Ipsha Panda et al.

/ International Journal on Computer Science and Engineering (IJCSE)

A New Approach To Fingerprint Recognition

Ipsha Panda
IIIT Bhubaneswar, India
[email protected]

Saumya Ranjan Giri

IL&FS Technologies Ltd.
Bhubaneswar, India
[email protected]

Prakash Kumar
IL&FS Technologies Ltd.
Bhubaneswar, India
[email protected]

Anjali Mohaptra
IIIT Bhubaneswar, India
[email protected]

Abstract—Fingerprint recognition continues to be the most widely used biometric system for security
purposes. Still it is a complex pattern recognition problem. In this paper, a new approach for fingerprint
recognition is developed. There are three steps: image preprocessing, minutiae extraction and minutiae
matching. In the first stage the fingerprint image is preprocessed to produce a thinned image with noisy
data deleted. Then for minutiae extraction, some heuristic rules have been developed based on the
concept of neighboring pixels. This stage results in removal of spurious minutiae points and true minutiae
points are detected. Finally an algorithm for minutiae matching has been designed to improve the
performance and have accurate results.

Keywords- Otsu thresholding; Central line thinning; radial distance; Minutiae

I. INTRODUCTION
Fingerprint recognition is one of the first applications of machine pattern recognition. It is still a challenging
problem due to the following reasons: smaller overlap between template and the input image due to rotation,
sweat or injury on finger cause incorrect minutiae detection, non linear distortion caused by finger plasticity. A
fingerprint consists of ridges and valleys. In general a ridge can either come to an end which is termed as ridge
end or it can split into two ridges which is termed as bifurcation point [4]. There are three stages in fingerprint
recognition. The first step i.e. image preprocessing is done to produce a noise free. second step involves
application of the heuristic rules developed to extract true minutiae points and finally the proposed algorithm is
applied for minutiae matching The paper is organized as follows: section II presents preprocessing stage, section
III presents proposed heuristics for minutiae extraction, section IV presents matching algorithm and section V
concludes and marks the future aspects.

II. IMAGE PREPROCESSI NG STAGE

The fingerprint images do need some preprocessing in order to accurately detect the minutiae points for
recognition. This step is done to reduce the noise and enhance the definition of ridges against valleys. The main
steps involved are shown in the figure below:

ISSN : 0975-3397 Vol. 4 No. 05 May 2012 909

 Ipsha Panda et al. / International Journal on Computer Science and Engineering (IJCSE)

Threshold Detection

Binarization

Dilation

Central Line Thinning

Removing unwanted
portions

Figure 1. Image Preprocessing Step

A. Threshold detection using OTSU method [1]

The key idea of recursive Otsu thresholding is to take the set of background pixels as determined by standard
Otsu thresholding and re-threshold these pixels. If let Otsu (a, b) denote the Otsu-selected threshold for all pixels
in the image between values a and b, the recursive Otsu thresholding can be written as
T1= Otsu (0, 255)
Tk = Otsu (Tk-1, 255)
The Recursive Otsu thresholding algorithm where T1 is the threshold determined from recursion k. The
following three stopping criteria for the algorithm:
1. Stop if the number of pixels added by the current recursive step exceeds the number of pixels selected by the
first threshold T1, which indicates that large portions of the true background are now being included.
2. Stop if the change in threshold between successive recursive steps is smaller than some small value d1, which
indicates that the algorithm is straining at very subtle detail (Tk-Tk-1 < d1),
3. Stop if the change in threshold between successive recursive steps is larger than some larger value d2, which
indicates that the algorithm is simply subdividing the actual background pixels. (Tk-Tk-1> d2).
Once any one of the stopping criteria is met, the recursion is stopped and uses the result from the preceding step.
B. Binarization of an image [5]
It is a process of converting grayscale image to black and white image i.e. binary image. The threshold value
that results from above step is used for this conversion. All the pixel values in the image is compared with the
threshold, if the value is above threshold then the pixel value is changed to one else the value is changed to zero.
C. Dilation of an image [5]
It is a morphological process which is used to reduce the width of the ridges. It thickens object in a binary
image. In this step, dilation is used to thicken the valleys in a binary fingerprint image. This results in erosion of
the ridges. A conservative structuring element consisting of four ones arranged in a two-by-two square is used
for the valley dilation to achieve some ridge width reduction.
D. Central line thinning of binary image [2]
This iterative step reduces the width of the ridges to one pixel at their central line. This method deletes pixels at
the outer boundaries of the ridges, so long as the width is greater than one pixel. First the image is scanned for
black pixels. For each pixel, a check is done to determine out of the twenty one rules which scenario is matched
and following changes are done. If none of the rules matched then the middle pixel remains unchanged. This
process is done continuously till no difference results in two consecutive iterations. Then the diagonal rules are
implemented to produce the final thinned image.

ISSN : 0975-3397 Vol. 4 No. 05 May 2012 910

 Ipsha Panda et al. / International Journal on Computer Science and Engineering (IJCSE)

E. Removal of unwanted portions from thinned image(using SAGAP) [3]

Short islands at the outer boundaries of the fingerprint image do not represent true ridge structure so these
unwanted portions need to be removed. This step begins with detecting terminations in the thinned image. At
each termination the ridge is traced one pixel at a time. The ridge is considered as island if it terminates before a
maximum trace of continuous twenty pixels and it is deleted otherwise no changes are done.

Now the original image is converted to a negative image by converting the black pixels to white pixels and vice
versa. These steps are repeated for the dual image constructed to get the thinned dual image After these steps the
resulting image shown in Fig. 2 and Fig. 3 is given as input to the next stage i.e. minutiae extraction.

Figure 2. Original Image thinned Figure 3. Dual Image Thinned

III. PROPOSED HEURISTICS FOR MINUTIAE EXTRACTION

After the preprocessing of fingerprint image, minutiae points are extracted from the both the thinned image
resulted. Some heuristic rules have been developed which are explained in details below:
A. Detection of Ridge end and Bifurcations on the basis of neighborhood pixels:
Each of the two dual images is scanned and for each pixel count the number of neighbors the pixel has. If the
count value of a pixel is one, then mark this pixel as ridge end (point ‘A’) [8], as shown in figure below. If the
count value corresponds to three, then the pixel is marked as bifurcation (point ‘B’) [8] provided its neighboring
pixel does not have count equal to three.

B A

Figure 4. Detection of ridge end and bifurcation point

If its neighboring pixel have count equal to three, as shown in Fig. 5, both point ‘A’ and point ‘B’ have
count equals three then point ‘A’ is considered as the bifurcation point. Point ‘A’ has no two neighboring pixels
adjacent with each other (there is at least one position gap among the neighboring pixels). So point ‘B’ is not
considered. The value of the pixel is changed from zero to 100 if it is a ridge end else it is changed to 200 if it is
a bifurcation. After this, the ridge ends at the outer boundaries are discarded for better results.

ISSN : 0975-3397 Vol. 4 No. 05 May 2012 911

 Ipsha Panda et al. / International Journal on Computer Science and Engineering (IJCSE)

A B

Figure 5. Bifurcation identification

B. Filtering of Detected Ridge ends

As a ridge end is detected, trace in backward direction for at least 20-25 pixels [8]. Suppose point ‘A’ is a
ridge end as shown in figure below. If another ridge end let point ‘B’, is detected before the trace ends then both
the points A and B are not considered as ridge end and the pixel value is changed from 100 to zero again.

Figure 6. Removal of false ridge end

If another bifurcation point let point ‘C’, is detected before the trace ends then the ridge end ‘A’ is discarded but
not the bifurcation point ‘C’ as it may be a spur.

Figure 7. Removal of false ridge end when it is connected to bifurcation point

C. Filtering of the Detected Bifurcations

As a bifurcation point is detected suppose point ‘A’, trace in all the three directions i.e. its neighboring pixels
are visited. While visiting in these directions if the branch ends before the trace is over or the trace length is less
than 20-25 pixels then it is a spur and point ‘A’ is not a bifurcation point. As shown in figure below while
tracing along the neighboring pixels point ‘B’ is the end point of the branch and whose trace length is less than
20-25 pixels. This shows that this branch is a spur. It is a false bifurcation point and the pixel value is changed
to zero from 200.

ISSN : 0975-3397 Vol. 4 No. 05 May 2012 912

 Ipsha Panda et al. / International Journal on Computer Science and Engineering (IJCSE)

Figure 8. Removal of spurious bifurcation points

D. Removing Bridges
When two parallel ridges are joined bridges are formed as shown in figure.

Figure 9. Bridges in a fingerprint image

The Fig. 9 shows that both point ‘A’ and point ‘B’ may be considered as bifurcation. Following the filtering
process explained in above step it may so happen that when we trace in all the three directions, each branch
trace may be less than 20-25 pixels then these points are not true bifurcation points. But if the trace length is
greater than 20-25 pixels then these points can be considered as bifurcation points. But in actual these form the
bridge, so in order to eliminate these points, need to differentiate between a true bifurcation and a false
bifurcation. A true bifurcation point has one acute angle and two obtuse angles while a bridge has two acute
angles and one obtuse angle. This heuristic is used to detect a bridge and then the point is discarded if it forms
two acute angles and one obtuse angle.

Acute angle

Obtuse angle

Acute angles Obtuse angles

Figure 10. Angles in Bridge Figure 11. Angles in true bifurcation

ISSN : 0975-3397 Vol. 4 No. 05 May 2012 913

 Ipsha Panda et al. / International Journal on Computer Science and Engineering (IJCSE)

E. Comparing with the dual image

The above steps are applied to both the images i.e. the original thinned image and the thinned dual image and
find the minutiae points in each image. Then scan the original image sequentially. When a bifurcation point is
detected in the original image, the particular coordinate points are searched in the dual image and a 10 X 10
window is constructed around the coordinates. If a ridge end is detected within the window then the bifurcation
point in the original image is a true minutiae, else discard the bifurcation point. The same process is repeated for
ridge ends in the original image i.e. bifurcation points is searched within the 10 X 10 window in dual image.
Finally all the true minutiae points are extracted and stored as a template.

IV. ALGORITHM FOR MINUTIAE MATCHING

A. Method

After minutiae extraction suppose the total number of true minutiae points in the template is 5 and the total
number of minutiae in the input image is 8 as shown in tables below. Then we calculate the radial distance
between two minutiae points taking one of them as reference point.
Suppose point ‘A’ is the reference point in template and point ‘P’ is the reference point for input. Then the radial
distance between point ‘A’ and point ‘B’ is 112 using the formula [8]:

√ (rowm – rowref )2 + (colm – colref )2 (1)

TABLE I. FOR TEMPLATE

Name Co ordinate Radial Minutiae
Distance Type

A 79,102 0 Bifurcation

B 75,214 112 Ridge end

C 64,110 17 Bifurcation

D 157,182 111 Ridge end

E 50255 156 Bifurcation

TABLE II. FOR INPUT IMAGE

Name Co coordinate Radial Minutiae
Distance Type
P 75,103 0 Bifurcation
Q 87,90 18 Bifurcation
R 172,146 106 Ridge end
S 108,206 108 Ridge end
T 101,254 153 Bifurcation
U 224,228 194 Ridge end
V 187,52 123 Ridge end
W 231,133 159 Ridge end

Next we pair the minutiae points based on the type of minutiae i.e. bifurcations should be paired with
bifurcations of other image and the same is for ridge ends. Suppose the reference point ‘A’ of template image is
paired with the reference point ‘P’ of input image.
Then we select two minutiae points and determine angles made between these two points with the reference
point as the center. Suppose we choose point ‘B’ and point ‘C’ and determine the angle between these two
points with reference point ‘A’ at the center. Similarly this is repeated for all combinations and is stored.

ISSN : 0975-3397 Vol. 4 No. 05 May 2012 914

 Ipsha Panda et al. / International Journal on Computer Science and Engineering (IJCSE)

TABLE III. ANGLES MADE BETWEEN TWO MINUTIAE POINTS FOR TEMPLATE
A B C D E

B - 60º 46º 9º

C 60º - 106º 51º

D 46º 106º - 55º

E 9º 51º - -

TABLE IV. ANGLES MADE BETWEEN TWO MINUTIAE POINTS FOR INPUT IMAGE
P Q R S T U V W
Q -
R 71º -
S 120º 48º -
T 128º 51º 8º -
U 87º 16º 32º 40º -
V 23º 48º 97º 105º 64º -
W 58º 13º 61º 69º 29º 35º -

Figure 12. Arrangement of minutiae points on the image

Next we calculate the difference between the radial distances between the paired minutiae points.

TABLE V. RADIAL DISTANCE COMPARISON

Minutiae Radial Minutiae Radial Difference

Points of Distance Points of Distance in radial
Template from Input from distance
Reference image Reference
Point point
A 0 P 0 0
B 112 S 108 4
C 17 Q 18 1
D 111 R 106 5
E 156 T 153 3

Suppose the threshold value for maximum difference between radial distances allowed is 5, then all the above
points are matched. So the resulted paired minutiae points based on radial distance comparison are:
A – P, B – S, C – Q, D – R , E – T

Now let us compare the angles made by each point with other points.

TABLE VI. ANGLE COMPARISON

A/P B/S C/Q D/R E/T
B/S -
C/Q 60º/120 º -
D/R 46º/48 º 106º/70º -
E/T 9º/8 º 51º/128º 55º/56º -

ISSN : 0975-3397 Vol. 4 No. 05 May 2012 915

 Ipsha Panda et al. / International Journal on Computer Science and Engineering (IJCSE)

Suppose the threshold value for angle difference is 5º. Consider the pairs A-P, B-S and C-Q. The angle made
between point ‘A’, ‘B’ and ‘C’ is 60º and the angle between point ‘P’, ‘Q’ and ‘R’ is 120º. As these are the
paired minutia so the angle difference must not be greater than threshold as shown in figure-11. So this results in
false pairing of s B-S and C-Q.
Similarly the angle difference for each pair is calculated.
Angle DAB and angle RPS, difference is 2º.
Angle EAB and angle TPS, difference is 1º.
Angle EAD and angle TPR, difference is 1º.
Finally the correct pairs are: A – P, D – R, B – S, and E – T.

Figure 13. Arrangement of minutiae points on the template image and the input image respectively

The above matched minutiae points can be diagrammatically viewed as follows.

Figure 14. Arrangement of all minutiae points on the template image

Figure 15. Arrangement of all minutiae points on the input image

These four minutiae points form similar structure in both input and template images. These steps are repeated
for each pair of reference minutiae points. The highest number of matching point is the result.
B. Algorithm
1. Repeat step-2 to step-5 until all minutiae points are consumed.
2. Take one minutiae point from template and one from input image as reference and repeat step-3 to
step-4 until all minutiae points are consumed.
3. Calculate radial distances for all points from the reference point for both the template and the input
image.
4. Pair one minutiae point from template with one from input image if the difference in radial distance if
the points ≤ certain threshold value and their type are same.
a. If the above condition is true, then find the angle between two minutiae of an image made with
the reference point. Find the same angle in the other image.

ISSN : 0975-3397 Vol. 4 No. 05 May 2012 916

 Ipsha Panda et al. / International Journal on Computer Science and Engineering (IJCSE)

b. If difference between the angles ≤ certain threshold value, then pair minutiae points from
template and input image.
5. Count the number of pairs (no repeating pairs) and this represents total number of matching minutiae
points for a particular pair of reference points. Store the count value.
6. Finally select the maximum count value. This is the final matching minutiae points.

V. EXPERIMENTAL RESULTS

We have developed a java program according to the explained algorithms and seen the output is exactly what
we explained in the theory. Two fingerprint images were taken one as a template image and the other as input
image. Then minutiae points were detected and finally two samples were compared with the detected minutiae
points.
We found all the matched points form similar structure in both samples. Lines were drawn by joining all
matched points among themselves in the java program.

Figure 16. Arrangement of all matched minutiae points on the input image. Total 8 points are matched out of 15.

Figure 17. Arrangement of all matched minutiae points on the template image. Total 8 points are matched out of 12.

It can be observed from the above two figures that two samples are matched only when all the points form
similar structure. The image in Fig. 17 has different orientation than the image in Fig. 16. This is because of the
punching angle of the finger on the biometric device. The vertices in the above figures represent minutia points.

VI. CONCLUSION
We have summarized all the steps involved in a fingerprint matching system. Some of the algorithms are
referred from some research papers and some are newly proposed by us. We have tested the algorithm and
found it efficient to a great extent. The algorithm is tad slower than some other algorithm but gives better result.
In future the time complexity of the algorithm can be reduced. In future more techniques may be identified to
detect correct minutiae.

ISSN : 0975-3397 Vol. 4 No. 05 May 2012 917

 Ipsha Panda et al. / International Journal on Computer Science and Engineering (IJCSE)

REFERENCES

[1] Oliver Nina, Bryan Morse, and William Barrett “A Recursive Otsu Thresholding Method for Scanned Document Binarization” IEEE,
2010
[2] M. Ahmed,R. Ward, “A rotation invariant rule based thinning algorithm for character recognition” IEEE Transactions on Pattern
Analysis and Machine Intelligence, 24, 1672-1678.
[3] P. Kumar, S. R. Giri, G. R. Hegde and K. Verma, “A Novel Algorithm to Extract Connected Components in a Binary Image of
Vehicle License Plates”, IJECCT 2012, vol. 2 (2), pp. 27-32.
[4] D. Maltoni, D. Maio, A.K.. Jain, S. Prabhakar, Handbook of fingerprint recognition. New York: Springer, 2003.
[5] R.C. Gonzalez and R.E. Woods, Digital Image Processing, 3rd Edition, Prentice Hall, 2002.
[6] Xuefeng Liang and Tetsuo Asano, “Fingerprint Matching Using Minutia Polygons”, 18th International Conference on Pattern
Recognition ,vol. 01.
[7] F.A. Afsar, M. Arif and M. Hussain, “Fingerprint Identification and Verification System using Minutiae matching”, National
Conference on Emerging Technologies, 2004.
[8] T. Graig, Diefenderfer, Thesis on “Fingerprint Recognition” at Naval Postgraduate School, Monterey, California, june 2006.
[9] Ravi J, K. B. Raja, Venugopal K. R., “Fingerprint Recognition using Minutiae Score Matching”, International Journal of Engineering
Science and Technology vol.1,pp. 35-42,2009.
[10] Yuliang He, Jie Tian, Xiping Luo, Tanghui Zhang. “Image enhancement and minutiae matching in fingerprint verification.” Elsevier,
Pattern Recognition Letters ,vol. 24, pp. 1349–1360, 2003.

ISSN : 0975-3397 Vol. 4 No. 05 May 2012 918