [4] [5] [6].
There are several types of regions in a fingerprint image. Hence, this paper proposes a new enhancement algorithm
The first type is the well-defined region. The fingerprint ridges in which contextual filtering is applied iteratively to a
and valleys in this region are contrasting with each other. The fingerprint image. The main idea of the algorithm is to iterate
second type is the recoverable corrupted region. The the output of the Gabor filter to the first step of the whole
fingerprint ridges and valleys in this region are corrupted by a algorithm. First, apply histogram equalization to eliminate
small amount of noise such as crease, smudges, etc. The last minor noise in the fingerprint image. Second, normalize the
type is the unrecoverable corrupted region. The fingerprint fingerprint image to reduce the variation in gray-level values
ridges and valleys in this region are heavily corrupted by such along ridges and valleys. Then find the region of interest of
a severe amount of noise and distortion that no ridges and the fingerprint image using the segmentation process to reduce
valleys are visible. The fingerprint enhancement algorithm has computation size. Third, before applying the Gabor filter to
a function to enhance these types of regions. In well-defined the fingerprint image, ridge frequency, and orientation needed
regions, the fingerprint pattern has good information. to be determined. Thus, going to be the components for
Therefore, there is no need to do a lot of enhancement in this creating a Gabor filter-bank. The Gabor filter outputs will be
region. In recoverable regions, there are a lot of fingerprint iterated respectively. For each iteration, the filtered images
patterns that can be improved due to noises. Then, the will enter the feature extraction and matching process. In the
algorithm should improve this area as best it can. In an end, the enhancement results and the equal error rate in each
unrecoverable region, there are a lot of fingerprint patterns that iteration are going to be analyzed.
cannot be improved due to the high disruption of noises, so
the algorithm should at least mark this region as unrecoverable
regions.
978-1-6654-0447-1/2021/$31.00
Authorized licensed use limited to: UNIVERSITAS©2021 186on February 09,2023 at 02:37:29 UTC from IEEE Xplore. Restrictions apply.
GADJAH MADA. Downloaded
II. METHODOLOGY the improvement of conventional histogram equalization.
The flowchart of the proposed method showed in Figure The method can limit the noise produced by the enhancement.
1 is divided into several sections: image enhancement, feature 2) Normalization: Normalization is one of the steps in
extraction, and matching. The following sub-sections explain fingerprint enhancement. It uses pixel components in its
the process of each step. The image enhancement process implementation. Therefore, the contrast level between the
contains several steps and will be explained in more detail. ridges and valleys structures does not change. Normalization
also aims to create a new efficient value of the gray level
along the fingerprint ridges and valleys, which facilitates the
next step of processing. The normalization approach used in
this algorithm [2] creates the new gray level values of each
pixel in a fingerprint image as:
, ∙ / ,
, 1
, ∙ / ℎ
where and are the mean and variance of all pixels in the
image. Then and are the desired mean and variance
after the normalization.
3) Segmentation: Segmentation is a step taken to separate
the foreground and the background of the fingerprint image.
The foreground is the part that has a fingerprint pattern, while
the background is the part where there is no fingerprint
pattern in it. Segmentation is necessary to eliminate the
Fig. 1. Flowchart of the proposed enhancement algorithm background of the image and areas that are too noisy. It is
A. Image Enhancement Block based on the calculation of the variance of gray levels. The
fingerprint image segmentation steps are described as
Figure 2 shows the breakdown process of the image
follows. For this algorithm, the image is divided into sub-
enhancement block. The block has four main steps: histogram
blocks of 16x16 size and for each block, the variance will be
equalization, normalization, segmentation, and Gabor filter.
determined. The gray-level variance for a block of size WxW
The block is used to create contextual filtering using the
Gabor filter. was calculated as shown below:
1 ()* ()*
! " # # $% , & " ' 2
-, +,
∅ 6 ∅ 9∅ 4
∅ 9∅ 6 ∅ 5
where ∅ is the orientation of the Gabor filter, is the
Fig. 2. Flowchart of the image enhancement block
1) Histogram Equalization: The fingerprint image has frequency of a sinusoidal plane wave, 3- and 3+ are the space
some minor noise in the first place. This image requires an constants of the Gaussian envelope along and axes,
enhancement to reduce minor noise and increase the clarity respectively. And ∅ , ∅ are the coordinates of , after a
between ridges and valleys from each other. Enhancement is clockwise rotation of the Cartesian axes by an angle of
made by Contrast Limit Adaptive Histogram Equalization 90> ∅ .
(CLAHE) by Pizer et al. [7] and Zuiderveld [8]. CLAHE is
187on February 09,2023 at 02:37:29 UTC from IEEE Xplore. Restrictions apply.
Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded
B. Iteration Block fingerprint is assessed based on a threshold. If the score is
The iteration block uses the Gabor filter as the basis for below the threshold value, it is considered a matched
the process. The output of the filter needs to be iterated to get fingerprint. If the score is higher than the threshold value, it
better results. As an illustration, Figure 3 explains how the is considered a non-matched fingerprint.
iteration process works before going to the next steps like
III. RESULTS AND DISCUSSION
feature extraction and matching.
Our experiment uses OpenCV and scikit-learn in python
on implementing iterative contextual filtering, feature
extraction, and matching. And our method using a computer
with the specification consists of windows 10 64-bit with
GT720m graphics, Intel core i5 gen 4 processor, and 8 GB
RAM.
A. Dataset
Fig. 3. The iterations process
The experimental results have been conducted on
fingerprint dataset FVC2004 set B to analyze enhancement
Gabor filters process the input fingerprint images % , performance. FVC2004 set B dataset contains 320 fingerprint
to produce enhanced images %* * , * , % , ,…, images: 40 fingerprints with 8 impressions each. Table I
%? ? , ? , respectively. Then, these iteration results enter the shows the overview of the dataset and its characteristics.
feature extraction and matching process. Some datasets have their characteristics in terms of
dimensions, resolution, and data origin.
Since the iteration process is based on the Gabor filter,
the filter equation is modified to allow for as many iterations TABLE I. FVC2004 SET B DATASET CHARACTERISTICS
as desired. Therefore, the modified equations are:
1 ∅B ∅B
Dataset Samples Dimensions Resolution Data Origin
/@A* @ , @ , ∅@ , @ 1 2 45 6 $27 @ ∅B ' 6
2 3-B 3+B
FVC2004 640 × 480
6 9 ∅" 7
80 500 dpi Optical sensor
∅"
DB1B (307 Kpx)
∅B " "
188on February 09,2023 at 02:37:29 UTC from IEEE Xplore. Restrictions apply.
Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded
further analysis regarding the matching performance is
necessary.
C. Fingerprint Matching Performance
The final purpose of the fingerprint enhancement is the
improvement of matching accuracy. Feature extraction and
Dry fingerprint images matching processes had performed on all enhanced images
from the proposed method. The accuracy measurement
techniques for fingerprint matching are False Acceptance
Rate (FAR), False Rejection Rate (FRR), and Equal Error
Rate (EER). EER indicates the accuracy of an algorithm.
Meanwhile, EER was obtained from the intersection point
between FAR and FRR.
IJK>LM>N LO>NP QM I
FGHI T 100% 9)
Wet fingerprint images QRR IJK>LM>N JQMOSPL
VPWXIWP LO>NP QM I
Fig. 5. Example of the wet and dry fingerprint images in FVC2004 set B
FHHI T 100% 10
B. Enhancement using Iterative Contextual Filtering QRR IJK>LM>N JQMOSPL
FVC2004
41.336 34.908 33.032 35.729 36.424
DB1B
FVC2004
27.867 30.902 32.465 34.097 35.174
DB2B
FVC2004
17.326 10.138 11.597 11.215 15.382
DB3B
FVC2004
37.357 37.118 36.319 34.861 35.138
DB4B
189on February 09,2023 at 02:37:29 UTC from IEEE Xplore. Restrictions apply.
Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded
Table III shows that the EER value of the proposed method
is better compared to the Gabor filter and the modified Gabor
filter. The average EER result from the Gabor filter is 30.972
%, the modified Gabor filter is 30.844 %, and the proposed
method is 27.893 %. In general, the proposed method
surpassed the Gabor filter by 3.08 % and the modified Gabor
filter by 2.95 %. Further analysis of the visual results between
these three methods is needed to get a better knowledge of
the enhancement results.
We compared the visual comparison between the two
enhancement methods as shown in Figure 8.
190on February 09,2023 at 02:37:29 UTC from IEEE Xplore. Restrictions apply.
Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded
to differences in the characteristics of each dataset. The images," 2017 8th International Conference on Computing,
Communication and Networking Technologies (ICCCNT), pp. 1-6,
dataset characteristics such as dimensions, resolution, and 2017.
data origin play a significant role in the experiments. Also, [4] Sherlock, B. G, D. M. Monro, and K. Millard, "Fingerprint
the proposed method shows an improvement regarding the enhancement by directional Fourier filtering." IEE Proc. Vis. Image
fingerprint matching accuracy. Experimental results show Signal Process.141.2(1994):87-94.
that the matching accuracy is improved using the proposed [5] S. Chikkerur, A. N. Cartwright, and V. Govindaraju, "Fingerprint
method. In general, the proposed method surpassed the Gabor enhancement using STFT analysis," Pattern Recognition, vol. 40, no.
1, pp. 198-211, 2007.
filter by 3.08 % and the modified Gabor filter by 2.95 %. The
[6] S. Jirachawang and V. Areekul, "Fingerprint Enhancement Based on
visual comparison also shows that the proposed method Discrete Cosine Transform," in Proc. Int. Conf. on Biometrics, 2007,
performed better restoration capabilities compared to the pp. 96-105.
Gabor filter and the modified Gabor filter. [7] S. M. Pizer et al., “Adaptive Histogram Equalization and Its
Variations,” Comput. Vision, Graph. Image Process., vol. 39, no. 3, pp.
ACKNOWLEDGMENT 355–368, 1987.
[8] K. Zuiderveld, “Contrast Limited Adaptive Histogram Equalization.
The authors would like to thank the Research Directorate Academic Press,” Graphics Germ IV, pp. 474-485, 1994.
of Gadjah Mada University for funding this publication [9] Harris, C. and M. Stephens, “A Combined Corner and Edge
through the Rekognisi Tugas Akhir (RTA) 2021 scheme. Detector.” Alvey Vision Conference, 1988.
[10] E. Rublee, W. Garage, and M. Park, “ORB : an efficient alternative to
SIFT or SURF,” 2011 International Conference on Computer Vision,
REFERENCES Barcelona, pp. 2564–2571, 2011.
[1] D. Maltoni, D. Maio, A. Jain, S. Prabhakar, Handbook of Fingerprint [11] A. Jakubović and J. Velagić, "Image Feature Matching and Object
Recognition, New York, NY, USA: Springer, 2009. Detection Using Brute-Force Matchers," 2018 International
Symposium ELMAR, Zadar, pp. 83-86, 2018.
[2] L. Hong, Y. Wan, and A. Jain, "Fingerprint Image Enhancement:
Algorithm and Performance Evaluation," IEEE Transactions on [12] BioLab. (2021, January) FVC2004 Web Site. [Online].
Pattern Analysis and Machine Intelligence, vol. 20, no. 8, pp. 777-789, http://bias.csr.unibo.it/fvc2004/
Aug. 1998.
[3] A. Garge, S. Karamchandani and Z. Saquib, "Performance evaluation
of variable and modified Gabor kernels for filtering of fingerprint
191on February 09,2023 at 02:37:29 UTC from IEEE Xplore. Restrictions apply.
Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded