This action might not be possible to undo. Are you sure you want to continue?

Feature Extraction Using Gabor-Filter and Recursive Fisher Linear Discriminant with Application in Fingerprint Identification

M. Dadgostar1, P. R. Tabrizi2, E. Fatemizadeh3, H. Soltanian-Zadeh2 1 Islamic Azad University, Science & Research Branch, School of Biomedical Engineering 2 University of Tehran, School of Electrical and Computer Engineering 3 Sharif University of Technology, School of Electrical Engineering mehrdad.dadgostar@gmail.com, p.roshani@ece.ut.ac.ir, fatemizadeh@sharif.edu, hszadeh@ut.ac.ir Abstract

Fingerprint is widely used in identification and verification systems. In this paper, we present a novel feature extraction method based on Gabor filter and Recursive Fisher Linear Discriminate (RFLD) algorithm, which is used for fingerprint identification. Our proposed method is assessed on images from the biolab database. Experimental results show that applying RFLD to a Gabor filter in four orientations, in comparison with Gabor filter and PCA transform, increases the identification accuracy from 85.2% to 95.2% by nearest cluster center point classifier with Leave-One-Out method. Also, it has shown that applying RFLD to a Gabor filter in four orientations, in comparison with Gabor filter and PCA transform, increases the identification accuracy from 81.9% to 100% by 3NN classifier. The proposed method has lower computational complexity and higher accuracy rates than conventional methods based on texture features. Therefore, it is appropriate to use Gabor filters as bandpass filters to remove the noise and preserve true ridge/valley structures [2]. Lee et al [3] used Gabor filter-based features for fingerprint recognition. Their method required an additional step to detect the reference point (core point) in the fingerprint image. Then the magnitude of Gabor features was extracted only from a small subimage centered in the core point. Tico et al [4] extracted features from the wavelet transform of the discrete fingerprint images on biolab database [5]. They also selected a subset of 104 fingerprint images from biolab database and tested the proposed method of [3]. Usually, features extracted by the Gabor filter are in very large dimensions. Traditionally, principal component analysis (PCA) and linear discriminator analysis (LDA/FLD) have been the standard approach to reduce the high-dimensional original pattern vector space into low-dimensional feature vector space [1]. Although these methods are the same in some aspects, FLD has proven to be more efficient than PCA in pattern recognition problems [1]. The main advantage of FLD over PCA is in extracting more discriminant features. A serious limitation in using FLD is that the total number of the features available from FLD is limited to c-1, where c is the number of classes. Xiang et al [1] suggested a recursive procedure for extracting FLD features called Recursive Fisher Linear Discriminant (RFLD). In this paper, images of biolab data base are used with size of 256×256 pixels. Our proposed method is built into four steps which include: 1. Decomposing images into 32×32 pixel cells and normalizing them using mean and variance values both equal to 100. 2. Applying Gabor filter in different directions and frequency equal to 10.

1. Introduction

Biometry is an important technique that identifies an individual based on specific physiological or behavioral characteristic. Fingerprint is one of the important physiological characteristics for personal verification and identification. Uniqueness of fingerprint has been examined and proved. In this paper, we propose a new method for fingerprint identification based on Gabor filter and RFLD (Recursive Fisher Linear Discriminate) algorithm [1]. Gabor filters have both frequency-selective and orientation-selective properties and have optimal joint resolution in both spatial and frequency domains.

978-0-7695-3520-3/09 $25.00 © 2009 IEEE DOI 10.1109/ICAPR.2009.64 217

x2 .θ ) = exp ⎨ ⎢ 2 + 2 ⎥ ⎬ cos( 2πfx ′). S B and S w matrixes are defined as follows: SB = ∑ n i ( m i − m )( m i − m )T . respectively. Suppose that we have a set of d-dimensional samples x1 . and N i ( x. Gabor filters are band-pass filters which have both orientation-selective and frequencyselective properties and also optimal joint resolution in both spatial and frequency domains. 6. a brief conclusion is given in section 4. i =1 c (6) 218 .. y ) > M i ⎪M 0 + Vi ⎪ N i ( x. A fingerprint convolved with a 0o-oriented filter accentuates those ridges which are parallel to the -axis and smoothes the ridges in the other directions. (4) (2) where f is the frequency of the sinusoidal plane wave along the direction θ from the x-axis.. For all the pixels in the cell. M i and Vi . The average inter-ridge distance is approximately 10 pixels. y ′ = x cos θ − y sin θ . and δ x and δ y specify the Gaussian envelope along x and y axes. Gabor filter Gabor filters are widely used because of their capability for removing noises and enhancing the ridge and valley structures. 4. 3. the normalized image is defined as: ⎧ (V0 ) × ( I ( x. These accentuated ridges and furrow structures constitute an efficient representation of a fingerprint image [6]. Viθ .. [7]. A bank of Gabor filters with different orientations is applied to each of the tessellated cells. y ) .. (5) S B is between-class scatter matrix and S w is within- class distribution matrix. 10. we normalize the gray level intensities in each cell separately to a constant mean and variance.. y) denote the gray level of pixel (x. where Ki is the number of pixels in each cell and Piθ is the mean of the pixel values in Fiθ . y ) = ⎨ ⎪ (V0 ) × ( I ( x. Normalization is a pixel-wise operation which does not change the clarity of the ridge and furrow structures. is defined as standard deviation of each cell with the following equation: Viθ = (1) ∑ ( F θ ( x. An even symmetric Gabor filter has the following general form in the spatial domain: ⎫ ⎧ ⎡ ′2 y ′2 ⎤ ⎪ ⎪−1 x G ( x. the estimated mean and variance of the cell. Let I(x. y ) − M i2 ) . The paper is organized as follows: Section 2 describes the feature extraction algorithm. If normalization is done on the entire image. y ) be the component image corresponding to θ for each cell. The standard deviation of intensity in each filtered cell is treated as a feature value. f . y ) − M i2 ) .. respectively. Feature extraction In this method. this results in 64 tessellated cells. Normalization is done to remove the effects of sensor noise and finger pressure differences [6]. Section 3 shows experimental results on the biolab data base. These Gabor filters are performed in 4. the filter frequency f is set to the average ridge frequency (1/K). Image classification using 3NN classifier and nearest cluster center point classifier with Leave-OneOut method. By applying properly tuned Gabor filters to a fingerprint image. 900 and 1350) with respect to the x-axis. y). In our algorithm. RFLD Algorithm The FLD algorithm selects w as a linear sub-area which maximizes the following Fisher indicator [1]: J ( w) = wT S B w wT S w w . 450.1. the true ridge and furrow structures can be greatly accentuated. y ) − P θ ) i i Ki 2 . i = 1. 2.. 8. x n belonging to c different classes with samples ni in the subset Di labeled wi . For a 256×256 image. Finally. We have used four different values θ (00. Normalization of each cell separately alleviates this problem [8].3. respectively. δ y ⎥⎪ ⎪ 2 ⎢δx ⎦⎭ ⎩ ⎣ where M 0 and V0 are the desired mean and variance values. Each image is tessellated into equal-sized nonoverlapping rectangular cells of predefined dimensions x′ = x sin θ + y cosθ . the normalized gray level value of pixel (x. y). Both values of δ x and δ y are empirically set to 4. if I ( x. feature extraction includes two steps: employing Gabor filter in different orientations and then using RFLD algorithm. Otherwise ⎪M 0 − Vi ⎩ (3) 2. and 12 orientations. y . where K is the inter-ridge distance.. A feature. Let Fiθ ( x. Before filtering the fingerprint image. c . then it cannot compensate for the intensity variations in the different parts of the finger due to finger pressure differences. Applying RFLD algorithm for extracting more discriminant features.

Derive the first feature direction w1 as the normalized eigenvector with the largest eigenvalue of − the matrix S w1S B .. . Wk = ⎢ ⎢. The following equation can be used to maximize J (w) : (8) S w = λS w → S −1S w = λw B w w B ( WNk −1) {w1.. we divided the fingerprint image into a set of 32×32 non-overlapping blocks.2% to 95. . 768 and 1024. 4. we calculate the mean point for each cluster. all the information represented by the “old” features are discarded and then new feature vector is calculated based upon all the feature vectors obtained previously..S w = ∑ ∑ ( x − mi )( x − mi )T .2% with no rejection option in comparison with Gabor filter and PCA transform applied. ( (k S w ) . we used RFLD algorithms. Then. The recognition rate for nearest cluster 219 in which w1 . 3. vk −1 . Discussion and Conclusion In the proposed method. the best results were achieved by using RFLD algorithms and Gabor filter in 4 orientations. ⎥ Bk = ⎢ ⎥ . c-1 features can be extracted at maximum. 8. 6. ⎢0 ⎥ ⎢ ⎦ ⎣ T ⎤ ⎥ ⎥ ⎥ ⎥. respectively. 10 and 12 orientations with f=0. We set the value of both M0 and V0 to 100. In order to reduce the dimension and also to extract more discriminant features.. First. ⎢. In order to maximize the value of J (w) . The results that obtained by nearest cluster center point classifier with Leave-One-Out method are shown in table 1. i =1 x∈Di c (7) ( ( ( ( ( S Bk ) = W Nk −1) (W Nk −1) ) T S BW Nk −1) (W Nk −1) )T ( S wk ) ( ( = W Nk −1) (W Nk −1) )T ( ( S wW Nk −1) (W Nk −1) )T (11) (12) m is the d-dimensional sample mean for the whole set. Thus. vd } constitutes an orthonormal basis = [vk . RFLD algorithm obtains the feature vectors step by step. In order to eliminate this upper bound on the total number of discriminant features.1 and δ x = δ y = 4. 4.. Gabor filterbanks were used in 4.. features were extracted using Gabor filter and RFLD algorithms. (10) 5.. we use the Euclidean distance of each test data from means to classify the data in clusters. mi is the sample mean for class labeled wi [1].. Use PCA to reduce the dimension of the original sample space n-1 to so that S w is nonsingular. ⎥ ⎢. S Bk ) are updated by (11) and (12) respectively. In nearest cluster center point classifier with Leave-One-Out method.. In this method. vk . Within-class scatter matrix. wk are the feature vectors obtained from previous k − 1 steps. perfect recognition with zero error rates were achieved with no rejection option. and the between-class scatter matrix. 640.. we normalized the gray level intensities in each block separately to a constant mean (M0) and variance (V0). Feature extraction procedure of first step is the same as that from the classical FLD algorithm.. In application of FLD and RFLD algorithm with Gabor filter. For identification process we used 3NN classifier and nearest cluster center point classifier with LeaveOne-Out method.. ⎥ ⎥ ⎥ ⎢wk −1 ⎥ ⎣ ⎦ (k ) (9) were 256 (64×4). Compute the kth feature vector wk as the normalized eigenvector with the largest eigenvalue from following eigenvalue equation: (WkT Wk ) −1WkT Bk wk = λwk . 5. It is evident from table 1 that the classifiers using RFLD with Gabor filter in 4 orientations have achieved the highest accuracy rate. As a result. Initialize S1 and S1 with the original scatter B w matrices S B and S w . vd ] where the set of vectors for R d . Table 2 indicates the results of using 3NN classifier. wk −1. k > 1 where Bk and Wk are defined as follows: ⎡S w ⎡ S (k ) ⎤ ⎢ B ⎥ ⎢ T ⎢0 ⎥ ⎢w1 ⎥ ⎢ ⎢.. The algorithm for RFLD is outlined as follows [1]: 1. a recursive procedure applying essentially the same basic idea of FLD is proposed by Xiang et al [1]. the selected features should maximize S B and minimize S w . As shown in table 1 and 2. 2.. w2 .. After first step. ⎥ ⎢ ⎢. the dimensions of feature vector Since the rank of S B is at most c-1. Features extracted by Gabor filter are usually high dimensional. Discard the extracted information from all samples and then calculate a new feature vector. The images are 256×256 pixels in 256 gray scale levels [5]. 512. Biolab database contains 168 fingerprint images from 21 persons (8 items per person). Experimental Results The method was evaluated on the biolab database images. Then. a new feature vector of length 21 was achieved by applying RFLD algorithm. the accuracy rate increased from 85..

IEEE Trans. Electron.9 1 100 1 100 6-Orientations Gabor Filter Number Accuracy of Rate (%) Features 21 84. no.. Prabhakar. X.6 19 92. 2097-2105. Tables 1 and 2 illustrate our method have higher accuracy rates than the proposed method in [4]. Apr. Tico et al [4] employed wavelet-based features for fingerprint recognition. 15. Saarinen. Jan. 37. “Fingerprint Feature Extraction Using Gabor Filter”. 8. 20. [3] C. A. In addition. 348-359.6 10-Orientations Gabor Filter Number Accuracy of Rate (%) Features 21 88. In comparison. H. no. Conf. We compared our method with the methods proposed by Lee and Wang in [3] and Tico et al in [4] which also used biolab database. 2006. 1999. “Face Recognition Using Recursive Fisher Linear Discriminant”. and S. [2] L.2% and 100% respectively. Prabhakar.6 8-Orientations Gabor Filter Number Accuracy of Rate (%) Features 20 86. Their method required reference point detection. [7] A.8 19 92. vol. and L. Ross. Their highest recognition rate of 95. Machine Intel. IEEE Trans. 21-22. [8] A.5 1 100 1 100 Gabor + PCA Gabor + FLD Gabor + RFLD 220 . Feature extraction methods 4-Orientations Gabor Filter Number Accuracy of Rate (%) Features 13 81. but table 1 shows that RFLD algorithm has more discriminant information than FLD algorithm. 282–285. vol.6 2 100 2 100 12-Orientations Gabor Filter Number Accuracy of Rate (%) Features 14 88. “Fingerprint Matching Using Minutiae and Texture Features”. Image Processing.csr. IEEE Trans. 5. The accuracy rates of FLD and RFLD algorithms are equal in table 2. Jain. Thessaloniki. Hong.8 20 92. and S. Lett.2 6-Orientations Gabor Filter Number Accuracy of Rate (%) Features 20 86. Xiang. our method obtains the same recognition rate by 3NN classifier with no rejection option and does not need to find the reference point. S. It achieved recognition rates of 100% by 3NN classifier with rejection option [4]. and J. 21. References [1] C. J. 1998. Lee. no. Lett. May 2000.2 19 91.center point classifier and 3NN classifiers are 95.. [5] Biometric System Lab. “Fingerprint Image Enhancement: Algorithm and Performance Evaluation”.K. Fan. Feb. Hong. Image Process.K. 2001. pp. pp.8 17 94. Agu. Pattern Anal. Lee and Wang used Gabor filter-based features for fingerprint recognition.. 5.3 19 91. 288-290. pp. L.8 12-Orientations Gabor Filter Number Accuracy of Rate (%) Features 20 89. Recognition rate in different feature extraction methods by nearest cluster center point classifier with Leave-One-Out method. 1999. pp.1 21 95. and S. Table 1. P. Jain.2 21 92.8 Gabor + PCA Gabor + FLD Gabor + RFLD Table 2. [4] M. 35.. IEEE Trans.K. vol.unibo. Wang. vol. 2001. Pankanti. Y.it/research/biolab/ [6] A. 8. Feature extraction methods 4-Orientations Gabor Filter Number Accuracy of Rate (%) Features 21 85. Jain. Pattern Anal.777-789. our method has lower computational complexity with higher accuracy rate.6 18 94. D. “A Multi-Channel Approach to Fingerprint Classification”. vol. in Proc. 4. S. A. pp. Recognition rate in different feature extraction methods by 3NN classifier. Tico. “Wavelet Domain Features for Fingerprint Recognition”. Kuosmanen.. Eleclron. Prabhakar.. no. no. and A. Greece. Hong. 9. Image Process.6 20 91. Jain. Internat.7 1 100 1 100 8-Orientations Gabor Filter Number Accuracy of Rate (%) Features 19 87. and T. 4.6 2 100 2 100 10-Orientations Gabor Filter Number Accuracy of Rate (%) Features 15 87. Lee. Oct. “Filterbank-based Fingerprint Matching”. 1. Machine Intel. pp.2% was achieved for the 3NN classifier with no rejection option. vol. Wan. Aug.K. http://www. pp. 846-859. no.

Sign up to vote on this title

UsefulNot useful- An Intelligent Approach to Recognize Touchless Written Bengali Charactersby Bikash Karmokar
- A Chemometric Study on the Analgesic Activity of Cannabinoid Compounds Using SDA, KNN and SIMCA Methodsby XiaohanJin
- Dimension Reduction by Mutual Information Feature Extractionby ijcsitjournal
- jicvol7no1paper04by tarillo2

- EigenFisherFaceby Osman
- 10.Decision Fulby TJPRC Publications
- Robust Human Emotion Analysis Using LBP, GLCM and PNN Classifierby ijtetjournal
- Comparación de técnicas de detección de cambios para monitoreo desmonte y regeneración en bosques tropicales en una serie de tiempo.by Armando Rodriguez Montellano

- Fingerprint Based Gender Classification Using 2D Discrete Wavelet Transforms and Principal Component Analysis
- A Note on Fingerprint Recognition Systems---IEEE2011
- Data Mining
- 100502
- A DCT Based Local Feature Extraction Algorithm for Palm Print Recognition
- fisherface-pami97
- IEEE Paper (Traffic sign detection)
- Detection and Rectification of Distorted Palm Print Images
- An Intelligent Approach to Recognize Touchless Written Bengali Characters
- A Chemometric Study on the Analgesic Activity of Cannabinoid Compounds Using SDA, KNN and SIMCA Methods
- Dimension Reduction by Mutual Information Feature Extraction
- jicvol7no1paper04
- plant classification survey
- e0b49526fa2d90928c
- Assessment of Image Classifications Using Compressed Multispectral Satellite Data (MSD)
- Henry Fingerprint Classification
- EigenFisherFace
- 10.Decision Ful
- Robust Human Emotion Analysis Using LBP, GLCM and PNN Classifier
- Comparación de técnicas de detección de cambios para monitoreo desmonte y regeneración en bosques tropicales en una serie de tiempo.
- An Improved K-Nearest Neighbor Classification Using Genetic Algorithm
- P74
- Project Report
- Survey on Efficient Feature Subset Selection Technique on High Dimensional Small Sized Data
- Luswishi Farmblock Land Classifciation Chisanga
- Detection of Spoilage Markers in Food Products Using a Mass Spectrometer
- 213-H282
- luswishifarmblocklandclassifciation18072012-121126020927-phpapp02
- Carl 08
- 04815224
- Gab Or

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd