This paper presents an efficient algorithm for iris recognition using phase-based image matching technique using phase components in 2D Discrete Fourier Transforms (DFTs) of given images. Experimental evaluation using the CASIA iris image databases (versions 1.0 and 2.0) and Iris Challenge Evaluation (ICE) 2005 database clearly demonstrates that the use of phase components of iris images makes it possible to achieve highly accurate iris recognition with a simple matching algorithm. In order to reduce the size of iris data and to prevent the visibility of iris images, we introduce the idea of 2D Fourier Phase Code (FPC) for representing iris information. The 2D FPC is particularly useful for implementing compact iris recognition devices using Digital Signal Processing (DSP) technology.
Index Terms-Phase-based image matching, phase-only correlation, phase-only matched filtering, biometrics, iris recognition.
I. INTRODUCTION
A major approach for iris recognition today is to generate feature vectors corresponding to individual iris images and to perform iris matching based on some distance metrics. Most of the commercial iris recognition systems implement a famous algorithm using iriscode, which was proposed by Daugman. In this algorithm, 2D Gabor filters are used to extract a feature vector corresponding to a given iris image. Then, the filter outputs are quantized to generate a 2 Kbit iriscode. The dissimilarity between a pair of iriscodes is measured by their
hamming distance based on an exclusive-OR operation. The iriscode is very compact and can be accommodated, even on the magnetic stripe implemented on the back of typical credit cards. In addition, exclusive-OR comparison allows us to perform extremely rapid recognition. On the other hand, one of the difficult problems in feature-based iris recognition is that the matching performance is significantly influenced by many parameters in the feature extraction process (for example, spatial position, orientation, center frequencies, and size parameters for 2D Gabor filter kernels), which may vary, depending on the environmental factors of iris image acquisition. Given a set of test iris images, extensive parameter optimization is required to achieve a higher recognition rate. Addressing the above problem, this paper proposes an efficient iris recognition algorithm using phase-based image matching, that is, an image matching technique using only the phase components in 2D Discrete Fourier Transforms (DFTs) of given images. The technique of phase-based image matching has so far been successfully applied to high-accuracy image registration tasks for computer vision applications where the estimation of subpixel image translation is a major concern. In our previous work on the other hand, we proposed an efficient fingerprint recognition algorithm using phase-based image matching and we have developed commercial fingerprint verification units for access-control applications. The original contribution of this paper is to show that the same matching technique is also highly effective for iris recognition (see our conference papers for earlier discussions of the proposed idea). Experimental evaluation using the CASIA iris image databases (versions 1.0 and 2.0) and Iris Challenge Evaluation (ICE) 2005 database clearly demonstrates that the use of the Fourier phase spectra of iris images makes it possible to achieve highly accurate iris recognition with a simple matching algorithm.
Fig. 1. Flow diagram of the proposed algorithm.
II. PREPROCESSING
An iris image contains some irrelevant parts (e.g., eyelid, sclera, pupil, etc.). Also, the size of an iris may vary depending on camera-to-eye distance and lighting condition. Therefore, the original image needs to be normalized.
A. Iris Localization
This step is to detect the inner boundary (the boundary between the iris and the pupil) and the outer boundary (the boundary between the iris and the sclera) in the original gray-scale image forg(m1,m2) shown in Figure 2 (a).
(a) (b)
(c)
Fig. 2. Iris image: (a) original image, (b) normalized image, and (c) normalized image with eyelid masking.
We use an ellipse as a model of the inner boundary. Let (l 1, l2) be the lengths of the two principal axes of the ellipse, (c1, c2) be its center, and θ be the rotation angle. We can find the optimal estimate (l1, l2, c1, c2, θ) for the inner boundary by maximizing the following absolute difference:
Here, Δl1 and Δl2 are small constants, and S is the N-point contour summation of pixel values along the ellipse:
Thus, we will detect the inner boundary as the ellipse on the image for which there will be sudden change in luminance summed around its perimeter. To reduce computation time, the parameter set (l1, l2, c1, c2, θ) can be simplified depending on iris images. In our experiments using the CASIA iris image database, assuming θ = 0 causes no degradation on its performance.The outer boundary is detected in a similar manner.
B. Iris Normalization
Next step is to normalize iris to compensate for the elastic deformation in iris texture. We unwrap the iris region to a normalized rectangular block of a fixed size (256-128 pixels in our experiments). In order to remove the iris region occluded by the upper eyelid and eyelashes, we use only the lower half of the iris region (Figure 2 (a)) and apply a polar coordinate transformation (with its origin at the center of pupil) to obtain the normalized image shown in Figure 2 (b), where n1 axis corresponds to the angle of polar coordinate system and n2 axis corresponds to the radius. In the transformed iris image, the irrelevant eyelid region should be masked as shown in Figure 2 (c). In general, the eyelid boundary can be
Fig. 3. Effective region extraction: (a) normal case, and (b) case when multiple sub-regions should be extracted.
modeled as an ellipse contour. Hence the same method to detect the inner boundary can be applied to eyelid detection.
III. MATCHING
This section describes the detailed process of effective region extraction, image alignment and matching score calculation. The key idea is to use phase-based image matching for image alignment and matching score calculation.
A. Effective Region Extraction
Given a pair of normalized iris images ˜ f(n1, n2) and ˜g(n1, n2) to be compared, the purpose of this process is to extract, from the two images, the effective regions f(n1, n2) and g(n1, n2) of the same size, which do not contain irrelevant regions as illustrated in Figure 3 (a). A problem occurs when the extracted effective region becomes too small to perform image matching. In such a case, we extract multiple effective sub-regions from each iris image as illustrated in Figure 3 (b) by changing the width parameter w. In our experiments, we extract 6 sub-regions from a single iris image by changing the parameter w as 55, 75 and 95 pixels.
B. Phase-Based Image Matching
Before discussing the image alignment and the matching score calculation, we introduce the principle of phase-based image matching using Phase-Only Correlation (POC) function. Consider two N1-N2-pixel images, f(n1, n2) and g(n1, n2), where we assume that the index ranges are n1 = −M1 · · ·M1 (M1 > 0) and n2 = −M2 · · ·M2 (M2 > 0) for mathematical simplicity, and hence N1 = 2M1 + 1 and N2 = 2M2 + 1. Let F(k1, k2) and G(k1, k2) denote the 2D DFTs of the two images. F(k1, k2) is given by (a) (b)
Fig. 4. Normalized iris image in (a) spatial domain, and in (b) frequency domain (amplitude spectrum), where K1 = 0.55M1 and K2 = 0.2M2
AF (k1, k2) is amplitude and θF (k1, k2) is phase. G(k1, k2) is defined in the same way. The cross-phase spectrumRFG(k1, k2) is given by
where is the complex conjugate of G(k1, k2) and θ(k1, k2) denotes the phase difference θF (k1, k2)−θG(k1, k2). The POC function rfg(n1, n2) is the 2D Inverse DFT (2D IDFT) of RFG(k1, k2) and is given by
When two images are similar, their POC function gives a distinct sharp peak. When two images are not similar, the peak drops significantly. The height of the peak gives a good similarity measure for image matching, and the location of the peak shows the translational displacement between the images. In our previous work, we have proposed the idea of BLPOC (Band-Limited Phase-Only Correlation) function for efficient matching of fingerprint images considering the inherent frequency components in fingerprint images. We have found that the same idea is also very effective for iris recognition. Our observation shows that (i) the 2D DFT of a normalized iris image contains meaningless phase components in high frequency domain, and that (ii) the effective frequency band of the normalized iris image is wider in k1 direction than in k2 direction (see Figure 4). The original POC function rfg(n1, n2) emphasizes the high frequency components, which may have less reliability. We observe that this reduces the height of the correlation peak significantly even if the given two iris images
Fig. 5. Example of genuine matching using the original POC function and the BLPOC function: (a) iris image f(n1, n2), (b) iris image g(n1, n2), (c) original POC function rfg(n1, n2), and (d) BLPOC function rK1K2 fg (n1, n2).
are captured from a common eye. On the other hand, BLPOC function allows us to evaluate the similarity using the inherent frequency band within iris textures. Assume that the ranges of the significant frequency band are given by k1 = −K1 · · ·K1 and k2 = −K2 · · ·K2, where 0≤K1≤M1 and 0≤K2≤M2 as shown in Figure 4 (b). Thus, the effective size of frequency spectrum is given byL1 = 2K1+ 1 and L2 = 2K2 + 1. The BLPOC function is given by
denotes note that the maximum value of the correlation peak of the BLPOC function is always normalized to 1 and does not depend on L1 and L2. Figure 5 shows an example of genuine matching using the original POC function and the BLPOC functionThe BLPOC function provides better discrimination capability than that of the original POC function.
C. Displacement Alignment
This step is to align the translational displacement between the extracted images (Figure 3). Rotation of the camera, head tilt and rotation of the eye within the eye socket may cause the displacements in normalized images. The displacement parameters can be obtained as the peak location of the POC function rfg(n1, n2). The obtained parameters are used to align the images.
Fig. 6. Distributions of matching scores.
D. Matching Score Calculation
In this step, we calculate the BLPOC function between the aligned images f(n1, n2) and g(n1, n2), and evaluate the matching score. In the case of genuine matching, if the displacement between the two images is aligned, the correlation peak of the BLPOC function should appear at the origin (n1, n2) = (0, 0). So, we calculate the matching score as the maximum peak value of the BLPOC function within the r-r window centered at the origin (r = 11 in our experiments). When multiple sub-regions are extracted as illustrated Figure 3 (b), the matching score is calculated by taking an average for effective sub-regions. If the matching score is close to the threshold value to separate genuines and impostors, we calculate the matching score with scale correction (see Figure 1).
IV. EXPERIMENTS AND DISCUSSIONS
This section describes a set of experiments using the CASIA iris image database (ver 1.0) for evaluating matching performance. This database contains 756 gray-scale eye images (320-280 pixels) with 108 unique eyes and 7 different images of each unique eye. We first evaluate the genuine (intra-class) matching scores for all the possible combinations of genuine attempts (7C2-108 = 2268 attempts). Next, we evaluate the impostor (inter-class) matching scores for 108C2 = 5778 impostor attempts, where we take a single image for each eye and make all the possible combinations of impostor attempts.
Figure 6 shows distributions of genuine and impostor matching scores. The figure shows a good separation of genuine and impostor matching scores, where the minimum genuine matching score is 0.1464, and the maximum impostor matching score is 0.1428. The score between these values can be chosen as a threshold to distinguish between the two classes. Thus, for this experiment, we can achieve EER=0%, where the EER (Equal Error Rate) is the error rate where the FNMR (False Non-Match Rate) and the FMR (False Match Rate) are equal. Some reported values of EER from using the CASIA iris image database are also shown in the same figure for reference. Note that the experimental condition in is not the same as our case, because the complete database used in is not available at due to the usage rights of iris images. The number of iris images in the database available at is smaller than the complete database. The result demonstrates a potential possibility of phase-based image matching for creating an efficient iris recognition system.
V. CONCLUSION
In our previous work, we have developed actual fingerprint verification units based on phase-based image matching. In this paper, we have demonstrated that the same approach can be highly effective also for iris recognition. The proposed approach will be useful for implementing a unified hardware/software engine for multimodal biometric system with iris and fingerprint recognition capability.
