Vous êtes sur la page 1sur 6

1134 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 8, NO.

8, AUGUST 1999

subspaces: noise and signal. The signal subspace contains vectors


of signal and noise components, and noise subspace contains the
noise process only. The enhancement algorithm using block-adaptive
Wiener Filtering was then formulated to estimate the clean signal
from the signal subspace. The estimation criterion developed here
simulates the human visual system. Our experimental results show
that the signal subspace approach can be successfully applied to
enhance images corrupted by uncorrelated additive noise.
Fig. 1. Automatic fingerprint identification system.
REFERENCES

[1] R. C. Gonzalez and P. Wintz, Digital Image Processing, 2nd ed.


Reading, MA: Addison-Wesley, 1987.
[2] T. S. Huang, G. T. Yang, and G. Y. Tang, “A fast two-dimensional
median filtering algorithm,” IEEE Trans. Acoust., Speech, Signal Pro-
cessing, vol. ASSP-27, pp. 13–18, Feb. 1979.
[3] H. L. Van Trees, Detection, Estimation and Modulation Theory, (Part I).
New York: Wiley, 1968.
[4] Y. Ephraim and H. L. Van Trees, “A signal subspace approach for
speech enhancement,” IEEE Trans. Speech Audio Processing, vol. 3,
pp. 251–266, July 1995.
[5] J. O. Limb, “Distortion criteria of the human viewer,” IEEE Trans. Syst.,
Man, Cybern., vol. SMC-9, pp. 778–793, Dec. 1979.
[6] B. Ramamurthi and A. Gersho, “Classified vector quantization of
images,” IEEE Trans. Commun., vol. COM-34, pp. 1113–1115, Nov.
1986.
[7] C. Bouman and K. Sauer, “A generalized Gaussian image model for
edge-preserving MAP estimation,” IEEE Trans. Image Processing, vol.
2, pp. 296–310, July 1993.
[8] W. H. Pun and B. D. Jeffs, “Adaptive image restoration using a
generalized Gaussian model for unknown noise,” IEEE Trans. Image
Processing, vol. 4, pp. 1451–1456, Oct. 1995.

Fig. 2. Portion of a typical fingerprint.

A Genetic Algorithm for the There is another popular method called the Galton method. It
Estimation of Ridges in Fingerprints is based on other features such as ridge endings, bifurcation, etc.
These features are called minutiae. This method is most suitable for
Ahmed S. Abutaleb and M. Kamel computer manipulations [3]. There are, in each fingerprint, around
100 minutiae; only 13–15 minutiae are needed to identify the person.
Minutiae detection is obtained after a lengthy process that involves
Abstract—A genetic algorithm is developed to find the ridges in paper image enhancement, edge or ridge detection, binarization [1], thinning
fingerprints. It is based on the fact that the ridges of the fingerprints are
parallel. When scanning the fingerprint, line by line, the ideal noise-free
[10], noise reduction, and finally minutiae detection.
gray level distribution should yield lines of black and white. The widths The following are several problems involved with minutiae detec-
of these lines are not constant. The proposed genetic algorithm generates tion:
black and white lines of different widths. The widths change until we get 1) poor image quality of the fingerprint, especially when obtained
the best match with the original fingerprint.
on paper;
Index Terms— Edge detection, fingerprint, genetic algorithm, image 2) over-inking or under-inking in many areas of the image;
processing. 3) old wounds or cuts in the hands appear as lines crossing the
fingerprint;
I. INTRODUCTION 4) sweat affects the continuity of the ridges and cause them to
appear as discontinuous curves.
Fingerprints have been used for personal identification since the
end of the 19th century. At the beginning, there was only one method Most of the reported work, however, starts with high quality
for fingerprint classification, the Henry method. It is based on the images and then proceeds to find the features, minutiae [2], [8]. The
shapes found in the fingerprint such as arches, whorls, loops, etc. classification could be achieved through syntactic pattern recognition
[14]. It is most suitable for manual manipulations. or other approaches such as neural networks [12]. Little work has
been done to enhance the image and improve the quality of the
Manuscript received June 20, 1995; revised May 28, 1997. The associate paper fingerprint. One could mention, however, the diffusion theory
editor coordinating the review of this manuscript and approving it for approach [13] or the phase method [11].
publication was Prof. Patrick A. Kelly.
The authors are with the Systems and Bioengineering Department, Cairo The major problem with the existing edge detection operators is
University School of Engineering, Giza, Egypt (e-mail: ataleb@idsc.gov.eg). that they do not take into consideration the special shape of the
Publisher Item Identifier S 1057-7149(99)06115-1. fingerprints. Fingerprints are made of almost parallel lines (the ridges)

1057–7149/99$10.00  1999 IEEE


IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 8, NO. 8, AUGUST 1999 1135

Fig. 3. Measured and estimated gray level values for a scanned line in the fingerprint.

which appear as dark area on the paper. There is another white area A portion of a typical fingerprint is shown in Fig. 2 where many of
between every two ridges. Thus, when scanning row-wise, we have, the mentioned problems are exemplified. If the ridges were accurately
ideally, black lines followed by white lines of different widths. These obtained, the task of minutiae detection starts. It is these minutiae,
lines do not follow any regular pattern except that they are dark their number and location, that differentiate between the different
followed by light lines. persons.
The estimation of the line widths is a nonlinear estimation problem This work is concerned with preprocessing the image to delineate
that is not easy to solve. In this correspondence, we propose to the ridges. When scanning the image line by line, we should, ideally,
use an adaptive genetic algorithm to solve the nonlinear estimation obtain black line followed by white line followed again by black
problem, and thus obtain the ridges in the fingerprint. This approach is line and so on. Due to different forms of noise, we obtain a set of
guaranteed to converge to the global minimum. This approach could gray levels for the pixels in the scanned line. Thus, a good algorithm
also be implemented on hardware since it is, basically, a parallel should be able to match a black and white set of lines to each scanned
algorithm. line in the fingerprint. Notice also that a black line should follow a
Genetic algorithms have been around since the 1960’s [4]. Recently white line, and the width of each is a finite number of pixels. This
there has been great interest in their use [9]. Genetic algorithms have is shown in Fig. 3.
the following advantages [5]: Thus, if the gray level of the scanned line at the ith location is
1) simple to use; defined as y (i), and if the assumed or estimated ideal black and white
2) gradient is not calculated; lines are defined as, then the correlation coefficient, r, is calculated
3) they have memory; through
4) a priori information about the unknown parameters is easily I 1 0
utilized as constraints; [y (i) 0 yavg ][ y
^(i) 0 ^avg ]
y

5) global minimum is achieved in finite number of steps. i=0


r = (2.1)
The basic idea behind such an approach is survival of the fittest. 0
I 1 0
I 1
Starting initially with a set of random solutions, one proceeds by [y (i) 0 yavg ]
2 [^
y (i) 0 ^avg ]2
y

selective mating among them. This results in new sets of solutions i=0 i=0
or offsprings. The process continues until convergence is obtained.
where
This correspondence presents an adaptive genetic algorithm to find
the ridges in the fingerprint. It is adaptive because the probabilities 1
I 1 0 1
0
I 1
yavg = y (i) and ^avg
y = ^(i):
y (2.2)
of crossover and mutation, parameters of the genetic algorithm, are I I
i=0 i=0
estimated on line. It is shown that the proposed approach performs
well and is able to delineate the ridges, even at low SNR values (poor The correlation coefficient r is the optimization criterion. The
quality images). It is also compared to two used methods of high optimization algorithm changes the widths of the black and white
accuracy, Sobel operator and the phase method [11]. The adaptive lines to obtain a maximum value of r. Notice that r has a maximum
genetic algorithm outperforms both methods especially for real poor of 1 and a minimum of 01.
quality images.
In Section II, the ridge estimation problem is formulated. In III. PROBLEM SOLUTION USING THE
Section III, problem solution using the proposed genetic algorithm PROPOSED ADAPTIVE GENETIC ALGORITHM
is introduced. In Section IV, we present and compare some real
The evolutionary computation approach offers a variety of ways to
fingerprint results obtained through the above mentioned methods.
solve the different parameter optimization problems. Each problem
has its own unique tuning parameters. In the case of the ridge
II. PROBLEM FORMULATION identification, the tuning parameters are the probabilities of cross
An automatic fingerprint identification system block diagram is over and mutation.
shown in Fig. 1. The fingerprint is obtained in a digitized format by The solution steps for the estimation of ridges are as follows.
scanning a paper that contains the fingerprint. The paper fingerprint 1) Each scanned line in the image is made of black and white
has several problems: 1) the gray level is not homogenous; 2) over- segments. The width of each segment, in pixels, and its value,
or under-inking in different sections; and 3) writing in the fingerprint black or white ridge or valley, are two unknown parameters and
paper obscures the original fingerprint, etc. These problems cause represent a gene. The genes are not mixed together, i.e., each
the ridges to be not clear enough and preprocessing of the image is gene is also a chromosome. Each parameter is, independently,
needed. treated in terms of mating, cross over, and mutations. Each
1136 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 8, NO. 8, AUGUST 1999

(a) (b)

(c) (d)
Fig. 4. Comparison between the estimates of the ridges: (a) original 642 64 frame; (b) magnitude of the Sobel operator; (c) ridges obtained by the
phase method; and (d) ridges obtained by the proposed adaptive genetic algorithm.

parameter is initialized at random with a population of size S . 3) The probability of reproduction for the ith solution, pri , is
Usually the population size is around 20 times the number calculated according to the formula
of unknown parameters. Thus, we have S vector solutions
for each scanned line in the image. Each parameter is coded
ff
through the binary code, and each parameter has minimum pri = S 01 i : (3.1)
and maximum values. For example, it was observed that the ffs
ridge/valley (segment) width lies between two and six pixels, s=0
if the ridges are vertical 6=4. If the code is 8 b wide, then
00 000 000 corresponds to two, and 11 111 111 corresponds to
If there is a great discrepancy between the maximum and
six. Notice that the black segment in the solution vector should
minimum values, other forms of the probability of reproduction
be followed and preceded by a white segment.
should be used.
2) Fitness function calculations are performed for the S vector
4) At each iteration, two solutions, for each parameter, are picked
solutions. For each solution, we calculate the correlation co-
at random according to their probabilities which are defined
efficient of (II.1) which represents the fitness function ffi . by (3.1). These two solutions represent the two parents. This
The cross over and mutation probabilities are both changing process is repeated S /2 times. Thus at the end we get a set of
each iteration. They are considered as unknown parameters, S parents.
and thus, are members of the solution vector. Initially, like any 5) The two parents will mate through the procedure of cross over.
other parameter, they have random values that are used in the The mating will occur with probability pc . The location of the
mating process. In each iteration, the solution that yields the crossover is chosen according to the uniform distribution. If the
maximum correlation value (maximum fitness function) has as size or code for each parent is 8 b, then the crossover position
its members the values of the probability of cross over and the would be between the second and the seventh position. The
probability of mutation. These values are the ones used in the result of the mating will be two offsprings. The two offsprings
mutation to produce new offsprings. represent two new solutions and they replace the two parents.
IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 8, NO. 8, AUGUST 1999 1137

(a) (b)

(c) (d)

Fig. 5. Comparison between the estimates of the ridges for another frame: (a) original 64 2 64 frame, (b) magnitude of the Sobel operator, (c) ridges
obtained by the phase method, and (d) ridges obtained by the proposed adaptive genetic algorithm.

6) Mutation will be applied with probability pm . In this operation, The horizontal, W1, and the vertical masks, W2, of the Sobel
every single bit in every solution is subject to change with operator are defined as
probability pm .
7) The new S vector solutions are used with (2.1) to find their cor- +1 0 01 and 01 02 01
responding fitness functions and the whole process is repeated +2 0 02 0 0 0
again. +1 0 01 +1 +2 +1:
8) Stop the iterations when the number of iterations exceeds a
limit or when the correlation coefficient does not change. The If the original image is defined as I (i; j ), then the output of the
obtained best solution, with maximum fitness (correlation), is horizontal and vertical masks are defined by
the desired solution.
k =+1 k =+1
y1 (i; j ) = w1 (k1 ; k2 )I (i + k1 ; j + k2 ) (4.1)
k =01 k =01
IV. RESULTS AND CONCLUSIONS and
k =+1 k =+1
The fingerprint is scanned using an HP scanner with resolution of
y2 (i; j ) = w2 (k1 ; k2 )I (i + k1 ; j + k2 ):
300 dpi. A typical fingerprint image generates 320 2 320 matrix of
(4.2)
=01 k =01
256 gray levels, see Fig. 2. Each image is divided into several 64 2 64
k

frames. This has the effect of reducing computational time and


The edge strength image, y (i; j ), is obtained as
easing memory handling. Each 64 2 64 frame is studied, through
edge detection, to determine the dominant orientation. It is rotated,
if need be, to make the dominant orientation near vertical. Three y (i; j ) = y1 (i; j )2 + y2 (i; j )2 (4.3)
algorithms are compared: 1) the proposed adaptive genetic algo-
rithm, 2) the phase based algorithm [11], and 3) the Sobel operator and the phase or edge direction, is obtained as
[15].
In the phase-based method, a Sobel mask operator is applied to the y2 (i; j )
(i; j ) = arctan : (4.4)
image. Instead of using the magnitude of the masks, we use the phase. y1 (i; j )
The binary image is obtained by setting the phase angle higher than
to one and else to zero. It was found that this method gives higher It is the edge strength, when plotted, that constitutes the edges of
accuracy than the regular Sobel operator. the image, after a proper thresholding. In the phase based method,
1138 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 8, NO. 8, AUGUST 1999

(a) (b)

(c) (d)
Fig. 6. Comparison between the estimates of the ridges for a third frame: (a) original 642 64 frame; (b) magnitude of the Sobel operator; (c) ridges
obtained by the phase method; and (d) ridges obtained by the proposed adaptive genetic algorithm.

it is  (i; j ) that has the information about the edges. The threshold determining the ridges or edges in the fingerprint. Currently we
is set at  . are investigating the problem of removing the writings from the
In the proposed adaptive genetic algorithm, as explained in fingerprint.
Section III, the minimum and maximum widths of the lines are set
to be two and six pixels, respectively. These values were obtained
REFERENCES
from the scanned image under the correct orientation. Thus, in 64
2 64 image, there are a maximum of 32 lines and a minimum of [1] A. Abutaleb, “Automatic thresholding of gray level pictures using two
11 lines. In the presented results, the ridges are made to be vertical dimensional entropy,” Comput. Vis., Graph., Image Process., vol. 47,
or at most inclined by 45 . If no orientation is allowed, then the pp. 22–32, 1989.
[2] M. Chong, R. Gay, H. Tan, and J. Liu, “Automatic representation
maximum widths should be increased to 12 or more pixels. This will of finger prints for data compression by B-spline functions,” Pattern
only increase the computational time. Recognit., vol. 25, pp. 1199–1210, 1992.
Three frames, each 64 2 64, are shown in Figs. 4–6. They all suffer [3] M. Eleccion, “Automatic finger print identification,” IEEE Spectrum,
from the problems of over-inking and under-inking in parts of the 1973.
[4] L. Fogel, A. Owens, and M. Walsh, Artificial Intelligence Through
frames. The results of the edges obtained through the three methods Simulated Evolution. New York: Wiley, 1996.
are shown. It is clear that the adaptive genetic algorithms performs [5] D. Goldberg, Genetic Algorithms in Search, Optimization, and Machine
better than the other methods especially for areas of over/under Learning. Reading, MA: Addison-Wesley, 1989.
inking. [6] J. Holland, “Adaptive plans optimal for payoff-only environments,” in
Proc. 2nd Hawaii Int. Conf. Systems Sciences, 1969, pp. 917–920.
In summary, we have presented, in this paper, a new approach [7] , Adaptation in Neural and Artificial Systems, 2nd ed. Cambridge,
to find the edges or ridges in the fingerprint. We used the fact that MA: MIT Press, 1992.
a fingerprint is made of white followed by black lines of bounded [8] A. Hrechak and J. McHugh, “Automated finger print recognition using
number of pixels. This enabled the problem formulation to be cast structural matching,” Pattern Recognit., vol. 23, pp. 893–904, 1990.
[9] “Special issue on genetic algorithms or evolutionary computation,” IEEE
as a parametric optimization problem. The parameters are the widths Trans. Neural Networks, vol. 5, 1994.
of the black and white lines in the scanned line in the fingerprint. [10] B. Jang and R. Chin, “One-pass parallel thinning analysis, properties
The proposed adaptive genetic algorithm proved to be effective in and quantitative evaluation,” IEEE Trans. Pattern Anal. Machine Intell.,
IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 8, NO. 8, AUGUST 1999 1139

vol. 14, 1992. solution is done via an appropriate optimization algorithm. There
[11] M. Kamel, A. Abutaleb, and M. Rasmy, “A gradient based preprocessing exist a number of optimization algorithms that are well studied in the
algorithm for finger print identification,” in Proc. 1st Int. IEEE Conf. literature. The commonly used algorithms [2] are stochastic simulated
Electronics, Circuits, and Systems, Cairo, Egypt, 1994, pp. 70–75.
[12] E. Mjoisness, “Neural networks, pattern recognition, and finger print annealing (SSA), iterated conditional modes (ICM), and mean field
hallucination,” Ph.D. dissertation, Calif. Inst. Technol., Pasadena, CA, annealing (MFA) [3]. In this paper, we investigate the use of relax-
1986. ation labeling (RL) [4], [5]. RL was originally introduced for context
[13] C. Price, W. Wambacq, and A. Oosterlinck, “Image enhancement and and constraint modeling in labeling problems, and it has seldom
analysis with reaction-diffusion paradigm,” Proc. Inst. Elect. Eng., vol.
137, pt. I, no. 3. been investigated for solving image-processing problems modeled by
[14] K. Rao and K. Black, “Type classification of fingerprints: A syntactic GMRF. Section II presents the GMRF model used for missing-data
approach,” IEEE Trans. Pattern Anal. Machine Intell., vol. 2, 1980. detection in image sequences. Section III outlines the formulation of
[15] M. A. Sid-Ahmed, Image Processing: Theory, Algorithms, and Archi- the optimization algorithm using relaxation labeling. The performance
tectures. New York: McGraw-Hill, 1995. of the proposed RL optimization algorithm is compared with other
optimization algorithms and presented in Section IV with respect
to quality and computational efficiency. Section V, summarizes the
work presented in this paper.

On the Computational Aspects of Gibbs–Markov Random II. GMRF MODEL FOR THE DETECTION
Field Modeling of Missing-Data in Image Sequences OF MISSING-DATA IN IMAGE SEQUENCES

Dilip Krishnan, M. N. Chong, and Showbhik Kalra The detection of missing-data in image sequences is an important
problem in image sequence restoration and the GMRF model has
been used for the mentioned problem [1], [6]. Blotches in degraded
Abstract— Gibbs-Markov random field (GMRF) modeling has been movie frames exhibit the temporal discontinuity characteristic, which
shown to be a robust method in the detection of missing-data in image is commonly used for the detection of blotches. The detection model
sequences for a video restoration application. However, the maximum developed by Chong et al. [6] addresses the problem of incorrect
a posteriori probability (MAP) estimation of the GMRF model requires
computationally expensive optimization algorithms in order to achieve
detection due to poor motion compensation at the moving edges. It is
an optimal solution. The continuous relaxation labeling (RL) is explored found that the degree of incorrect detection (at the moving edges) in
in this paper as an efficient approach for solving the optimization image sequences is reduced significantly by incorporating a moving-
problem. The conversion of the original combinatorial optimization into edge detector into the MRF model. This edge preserving MRF model
a continuous RL formulation is presented. The performance of the RL [6] is adopted here for the detection of missing-data (blotches):
formulation is analyzed and compared with that of other optimization
methods such as stochastic simulated annealing, iterated conditional
P (D = djX = x) = P (D = d)P (X = xjD = d)
01
modes, and mean field annealing. The results show that RL holds out
[0 1 + (xi ; vi ))f (di )
promise as an optimization algorithm for problems in image sequence 1
= exp
processing. Z T i2S
Index Terms— Gibbs–Markov random field, missing-data detection, + ( 2 + (xi ; vi )) (1 0 di )
relaxation labeling, simulated annealing.
+ (1 0 di )(xi 0 xi +v
2
) ] (1)
I. INTRODUCTION
where S denotes the pixel lattice of two adjacent frames from a
This paper discusses the development of an efficient Gibbs–Markov sequence; xi is the observed intensity at site i of the lattice S ; D
random field (GMRF) model for the detection of missing data in denotes the blotch detection frame which is to be estimated using
image sequences [1]. The main application of the proposed model MAP formulation; di = 1 indicates the presence of a blotch at site
is the restoration of old degraded movies without causing distortions i and di = 0 denotes no blotch at site i; X denotes the observed
to nondegraded areas of image frames. The general procedure in image frame with the intensity of each pixel, xi ; xi+v denotes the
the solution of a problem using GMRF modeling is to specify an single motion-compensated neighbor pixel from the neighbor frame;
objective function, U (x); that quantifies the a posteriori assumptions vi denotes the motion vector at site i which is computed using a
of the problem at hand. A Gibbs distribution can then be specified multiresolution full-search block matching algorithm; f (di ) is the
as P (x) = e0U (x)=T where T is the global control parameter called number of the four neighbors (first-order clique) of di with the same
temperature and therefore P (x) is the a posteriori distribution. value as di ;  (1) is the delta function, and function (xi ; vi ) is
The MAP estimation is carried out on P (x) to determine the a moving edge detector. The moving edge detector (xi ; vi ) can
mode of the distribution. This mode represents the highest probability be seen as a weighting function to alleviate the false detection of
(lowest energy) solution of the problem. The MAP estimation is the temporal discontinuity that is due to poor motion estimate; the
exceedingly tedious, especially as the number of unknowns in the prior pdf will be weighted more than the likelihood function when
objective function increases. The attempt at finding an optimal MAP a moving-edge is found. ; 1 ; 2 are the parameters used in the
Manuscript received September 9, 1997; revised June 19, 1998. The estimation.
associate editor coordinating the review of this manuscript and approving An optimization algorithm will be required to solve (1). Nonde-
it for publication was Dr. Mark R. Brown. terministic (stochastic) optimization techniques such as SSA seek
The authors are with the School of Applied Science, Nanyang Tech-
nological University, Singapore 639798, Republic of Singapore (e-mail: to maximize the a posteriori probability based on the controlling
asmnchong@ntu.edu.sg). parameter (temperature) T: Deterministic optimization techniques
Publisher Item Identifier S 1057-7149(99)06007-8. such as relaxation labeling, on the other hand, would deal with the

1057–7149/99$10.00  1999 IEEE

Vous aimerez peut-être aussi