International Journal of Physical Sciences Full Length Research Paper

Occlusion, reflections and iris shape deformations are the obstacles that stand in the way of a complete solution to iris localization problem. How to reject outliers caused by occlusion and reflections as much as possible before ellipse or spline fitting is a key challenge. For this reason, we proposed a Hough clustering method, which utilizes the shape configuration of iris edge points and their local appearance characteristics to distinguish iris from non-iris edge points. The experimental results show an improved localization performance of the proposed algorithm on CASIA2.0 and 3.0 databases. 
 
   
 
 Key words: Hough clustering, local edge point experts.


INTRODUCTION
Iris recognition is one of the most reliable biometric identifier and is the preferred mode for biometric recognition (Daugman, 1993).As an emerging subfield in biometric recognition, iris recognition has enjoyed much attention in recent years (Daugman, 1993;Du et al., 2010;He et al., 2009;Wildes, 1997), and substantial progress has been made since the initial works by Du et al. (2009).Nonetheless, many aspects of iris recognition remain only partially solved.
One crucial, yet elusive, component of iris recognition is the iris localization problem.As with other biometric recognition problems, registration is a key to accurate iris recognition since it allows "apples-to-apples" comparisons.Imprecision in iris localization is the principal cause of miss-match errors (He et al., 2009).
The difficulty of iris localization stems from the complexities of real-world images, such as, the presence of eyelid/eyelash occlusion, specular reflections and outof-axis gazing.The difficulty of the problem is compounded by the typical expectation of real-time processing speeds, which limits the range of solutions that can be employed.
Most existing methods to iris localization effect a compromise between the goals of robustness and efficiency.A common theme amongst these methods is the use of a simplifying assumption about iris geometry coupled with an exhaustive search procedure.An example includes the early work of Daugman (1993), where the integro-differential operator was proposed for robust iris localization.Another case is that of Wildes (1997), where the use of Hough voting was proposed for iris localization.The drawbacks of these methods are: (1) a strict circular shape hypothesis cannot account for person specific iris characteristics as well as perspective deformations; (2) an exhaustive search over a 3D parameter space is computationally expensive.Remedies for these drawbacks have been proposed in recent works (Du et al., 2009;Zuo and Schmid, 2010;Proença, 2010).These methods typically precede two stages.First, the iris region is detected by some learning based method (that is, adaboost, neural network, etc.).This significantly reduces the number of candidates for iris edge points.Second, an ellipse or smoothing spline is applied to fit these edge points.Although, this two stage scheme typically incurs low computational cost, it depends on the accurate delineation between iris and non-iris edge points in the first stage.As such, in this work, we propose an approach that more accurately detects relevant edge points and their physiological assignment (that is, iris inner and outer edges).

PROPOSED METHODS
Here, we will first describe how to detect image edge points using multi-scale theory.Next, we introduce the local edge point experts' model for estimating the confidence of one edge point belonging to iris edge only according to its local patch appearance.Also, we explain Hough clustering for iris edge point selecting from all edge points.Finally, we present smooth spline fitting as our curve fitting method.

Edge point detection
We firstly detect image edge points based on multi-scale theory (Lindeberg, 1998).Based on multi-scale theory, detecting edges at a coarse scale can exclude noise and high frequency image texture as much as possible, while localizing edges at a finer scale has high localization precision.Therefore, we first detect edge points at a detection scale (coarse scale) to avoid the influence of eyelash and noise, and then localize the edge points detected at a coarse scale to a finer scale.In this paper, the multi-scale edge detection method based on local directional derivatives (Lindeberg, 1998) was adopted to obtain a set of edge points n (x 1 ; y 1 ) T ; (x 2 ; y 2 ) T ¢¢¢(x N ; y N ) T o along with their gradient directionsf Á 1 ; Á 2 ; N g.In addition, the gradient direction of every edge point was calculated at a coarse image scale to avoid the disturbing of noise and iris texture.

Local edge point experts
Local discriminative model is an efficient way to measure the similarity between the testing local feature and the learned features; however, it is seldom used in iris localization.In this work, we employ it to estimate the likelihood of an edge point belonging to iris inner or outer edge only according to its local patch feature matching.Let f f 1 ; f 2 ; ¢¢¢f N g denotes the local patch features of edge points.They are extracted from iris image, in which the specular reflections have been removed.The local image patch feature f i 2 R M of the i -th edge point refers to the vector concatenation of image intensity values within a P £ P region around this position x i = (x i ; y i ) T .The dimension M is equal to P £ P .Given local image patch feature of every edge point, its local expert (the likelihood of f i being iris inner or outer edge according to local patch feature matching) can be calculated by employing a support vector machines (SVM) classifier.Similar local experts have been used to creat effect in deformable face fitting (Saragih et al., 2009).In this work, it was used to approximate the probability of an edge point belonging to iris inner or outer edge through local patch feature matching.The local edge point experts of the i -th edge point (located atx i ) is defined as below: out er was similar.To reduce the intra-class variations, every patch was normalized with respect to its dominant gradient direction.

Hough clustering
We cannot determine whether an edge point is iris inner or outer edge accurately from the local image patch feature only.However, combing the global spatial configuration of edge points, we can distinguish iris from non-iris edge points accurately.As we know, the shape of iris edge is circle-like, so we can first detect the center location and scale of iris inner or outer edge, and then, discriminate iris and non-iris edge points according to its local patch appearance coupled with their spatial distribution relative to the detected center location and scale.This method is called Hough clustering.
The center and scale detection for iris inner or outer edge can proceed through maximizing the following Hough score (Leibe et al., 2008;Maji and Malik, 2009) formulated under a probabilistic framework that is robust against small iris shape deformations and noise.This formulation also enables the optimal parameter search in a continuous parameter space by using mean-shift algorithm.The Hough score is obtained by marginalizing over all the edge points: scale can be described by µ e = (x ce ; y ce ; ¾ e ).We denote this probability by ¼ i e .The triple l i = (x i ; Á i ; f i ) denote the observed local features of the i -th edge point, where x i is its 2D location in image domain, Á i is its gradient direction and f i is its local patch feature.The discrete random variable o i denotes whether the i -th edge point is on a circle-like shape edge (O) or not.c i is another discrete random variable denoting whether the ith edge point have iris inner edge or outer edge appearance feature.Assuming a uniform prior over all edge point observations: ¼ i e / p(o i = O; c i = C e ; µ e j l i ) = p(µ e j o i = O; x i ; Á i ) p(o i = O j x i ; Á i ) p(c i = C e j f i ) (3) Where Equation 3 utilizes the independence of o i and c i , and that ofx i , Á i and f i is assumed.In Equation 3, the first term is the probabilistic Hough vote for a circular like shape edge's center position and scale.The second term is the probability of an edge point (locatedat x i with gradient direction Á i ) being a circular like shape.The last term represents the likelihood of this edge point being iris inner or outer edge through local image patch feature matching.Assuming that the probability p(o i = O j x i ; Á i ) satisfies uniform distribution.
S C e ; µ e ) / N X i = 1 p(µ e j o i = O; x i ; Á i ) p(c i = C e j f i ) (4)   All that Equation 4 says is that, the Hough score is proportional to the sum of all the probabilistic votes about possible center position and scale of iris inner or outer edge and every vote is weighted by the likelihood of that edge point having iris inner or outer edge appearance feature.This likelihood is denoted by p(c i = C e j f i ) which we get previously.Maximizing the Hough Score S(O; C e ; µ e ) is equal to maximizing the right hand side of Equation 4. To reduce computational complexity, we leverage the gradient direction of every edge point to formulate every probabilistic vote.If knowing the scale of iris inner or outer edge is ¾, based on the fact that the gradient direction of an edge point of iris points towards its center, we can infer that the center position of iris inner or outer edge may be distributed around the below pointx 0 Where x i is the location of the i -th edge point and Á i is its gradient direction.In Figure 2b, the red points and blue points are edge points which have high confidence of being iris inner or outer edge through local appearance feature matching, and the center projection for these edge points at different scale ¾ are shown in Figure 2c.The probabilistic vote about possible parameter of iris inner or outer edge in Equation 4 is approximated by a kernel density estimate (Comaniciu and Meer, 2002) as follows: Where weight (w i ) is equal to the local edge point expert p(c i = C e j f i ).Using a kernel density estimate to approximate the vote for every possible parameter p(µ e j o i = O; x i ; Á i ) is to tolerate a degree of iris shape deformations, because iris inner or outer edge is not a strict circle due to person specific iris characteristics as well as perspective deformations.This effectively smoothes the underlying distribution, where the degree of tolerance to shape deformation is effected through the kernel bandwidth.For each scale ¾ , we find the mode of the kernel density estimation (KDE) using the mean-shift algorithm (Comaniciu and Meer, 2002).The optimal center position (x ¤ ce ) and scale (¾ ¤ e ) of iris inner or outer edge are obtained by comparing the maximum of Equation 6 over all possible scales.Because the exhaustive search is only over scale (that is, one dimensional exhaustive search), the computational speed is very fast.The correspondence set of iris inner and outer edge points (Figure 2d) can be determined by selecting points that satisfy p(µ ¤ e j o and ¿ 2 are two thresholds).That means we select the edge points which have high confidence of being iris inner or outer edge through local appearance feature matching and also distribute under the circle-like shape configuration.

Smoothing spline fitting
After selecting the corresponding set of iris inner and outer edge from all image edge points, the cubic smoothing spline (He et al., 2009) is employed to fit iris inner or outer edge points on polar coordinate system.This is shown in Figure 2e.The first experiment was designed to test the performance of our method when using different local image features.In this work, we propose two types of local image features namely: raw pixel intensity and intensity gradient.Raw pixel intensity is defined as a vector constructed by the intensity value of the very pixel, and the intensity gradient is defined as a vector that is composed by the intensity gradient of every pixel.By adopting different feature types, we can get different equal error rates (EER) from our method.To calculate the EER, we use our proposed method in the localization part and Daugman's method in normalization and matching parts.The result listed in Table 1.

EXPERIMENT
On these datasets, we also compared four types of iris localization methods: (1) the integro-differential operator proposed by Daugman (1993), (2) the traditional Hough transform proposed by Wildes (1997), (3) pulling and pushing method proposed by He (2009), and (4) the edge point subset selection combined with Smoothing spline fitting (EPSS+SSF) proposed in §2 (Figure 3).To give a fair comparison, we adopted different methods in the iris localization part and the same feature extraction and matching methods proposed by Daugman (1993) to  2. All algorithms were implemented in Matlab on a 3.98 GHz duo CPU (1.99 G per core) and 2 GB of memory.The average running time for the 4 algorithms on CASIA-V3.0database are: 10.20, 14.44, 1.99s and 3.31 s respectively.From the accuracy and efficiency comparison, the proposed algorithm was the most accurate localization algorithm though it is slightly slower than that proposed by He (2009).The proposed method's performance is obviously better than that of Daugman's and Wildes' as our method takes into account the shape deformations caused by person specific iris characteristics as well as perspective deformations while Daugman's and wildes' do not.The strict circular shape hypothesis of iris edge in Daugman's and Wildes' method cannot account for any shape deformations leading the non-ideal performance.Moreover, our method's performance was also a little better than He's because our method can determine whether an edge point belongs to iris edge accurately according to its location, intensity gradient direction and local image feature.But He's method discriminate iris and non-iris edge points only according to its spatial distribution relative to the detected iris region and this strategy is not as effective as our proposed method.

Conclusion
In this paper, we present a novel method for iris detection and localization using a weighted kernel voting framework.It improves the performance of iris localization by selecting the right corresponding iris edge point for fitting.A mean-shift search strategy was used to guarantee that the method can tolerate some shape deformations.The validity of the proposed algorithm was verified by testing it on three different iris databases.
are the weight vector and bias of SVM respectively.Figure1illustrates the details of local edge point experts model.the off-line training, we first manually label the pupillary edge points, limbic edge points and outliers.When training w inner and w inner , positive patch examples are obtained by extracting local patches centered at pupillary edge points, while negative patch examples are obtained by sampling patches centered at other remaining edge points.For the training of w out er and b

Figure 2 .
Figure 2. Illustration about selecting corresponding iris inner and outer edge points from non-iris edge points using Hough clustering: (a) the edge map of iris image, (b) the local edge point experts for iris inner or outer edge, (c) the center projection for all the edge points which have high probability of being iris inner and outer edge points through local appearance feature matching, and the Hough clustering results after mean-shift mode-finding, (d) the corresponding iris inner and outer edge points, (e) the smoothing spline fitting results.
This experiment was carried out to evaluate the performance of the proposed method on the problem of iris localization on different databases: (1) the CASIA version 2.0, (2) the CASIA version 3.0(Lamp)(CASIA, 2008), and (3) the real-world database founded by us in natural environment.The CASIA version 2.0 database contains 1200 images with resolution of 640 × 480 from 60 live human test subjects (20 different images for each subject) with two separate visits (40 days apart).The CASIA Version 3.0 (Lamp) contains 411 human test subjects' 16213 images (640 × 480 pixel resolution).The real-world database contains 3300 iris images (640 × 480 pixel resolution) of 126 persons and the image acquisition device is OKI's IRISPASS-h.This database includes a number of low quality images, including those with eyelid/eyelash occlusion, glasses reflections and out of axis gazing.

Table 1 .
The EER comparison of our method using different local image features on CASIA-V3.0,CASIAV2.0 and our real database.