Multi-algorithmic Palmprint Authentication System Based on Score Level Fusion

Automatic palmprint recognition systems are classified into two categories: online and offline. An online system acquires palmprint images using a digital camera or a special capture device that is directly connected to a computer for real-time processing. In offline, palmprint images are obtained using digital scanner which can extract the ridges, singular points, and minutiae point features (Wong et al., 2005). The palmprint images are captured using various devices such as web camera, digital camera, mobile phones, and digital scanners.

The biometric research centre at the Hong Kong Polytechnic University has developed multispectral palmprint image acquisition device to acquire high quality palmprint images of small sized 384 × 284 image pixels at reasonable cost for the recognition system. Figure 2 shows the sample low resolution and high resolution palmprint images.

Fig. 1

Fig. 2

Pre processing is used to segment the region of interest (ROI) and to extract only the required features, thus making it necessary to obtain a subimage from the captured palmprint image. The palmprint images are orientated and normalized before feature extraction to eliminate the variations caused by rotation and translation. Segmentation is a major task in palmprint recognition which aids in enhancing the quality of the palmprint images. ROI provides more information suitable for efficient recognition. Two techniques used in the process are square-based segmentation and circle-based segmentation (Zhang, 2004).

The palmprint images are oriented and translated before the segmentation process. The steps followed to acquire the square-based segmentation (Wang et al., 2007) are given below:

Step 1 Median filter is used to remove noise in the palmprint image.

Step 2 Thresholding is done to convert gray scale image I (i, j) into the binary image. The symbol ‘α’ is the threshold value set as 100. (1) I(i,j)={0,α<0,1,α⩾0.

Step 3 Contour of the palmprint image is obtained using the algorithm of contour tracing. The two valley points C₁, C₂ are located as shown in Figure 3(a).

Step 4 The two valley points are joined using the straight line Eq. (2) (2) y−y1=(y2−y1x2−x1)(x−x1).

Step 5 The points on the contour of palmprint at an angle of 45° and 55° are plotted with the reference straight line L₁ and L₂. Figures 3(b) and (c) illustrate the reference points and the angles made on the reference line. ROI is used to segment the centre part of the palmprint image as shown in Figure 3(d).

The angle can be extended from the reference line up to ±40°. As the valley point is different for every person, this method is highly stable.

Fig. 3

The circle-based segmentation technique is the most efficient technique used to extract the central part of the palmprint image. This technique can also be used independently without feature extraction and classification. The gray image is converted to binary image using Eq. (1). The two valley points C₁, C₂ are obtained using the contour tracing algorithm as shown in Figure 4 (a). C₁, C₂ as shown in Figure 4 (b) is joined using the straight line Eq. (2).

The steps for the circle-based segmentation are given below:

Step 1 The radius of the circle ‘r’ from the midpoint value of the straight line is calculated using Eq. (3). x₁, x₂, y₁, and y₂ are the coordinate points along the X and Y axis, respectively: (3) r=(x2−x1)2+(y2−y1)2.

Step 2 Using the ‘strel’ command, the neighborhood region of the centre point is obtained and image is cropped as shown in Figure 4(c).

It is a very stable phenomenon as the valley points of palmprint differs from person to person. Simultaneously, the radius of the circle also varies for each person. A test database is created using the radius of the circle and the test image of palmprint is checked with the database created. If the radius of the palmprint image matches with the image in the database, then the palmprint belonging to a particular person can be identified easily. FRR and FAR are highly reduced to <0.001%.

Fig. 4

DT-CWT is a form of Discrete Wavelet Transform (DWT),which generates complex coefficients by using a dual-tree of wavelet filters to obtain their real and imaginary parts. It is a pyramid-like structure having a Gabor-like frequency response with six multiscale directional subbands (±15°, ±45°, ±75°). It provides two main advantages over the real DWT image representation namely, approximate shift invariance and improved directional selectivity, which renders DT-CWT nearly rotation invariant.

In dual-tree, two real wavelet trees are used with each capable of perfect reconstruction. One tree generates the real part of the transform and the other is used in generating complex part. {H₀(z), H₁(z)} is a quadrature mirror filter (QMF) pair in the real-coefficient analysis branch. For the complex part {G₀(z), G₁(z)} is another QMF pair in the analysis branch (Selesnick et al., 2005). All filter pairs are orthogonal and real valued. This realization laid the foundation of utilizing to further enhance and improve the recognition rates already achieved by other methods. The extracted feature vectors are given as input to a classifier. The DT-CWT can be used for a variety of applications such as edge detection, image restoration, enhancement, and image compression.

The 1D dual-tree wavelet transform is implemented using a pair of filter banks operating on the same data simultaneously. The upper iterated filter bank represents the real part of a complex wavelet transform. The lower one represents the imaginary part. The orthogonal two channel filter bank with analysis low-pass filter H₀(z),analysis highpass filter H₁(z) and synthesis filters G₀(z) = H₀(z − 1), G₁(z) = H₁(z − 1) are shown in Figure 5.

For an input signal X(z), the analysis part of the filter bank inclusive of subsequent upsampling produces the low-pass cofficients C¹ and the high-pass coefficients D¹ (Zhang et al., 2007): (4) C1(z2)=12{X(z)H0(z)+X(−z)H0(−z)}, (5) D1(z2)=12{X(z)H1(z)+X(−z)H1(−z)}.

The decomposed low frequency Xl1(z) and a high frequency Xh1(z) parts are described in Eqs. (7) and (8). (6) X(z)=Xl1(z)+Xh1(z), (7) Xl1(z)=C1(z2)G0(z)12{X(z)H0(z)G0(z)+X(−z)H0(−z)G0(z)}, (8) Xh1(z)=D1(z2)G1(z)=12{X(z)H1(z)G1(z)+X(−z)H1(−z)G1(z).

However, this decomposition is not shift invariant as X(z) are aliasing terms introduced by the down sampling and up sampling operations. The 2D DT-CWT images are computed by applying the pair of trees to the rows and columns of the images as in the basic DWT. The real DT-CWT of an image x(n) is implemented using two critically sampled separable 2D DWT in parallel. For each pair of subbands, the sum and differences are obtained as shown in Figure 6. Each level of the tree produces six band-pass subimages as well as two low-pass subimages, on which subsequent stages iterate. In the case of real 2D filter banks the three high-pass subbands have orientations of 0°, 45°, and 90° compared to the complex filters which have six high-pass subbands at each level which are oriented at ±15°, ±45°, ±75°. The shift invariance and directionality of the DT-CWT is applied to the palmprint image to extract the optimal features as shown in Figure 7.

Fig. 5

Fig. 6

Fig. 7

The 2D DWT are good at detecting point discontinuities at edge points but are not effective in representing the geometrical smoothness of the contours. Contourlet transform used for feature extraction provides a flexible multi-resolution and directionality to capture smooth contours in the palmprint images. LP combined with Directional Filter Bank (DFB) decomposing the palmprint image into directional subbands at multiple scales is shown in Figure 8. LP decomposition generates down sampled low-pass version of the original image and would obtain the band pass of image by taking the difference between original and prediction values (Burt and Adelson, 1983). Double filter bank is constructed using two blocks, Quincunx filter banks with fan filters and shearing operator. The first block divides into vertical and horizontal directions, and the second block reorders the image samples in the binary level of tree decomposed structure to obtain the desired 2D spectrum division (Do and Vetterli, 2005).

The band-pass image from the LP is fed into DFB to extract 2D directional information of an image (Bamberger and Smith, 1992). The main reason for combining the DFB with LP is to overcome the lack of low frequency in several directional subbands. Figure 9 depicts the contourlet features of palmprint image without preprocessing at pyramid level four decomposition.

Contourlet transform provides 2ⁿ subbands at multiscales. The pyramidal level three and four decomposition of palmprint image produces eight and sixteen subbands. Figure 10 illustrates the contourlet features of palmprint image with preprocessing at pyramid three and four level decompositions.

The contourlet features for classification are obtained by calculating energy for each subband and selecting the subset of feature sequence from a feature set with high dimension. The statistical features are extracted from each subband to create a feature vector. Energy, contrast, correlation, and homogeneity are the discriminating features calculated by using the various properties of Gray Level Co-occurrence Matrix (GLCM) techniques (Ardabili et al., 2011). The properties are as follows: (9) Energy=∑i,jp(i,j)2, (10) Contrast=∑i,j|i−j|2p(i,j), (11) Correlation=∑i,j((i−µi)(j−µj)p(i,j))/(σiσj), (12) Homogenity=∑i,jp(i,j)/(1+|i−j|). where µ, σ are the mean and standard deviation of co-occurrence matrix. The subbands at each level were used to form a matrix. The unwanted coefficients are removed and the feature vector is obtained using the properties of GLCM.

PCA is an appearance-based statistical method used for feature extraction, data compression, redundancy removal, and prediction. The dimensionality is reduced by projecting the feature vector of the palmprint image into a small set of feature space called eigen palms. The projected weights forming eigenpalm vector can represent the principal components of palmprint image. By the integration of contourlet and PCA, the coefficients generated by contourlet transform at multiscales and multidirections are given as input to PCA. The discriminating features of all the subbands are reduced using PCA. The dimensionality after contourlet transform is reduced from 5178*5178 to 6*15 using PCA.

Fig. 8

Fig. 9

Fig. 10

In sum-rule based score level fusion scheme (Jain et al., 2005), the fused score ‘S’ is computed using (14) s=s1+s2, where S₁ and S₂ are the matching scores obtained from matching modules 1 and 2 of multi-algorithm palmprint authentication system after score normalization.

In weighted sum-rule score level fusion (Cui and Yang, 2011), the fused score ‘S’ is computed using (15) s=w1s1+w2s2, where s₁ and s₂ are the matching scores after score normalization, w₁ and w₂ are the weights. The weights w₂ and w₂ can be found using an empirical weighting scheme. It can be varied from 0 to 1 in steps of 0.02 such that w₁ + w₂ = 1. The fused score compared to a decision threshold decide a person to be genuine or an impostor.

Support Vector Machine (SVM) is based on the principle of structural risk minimization which results in fusing the matching scores obtained from the classifiers after score normalization. It yields satisfying result by deploying a nonlinear optimal separating hyper-plane which could separate the datasets more efficiently.

The training data x_i = x_i1, x_i2, …, x_iN correspond to input vector of the normalized matching scores, where i=1,…, N is the number of matching scores used for training and the class y_iϵ(1, -1) (Wang and Han, 2009). In the testing phase, the fused score F_s of test input pattern is calculated using (16) Fs=wxt+b, where w is a weight vector, x_t is the test pattern of the matching scores obtained from classifiers and b, the bias estimated on the test set. Finally, the test pattern x_t belongs to a genuine class if the value of y_i is set to ‘1’ or an impostor class if the value of y_i is set to −1 (Vatsa et al., 2008). The SVM fusion rule enhanced the recognition performance and reduced the error rates compared to the sum rule for the proposed multi-algorithmic palmprint authentication system.

Multi-algorithmic palmprint score level fusion schemes performance was evaluated by plotting the ROC curves in terms of GAR and FAR as shown in Figure 11. GAR is equivalent to (1-FRR). For a FAR of 0.1%, the GAR of SVM fusion method is 98.00%, the GAR of weighted sum-rule method is 96.75% and the GAR of sum-rule fusion method is 96.00%. Among all of these fusion approaches, SVM score level fusion method has given the best GAR of 98.00%.

Equal error rate (EER) is the point in a ROC curve where FAR and FRR assume the same value. ROC curves of SVM, weighted sum-rule and sum-rule score level fusion schemes in terms of FRR and FAR were plotted and shown in Figure 12. From the ROC curve, EER of 1.5% was obtained for the SVM fusion method. For weighted sum-rule and sum-rule match score level fusion approach, EER obtained was 2.20% and 2.50%, respectively. The experimental results demonstrate that SVM match score level fusion outperforms sum-rule and weighted sum-rule method.

The execution times for single algorithmic and multi-algorithmic approaches for palmprint recognition are shown in Tables 1–3. It can be observed that multi-algorithmic approach takes longer time than single algorithmic approaches.

Fig. 11

Fig. 12