Segmentation of Low Quality Fingerprint Images Diala Jomaa

Computer Science, DalarnaUniversity Planet Gatan1B 78453, Borlänge

djo@du.se

ABSTRACT

A fingerprint image is a pattern which consists of two regions,

foreground and background. The foreground contains all

important information needed in the automatic fingerprint

recognition systems. However, the background is a noisy region

that contributes to the extraction of false minutiae in the system.

To avoid the extraction of false minutiae, there are many steps

which should be followed such as preprocessing and

enhancement. One of these steps is fingerprint segmentation. The

aim for fingerprint segmentation is to separate the foreground

from the background. Due to the nature and the poor quality of

fingerprint image, the segmentation becomes an important and

challenging task.

This paper presents a new algorithm to segment fingerprint

images. The algorithm uses four features, the global mean, the

local mean, variance and coherence of the image to achieve the

fingerprint segmentation. Based on these features, a rule based

system is built to segment the image.

The proposed algorithm is implemented in three stages; pre-

processing, segmentation, and post-processing. Gaussian filter and

histogram equalization are applied in the pre-processing stage.

Segmentation is applied using the local features. Finally, fill the

gaps algorithm and a modified version of Otsu thresholding are

invoked in the post-processing stage.

In order to evaluate the performance of this method, experiments

are performed on FVC2000 DB1. Segmentation of 100 images is

performed and compared with manual examinations of human

experts. It shows that the proposed algorithm achieves a correct

segmentation of 82% of images under test.

General Terms

Algorithms, Performance, Design, Reliability, Experimentation.

Keywords

Fingerprint Recognition System, Fingerprint, Otsu, Split and

Merge, Histogram Equalization, mean, variance and coherence.

1. INTRODUCTION Fingerprint is fully created at about seven months of fetus

development. It is unique and unchangeable during individual’s

life excluding the situation of accidents in the finger such as cuts

or injuries. A captured fingerprint image usually consists of two

components; the foreground and the background. The foreground

is the area of scanner surface which is in contact with a finger

surface. It includes the necessary information needed for

fingerprint recognition, while the background is the noisy area

which is located at the borders of the image.

Fingerprint segmentation is the process by which the foreground

is separated from the image background. The result of fingerprint

segmentation is a fingerprint image in which the background is

removed [2, 9].

Fingerprint segmentation is an important step in the automatic

fingerprint recognition systems because it improves the fingerprint

images so that features can be extracted from these images by the

automatic fingerprint recognition systems. Features such as

minutiae and singular points are especially important for the

reliable identification of fingerprints. Minutiae means small

details that can determine important local features in the

fingerprint image and singular points means regions that represent

distinctive shapes. When fingerprint images include a noisy

background, feature extraction algorithms extract a lot of false

features. Hence, developing good fingerprint segmentation

algorithm helps to discard the background, and thus reduces the

number of false features.

However, the presence of dust and grease in the scanner’s sensor,

the presence of some traces from previous image acquisition or

the dryness or the wetness of the finger can cause the

segmentation to be a challenging task.

Fingerprint segmentation can be achieved by two approaches:

block-wise based and pixel-wise based [12]. In the block-wise

approach, the fingerprint image is divided into blocks and each

block is classified into foreground or background based on

features calculated for the block. While in the pixel- wise method,

segmentation is achieved on the pixel level.

Based on the hypothesis that a small fingerprint fragment

resembles a two-dimensional sinusoidal function Marques and

Thome [10] used a neural network to detect the region of interest

in fingerprint images. Bazen and Gerez [3] invoked an optimal

linear classifier trained with three pixels features; mean, variance

and coherence to achieve segmentation. Weixin [13] presented an

improved active contour for extracting salient object contours in

fingerprint images. Helfroush and Mohammadpour [7] used a

combination of three variance mean and ridge orientation features

and also employs the median filter as a post processing step.

In this paper, a new algorithm for the segmentation of fingerprint

images is presented. Segmentation is achieved by using a number

of rules related to global mean, local mean, local variance and

coherence. It includes pre-processing and post-processing stages

to enhance segmentation.

The rest of the paper is organized as follows. Automatic

fingerprint recognition system is described in section 2.

Segmentation methods and enhancement techniques are given

respectively in section 3 and 4. The features for fingerprint

segmentation are shown in section 5. In section 6 the proposed

algorithm is illustrated. The experimental results based on the

proposed method are displayed in section 7 and in section 8, the

conclusion is presented.

2. Automatic Fingerprint Recognition System Fingerprint recognition system is the oldest recognition system.

In the early twentieth century, fingerprint recognition becomes

accepted as a personal identification system in forensic.

Afterwards, different fingerprint recognition techniques, like

latent fingerprint acquisition, fingerprint classification, and

fingerprint matching were developed. At present, automatic

fingerprint recognition is in progress day after day not just in

forensic applications, but even in civilian applications.

Fingerprint recognition systems consist of the following parts

(figure 1):

- Sensing or Image acquisition

- Pre-processing

- Feature or minutiae extraction

- Matching

Figure 1. Fingerprint Recognition System

2.1.1 Sensing or Image Acquisition The acquisition of a fingerprint images was accomplished by

using off-line sensing or live-scan. Off-line sensing is defined as

ink-technique. An individual place his finger in black ink then his

finger is pressed in a paper card. However, live-scan scanners

become presently more frequent, because of its simplicity in

Permission to make digital or hard copies of all or part of this work for

personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that

copies bear this notice and the full citation on the first page. To copy

otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee.

Conference’10, Month 1–2, 2010, City, State, Country.

Copyright 2010 ACM 1-58113-000-0/00/0010…$10.00.

usage. There is no need for ink. The digital image is directly

acquired by pressing against the surface of the scanner.

2.1.2 Pre-processing To simplify the task of minutiae extraction and make it more easy

and reliable, some preprocessing techniques are applied to the raw

input image. Enhancement and segmentation of the fingerprint are

the most commonly methods performed in the preprocessing step.

The principal aim of enhancement is to improve the clarity of

ridge in the recoverable area in the image and to assign the

unrecoverable ridges as a noisy area. Recoverable region is

considered when ridges and valleys are corrupted by a small

amount of dirt, ceases, or other kind of noise. Unrecoverable

region are the regions which are impossible to recover them from

a very corrupted and noisy image [14].

However the primary purpose of segmentation is to avoid

extraction of feature in the background that is in reality considered

as a noisy area [8]. Segmentation indicates the separation of

fingerprint area or foreground from the image background. Due to

the streaked nature of the fingerprint area, a simple thresholding

technique is not sufficient. In addition to the presence of noise in a

fingerprint image, fingerprint segmentation requires more robust

and strong techniques [1].

2.1.3 Feature extraction After preprocessing step, the segmented and enhanced fingerprint

is further processed to identify the main and distinctive minutiae.

Most of the minutiae extraction methods necessitate the

fingerprint gray-scale image to be transformed into a binary

image. The acquired binary image is forwarded to a thinning stage

to reduce the thickness of the ridge to one pixel ridge. Afterwards,

the minutiae are simply detected by a simple image scan.

Due to the characteristic of the pixel that corresponds to minutiae,

the simple scan image is one of many methods developed for

minutiae detection. It depends on calculating crossing number of a

pixel. The crossing number is the half sum of the differences

between pairs of adjacent pixels in the 8-neighborhood of p. Since

the minutiae pixel can be bifurcation, crossover, termination, and

so on. Therefore, the crossing number for minutiae must be

different from 2 [9].

To avoid the problems related to fingerprint binarization and

thinning, many methods have been proposed. Direct gray-scale

minutiae extraction is one of these methods. The basic idea of this

algorithm is to track the ridge lines in the gray-scale image by

going according to the local orientation of the ridge. When a ridge

line terminates or intersects another line, the algorithm detects this

location as a minutiae point [9].

2.1.4 Matching Algorithms that extract important and efficient minutiae, will

improve the performance of the fingerprint matching techniques.

The features extracted of the input image are compared to one or

more template that was previously stored in the system database.

Therefore the system returns either a degree of similarity in case

of identification or a binary decision in case of verification.

Minutiae-based and correlation-based matching techniques are the

most common techniques in fingerprint matching. In Minutiae-

based techniques, first systems extract the minutiae in both images

then the decision is based on the correspondence of the two sets of

minutiae locations. However in correlation-based techniques

compare two fingerprints based on their gray level intensities.

First it selects relevant templates in the primary fingerprint then it

uses template matching to locate them in the secondary image and

compare positions of both fingerprints [4]. In correlation-based

techniques, the propagation of errors in minutiae extraction step is

avoided. It doesn’t require many preprocessing steps.

Nonetheless, minutiae-based technique is the most widely used

technique in fingerprint matching.

3. SEGMENTATION METHODS Generally, Image segmentation methods are classified into two

categories, discontinuity and similarity of intensity value. In the

discontinuity-based categories, segmentation can be defined as

edge-based segmentation that subdivides an image based on

abrupt changes in the intensity. In the similarity-based categories,

the segmentation is related to partitioning an image into regions

according to their similarity. The similarity is a measure that is

defined in advance depending on the fundamental problem in the

image. This measure can be a specific intensity level, mean value,

variance value, and so on. Point, line and edge detection are

examples of discontinuity methods. Also, threshold, Otsu,

splitting and merging and region growing are examples of

similarity-based methods. The combination between different

methods can give an improvement in segmentation performance.

3.1 Threshold method The main idea in threshold methods is to select a threshold T that

can separate objects from the background. This threshold can be

specified according to the intensity histogram. Histogram of an

image displays the gray-level values versus the number of pixels at

that value. Any pixel with gray level Tyxf ),( is assigned as a

foreground; otherwise the pixel is assigned as background see

formula 3.1.

255),( yxG If Tyxf ),(

0),( yxG If Tyxf ),(

(3.1)

For fingerprint images, the histogram shows the contrast of the

image and the distribution of the gray level. As shown below in

figure.2 the image is a bright image and no obvious gray level

point can be as thresholding point. Because of the nature of

fingerprint images, this algorithm cannot apply a simple

thresholding technique.

Figure 2. Histogram for bright fingerprint

3.2 Optimum Global Thresholding, Otsu’s

method The basic idea of Otsu’s method is that an optimum threshold

maximizes the separation between classes with respect to intensity

value. According to maximizing the between-class variance, this

method is optimum. Otsu’s algorithm can be presented as the

follows [12]:

- Compute the normalized histogram of the input image by

using equation (3.2)

MN

nP i

i

(3.2)

Where in

is the number of pixels with intensity i, MN is total

number of pixel

- Set through possible threshold as shown in equation (3.3) kkT )( (3.3)

Where L is the maximum intensity level in the image and is

between 0 and L-1

- Compute the cumulative sum by using equation (3.4) k

i

iPkP0

)( (3.4)

- Compute the cumulative means by using equation (3.5) k

i

iiPkm0

)()( (3.5)

- Compute the global intensity mean by applying equation

(3.6) 1

0

)(L

i

iG iPm (3.6)

- Compute the between class variance by using equation

(3.7)

))(1).((

)))(1).((( 22

kPkP

kPkPB

(3.7)

- Compute the Global variance by using equation (3.8) 1

0

2)(

L

i

iGG Pmi (3.8)

- Obtain the Otsu threshold as in equation (3.9)

)max(2

BK (3.9)

- Obtain the seperability measure by using equation (3.10)

2

2

G

B (3.10)

As explained before, the objective of this method is to find one

threshold to give the best separation between classes. Due to the

quality and nature of fingerprint, Otsu’s method cannot produce a

good separation between the foreground and the background. In

addition, the presence of dark contrast in the foreground, the

algorithm can destroy the details for ridges and valleys. Also the

formal Otsu’s Algorithm may not preserve the content of the

foreground (figure 3).

Figure 3. (a) Original image 1.4, (b) Segmented image by using

original Otsu

3.3 Split and Merge In this section, Region Splitting and Merging technique is

introduced here to segment the image by finding the regions

directly. The principal idea of this method is to iteratively split no

similar regions and merge similar regions. The approach for this

algorithm is to subdivide first the image into smaller and smaller

quadrant regions. This type of splitting is called quad trees, each

node has exactly four descends. The root for this tree corresponds

to the entire image and each node corresponds to the subdivision.

The region of interest has some clear characteristics that should be

taken as a predicate for this region segmentation. This predicate

can be considered as mean, variance or color. The splitting takes

place if the region doesn’t satisfy the predicate, otherwise these

splitting stops (figure 5). If only splitting used, the algorithm

contains only adjacent regions with same properties. Therefore,

the merging technique is needed to combine two adjacent regions

with identical properties.

Figure 5.Quad tree splitting

Split and merge algorithm is summarized by the following steps

[5]:

- Define a typical predicate to segment the image

- Split the image into four disjoint quadrants in the case of

the region doesn’t satisfy the predicate

- Merge two adjacent regions if they satisfy the predicate.

- Merge two regions in different level if they satisfy the

predicate

- Merge small region with the most similar adjacent region

- Stop merging and splitting until no regions remain

unchecked.

(a) (b)

Figure 6. (a) Original image 1.4, (b) Segmented image by using

Split and merge with variance 100.0 as a predicate

In the figure above (figure 6), it shows the result obtained by

applying Split and Merge algorithm in a fingerprint image. This

algorithm segment the image by using the variance as a predicate.

The algorithm starts with image. If the variance of the image is

greater than 100.0, the algorithm divides the image into smaller

quadrant regions. If the variance is greater than 100.0 for any

quadrant, the algorithm subdivides that quadrant into sub quadrant

and so on. After splitting the image, merging technique starts by

combining two regions if their variance is less than 100.0. As seen

in the resulting image, Split and Merge helps to nearly extract the

foreground from the background. Due to the predicate used in this

segmentation, the foreground has not effectively partitioned.

Variance is not a sufficient predicate to segment an image without

any problems. Looking for efficient feature or suitable predicate is

needed.

For that reason, the explanation of some important feature in

fingerprint segmentation is important to discuss.

4. Image Enhancement As described in the previous section, when the image is acquired,

the quality of fingerprint image can be affected by different

factors. However, fingerprint images with low contrast or false

traces ridges or noisy complex background cannot be segmented

correctly by segmentation methods. Therefore, it is required to

improve the quality of the image by applying some enhancement

techniques. Some of these techniques that were used in this thesis

are Gaussian Filter, Histogram Equalization and morphological

reconstruction.

4.1 Gaussian Filter Gaussian Noise appears in the image caused by factors such as

poor illumination and high temperature for sensor. Gaussian filter

is used to smooth the image and remove these noises. This filter is

similar to mean filter which is called averaging filter. The degree

of smoothing of Gaussian is expressed in term of which the

standard deviation of the distribution. In 2-D the Gaussian

distribution has the form of (4.1).

G(x,y) = 2

22

2

22

1yx

e

(4.1)

The goal of Gaussian Filter is to use this distribution as a point

spread function which can be performed by convolution mask.

The meaning of convolution is the process of moving the mask

from the upper-left corner to the lower-right corner and replacing

the value of the center pixel in the image by the value of g(x,y).

G(x,y) is calculated by the sum of products of the filter

coefficients and the corresponding image pixels in the area

spanned by the filter mask. Normally, the filter mask is a two

dimensional array in which the values of the mask coefficients

affect the nature of the image [5]. Therefore Gaussian filters

smoothes the image by using the following mask based on

discrete approximation to the Gaussian function (Figure 7).

273

1

Figure 7: Gaussian mask

Figure 8 shows the result of using Gaussian filter with = 5.

This filter blurs the fingerprint image by reducing some noises in

the background.

1 4 7 4 1

4 16 26 16 4

7 26 41 26 7

4 16 26 16 4

1 4 7 4 1

(a) (b)

Figure 8. (a) Original image 1.4, (b) Noise reduction with

Gaussian Filter

4.2 Histogram Equalization The usage of Gaussian Filter in fingerprint images help efficiently

to enhance the image in case of the presence of noise such that

image acquisition noises. However, this filter cannot enhance an

image which is affected by contrast problem. Therefore Histogram

Equalization is a good solution for this problem. Histogram

Equalization is the most common technique for improving the

appearance of a poor image. It is the technique to get the

histogram for the destination image as flat as possible. Histogram

Equalization defines a mapping of gray level p into gray level q

such that the distribution of gray level q is uniform. This mapping

stretches the contrast of gray level near the maxima in the

histogram [6]. The probability density function of a pixel intensity

level kr is yield by formula (4.2).

n

nrp k

kr )( (4.2)

Where kr is between 0 and 1, k=0, 1… 255, kn is the number of

pixels at intensity level kr and n is the total number of pixels.

The new intensity value ks for level k is derived by formula (4.3).

k

j

jr

k

j

j

k rpn

ns

00

)( (4.3)

By applying histogram equalization in a fingerprint image, the

contrast is increased in most of fingerprint pixel. The first image

in figure 9 shows the original image with its corresponding

histogram and the equalized image with its corresponding

histogram. As expected, the histogram of the original image (a) is

concentrated on the light side of the intensity scale. The result of

histogram equalization show important improvement in the

contrast. In the equalized histogram (d), the intensity values cover

the entire gray scale. Therefore, the significant contrast

differences between the original histogram and the equalized one,

illustrate the power of histogram equalization as a principal

contrast enhancement tool.

(a) (b)

(c) (d)

Figure 9. (a) Original image, (c) histogram of the original

image, (b) equalized image, (d) histogram of the equalized

image

5. FEATURES FOR FRINGERPRINT

SEGMENTATION Fingerprint features must reflect both the gray level of fingerprint

and the direction of ridge lines. However, the complicated

construction of fingerprint pattern and the imbalance in the

contrast, require local feature instead of the global feature. Local

Mean and variance are the features for gray-level based methods

while coherence is the feature for direction based method. The

combination for these features in one algorithm show efficiently

the distribution of the pixels for ridges and valleys in the image.

Coherence feature indicates the strength of the local window

gradients centered on the processed point along the same

dominant orientation.

Local mean, local variance and local coherence [7] are calculated

as the following formulas (5.1) and (5.2).

w

meanIw

Variance 2

2)(

1

w

Iw

mean2

1

(5.1)

Where I is the intensity and w is the window size centered on the

processed pixels.

)/()(4)( 22

yyxxxyyyxx gggggcoh

w

xxx Gg 2 ,

w

yyy Gg 2 , y

wxxy GGg

(5.2)

Where xG and

yG are corresponding horizontal and vertical

gradient components which are given by Sobel operators.

Background of a fingerprint image is the area where the finger

does not touch the sensor which means that background is the

white pixel in the image. Therefore, the local mean in the

background is higher than that of the foreground. In contrary to

mean, local variance is higher in the foreground parts compared

with background. This is applied to coherence which is high in the

foreground and low in the background. The spatial distribution of

these three features is shown in figures 10, 11 and 12.

Figure 11. Distribution of mean in image 1_4.

Figure 12. Distribution of variance in image 1_4.

Figure13. Distribution of coherence in image 1_4.

Another interesting relationship among mean, variance and

coherence is shown in figure 14. This 3D distribution of the local

mean, variance and coherence of sub-images extracted from

image 1_4 of FVC2000 DB1 shows clearly that a fingerprint

image has two distinctive spaces of foreground and background.

Figure 14. Relationship among mean, variance and coherence

of image 1_4.

6. THE PROPOSED METHOD The detailed steps of the fingerprint segmentation algorithm are

depicted in figure 15. In this paper, fingerprint segmentation is

achieved by three stages:

Figure 15. The proposed algorithm.

6.1 Pre-processing Fingerprint images with low contrast, false traces ridges or noisy

complex background cannot be segmented correctly. Therefore, it

is required to enhance the image. Techniques used in this paper

are Gaussian Filter, and Histogram Equalization. Gaussian Filter

is used to smooth the image and hence background areas. This

step together with split and merge technique, which is applied in

the next stage, will collect pixels with similar gray levels into big

areas. Histogram Equalization is invoked in this stage too. When

the global mean of the image under consideration is higher than a

certain threshold, which mean a bright image, histogram

equalization is used to enhance the image by reducing the

brightness of the image.

6.2 Segmentation In this stage, split and merge technique is applied to collect

similar background areas after the smoothing process. In order to

separate the foreground from the background, the image is divided

into a number of non-overlapping sub-images of size 10x10 pixels

and the local mean, local variance, and local coherence are

computed for each sub-image.

The global mean together with the three aforementioned

parameters are used to build a rule based system to segment the

foreground and the background of the image under consideration.

The result of this rule based system is to decide whether a certain

block is a foreground or a background. When a certain block is

decided to be a background, a set of white pixels are placed in this

block, otherwise the image pixels are kept.

The segmentation is achieved by dividing the active space shown

in figure 14, which is specified by minimum and maximum values

of local mean, variance and coherence, into a number of

subspaces. Then the global mean is invoked to build a simple

hierarchy of if then rules to decide whether a certain block

belongs to a background or a foreground region.

6.3 Post-processing The segmented fingerprint image may contain isolated

background blocks which are surrounded by foreground blocks.

Obviously, these background blocks are foreground in the original

image. A simple post-processing technique is proposed to

eliminate the presence of these isolated blocks. It is as follows, see

figure 16:

For all blocks in the image, if a background block (i,j) is found

o Check the N4 neighbor blocks which are located at (i, j-1),

(i, j+1), (i-1, j) and (i+1, j) for the presence of foreground.

o If two or more of these neighbors are foregrounds,

change all of the pixel in the block (i,j) back to their

original value before any segmentation.

After filling the foreground with the missing blocks, a modified

version of Otsu thresholding [11] method is applied. The modified

algorithm works locally in the blocks of the image to find the

corresponding optimal threshold for each block and segments the

image into white background and black foreground.

Figure 16. Filling gaps in a block image

6.4 Modified Otsu’ method Noise and non uniform illumination in the foreground and

background play a principal role in the performance of original

Otsu’s algorithm. Also, the algorithm goes through the entire

image to find an optimal threshold to further use in segmentation.

In this paper, a new version of Otsu method is developed. The

developed algorithm seeks locally in small regions of the image

and find corresponding optimal threshold for every region.

The main idea of this method is to extract the foreground from the

background by applying the global Otsu’s method and to preserve

ridges lines in the foreground by using Otsu’s in image

partitioning (figure 4).

(a) (b) (c)

Figure 4. (a) Original image 1.4, (b) Image segmented into 900

regions using Otsu thresholds in each region, (c) Segmentation

of (a) using modified Otsu

7. EXPERIMENTAL RESULTS The proposed algorithm is tested on 100 fingerprint images which

are selected randomly and without repetition from FVC2000

database DB1. These images are collected by using two small-size

and low-cost optical sensors.

To evaluate the efficiency of this algorithm, human expert

examines the results of the segmentation algorithm from these

random images. Figure 17 shows a number of successful

segmentation achieved by this algorithm.

Figure 17. Correctly segmented images from FVC2000

fingerprint database.

In order to evaluate this method quantitatively, a four-level

segmentation scheme is suggested. The number of correctly

segmented blocks in the image is measure on which this scheme is

based. This scheme is defined as follows:

Good, when more than 90% of the background blocks are

segmented correctly.

Almost Good, when 75% - 89% of the background blocks are

segmented correctly.

Almost Bad, when 60% - 74% of the background blocks are

segmented correctly.

Bad, when less than 60% of the background blocks are

segmented correctly.

The results of classification into four categories are shown in

figure 18.

Good

Almost good

Almost bad

Bad

Figure 18. Four samples from FVC2000 which show the good,

almost good, almost bad and bad categories.

This scheme is applied on the tested images extracted from

FVC2000 DB-1 and the results of segmentation are depicted

Table 1. According to this scheme, Good and Almost Good results

can give 82% of the images under test.

Table 1. Results of different categories

Result Percentage

Good 66%

Almost Good 16%

Almost Bad 4%

Bad 14%

8. CONCLUSIONS This paper presents a new algorithm for the segmentation of

fingerprint images. The proposed algorithm is implemented in

three stages; pre-processing, segmentation, and post-processing.

Segmentation is achieved by using a number of rules related to

global mean, local mean, local variance and coherence.

Experiments of testing this algorithm show that it is able to

segment 82% of images used for testing. Further analysis to these

results can reveal to a set of modifications of this algorithm to be

implemented in the future. These modifications can be either the

use of Support Vector Machine (SVM) classifier or Fuzzification

techniques to replace the rule based system, which can lead to

better segmentation results. The scheme was applied on the tested

images extracted from FVC2000 DB-1. More intensive tests can

be done in other FVC2000 databases.

REFERENCES

[1] Afsar, F.A., Arif, M. and Hussain, M. 2004. Fingerprint

Identification and Verification System using Minutiae

Matching. In proceedings of the National Conference on

Emerging Technologies. 141-146.

[2] Akram, M., Nasir, S., Tariq, A., Zafar, I. and Khan, W. S.

2008. Improved Fingerprint Image Segmentation Using New

Modified Gradient Based Technique. In proceedings of the

2008 Canadian Conference on Electrical and Computer

Engineering. Niagara Falls, Canada, 001967 - 001972

[3] Bazen, A. M. and Gerez, S. H. 2001. Segmentation of

Fingerprint Images. In proceedings of the ProRISC 2001

Workshop on Circuits, Systems and Signal Processing.

Veldhoven, the Netherlands, 276 - 280.

[4] Bazen, A. M., Verwaaijen, G. T. B., Gerez, S. H., Veelenturf,

L. P. J. and Zwaag, B. J. 2000. A Correlation-Based

Fingerprint Verification System. In proceedings of the

ProRISC Workshop on Circuits, Systems and Signal

Processing. Veldhoven, the Netherlands.

[5] Gonzalez, R. C. and Woods, Richard E. 2008. Digital Image

Processing.

[6] Greenberg S. , Aladjem, M., Kogan, D. and Dimitrov, I.

2000. Fingerprint Image Enhancement using Filtering

Techniques. In proceedings of the 15th International

Conference on Pattern Recognition. Barcelona, Spain, 227--

236.

[7] Helfroush, M. and Mohammadpour, M. 2008. Fingerprint

segmentation. In proceedings of the 3rd International

Conference on Information and Communication

Technologies: From Theory to Applications. Damascus,

Syria.

[8] Hong, L., Jain, A., Pankanti, P. and Bolle, R. 1996.

Fingerprint Enhancement. In Proceedings of the IEEE

WACV. Sarasota, Florida.

[9] Maltoni, D., Maio, D., Jain, A. and Prabhakar, S. 2003

Handbook of fingerprint recognition: Springer.

[10] Marques, P. B. and Thome, G. 2005. A Neural Network

Fingerprint Segmentation Method. In proceedings of the

IEEE Fifth International Conference on Hybrid Intelligent

Systems. Rio de Janeiro, Brazil.

[11] Otsu, N. 1979. A threshold selection method from gray level

histogram. In Proceedings of the IEEE Trans. Syst.Man

Cybern, vol. SMC-9, pp. 62 - 66.

[12] Ren, C.,Yin, Y., J. Ma, and Yang, G. 2008. A Linear Hybrid

Classifier for Fingerprint Segmentation. In proceedings of

the IEEE Fourth International Conference on Natural

Computation. Jinan, China, 33 - 37.

[13] Weixinand, B., Deqin, X. and Yi-wei, Z. 2009. Fingerprint

Segmentation Based on Improved Active Contour. In

proceedings of the IEEE Computer Society International

Conference on Networking and Digital Society. Guiyang,

Guizhou, 44 - 47.

[14] Wu, C. 2007. Advanced Feature Extraction Algorithms for

Automatic Fingerprint Recognition Systems. Doctoral Thesis.

UMI: ProQuest Digital Dissertations, State University of

New York at Buffalo.