Fast Iris Localization Based on Image Algebra and Morphological Operations

The localization of the iris is the most significant factor in biometrics of the iris, which traditionally assumes strictly controlled environments. The proposed method includes the pupillary and limbic iris boundaries localization. A primary advantage of image arithmetic operations is that the process is straightforward and therefore fast, these characteristics are employed and combined with the morphological operators in the designing of the proposed algorithm. The proposed algorithm takes into account the noise area which is found in various parts of the eye image such as light reflections, focus and small visible iris. The experimental results are conducted on a collection of iris images consist of 756 images belong to Chinese Academy of Sciences Institute of Automation (CASIA V-1) and 450 images belong to Multi Media University (MMU V-1) databases. The results indicate a high level of accuracy using the proposed technique. Moreover, the comparison results with the state-ofthe-art iris localization algorithms expose considerable improvement in segmentation accuracy while being computationally more efficient.


Introduction
For the automated personal identification, iris recognition is one of the most accurate and broadly employed methodology.A typical iris recognition system comprises segmentation, normalisation, feature extraction and matching.In an excellent iris authentication system, the iris segmentation is a very significant part, which must be able to cope with many adverse circumstances such as the blurred outer border of iris, light reflex points in pupil, eyelids and effected eyelashes [1].However, traditional iris localization methods and matching features depend only on near-infrared illumination and require samples to be taken under strictly restricted conditions [2], which is the main difficulty of deploying an iris recognition system in civilian and surveillance applications.Several ways to address the problem were suggested in the literature.The accuracy and computation time for currently available iris localization algorithms are still below the expectations and requires further improvement for the deployments.Therefore, the proposed method was designed to take the advantage of the image arithmetic operations features with morphological operators features for localizing the iris.
This paper is organised as follows: Section 2 overviews the existing literature for iris localization.The proposed method for iris localization is explained in section 3. The experimental results of the proposed system are provided in section 4, and final conclusions are shown in section 5.

Literature Review
The localization of the iris represents the core of iris recognition system.Most iris localization is based on techniques proposed by Daugman [3], Wildes [4] and active contour models [5].
Daugman's integro-differential operator [6], [7], [8] searches the image domain for the maximum in the blurred partial derivative, concerning the increasing radius r of the normalised contour integral of I(x, y) along with a circular arc ds of radius r and centre coordinates (x 0 , y 0 ).In Eq. ( 1), G σ (r) is a smoothing function such as a Gaussian of scale σ.The full operator executes a frequent search for a maximal contour integral derivative with increasing radius in successive accurate scales of analysis through the three parameter centre space coordinates and radius (r, x 0 , y 0 ) determining the contour integration path [3].
Despite its high efficiency, Daugman's operator does not always segment the iris accurately when the input image of the iris is noisy.This noise can be in the form of a reflection, or that the eye in the image is not fully open or the eyelashes covering part of the iris area.Low-quality iris images cause a poor recognition performance as long as the traditional processing of iris images is performed On the other hand, Wildes proposes the circular Hough transform.In this method, the boundary is estimated by rediscovering the circular structures within the binary edge map of the points (xj, yj), j = 1, .., n, over the image gradient.The goal is to find a triplet of two-dimensional origin coordinates and radii that maximises the expression H (xc, yc, r) in Eq. ( 2) [4]: where h(x j , y j , x c , y c , r) = 1 when (x j -x c ) 2 + (y j -y c ) 2 -r 2 = 0.
The appropriate majority methods are edge-based methodologies [9][10][11][12][13][14][15], at first; the edge features are utilised to detect the outer and inner iris borders.Images are taken under the visible wavelength spectra but in uncooperative environments of recognition might be lead to light reflections and specular [16], which could produce false edges in edge maps.Beside, eyelashes, poor focus and blurring motion might destroy the edges.Significant experience exposes that the techniques which have been depending on the edge-based detection are susceptible to these false edges.Secondly, the Hough transform is computationally exhaustive through the "brute force" approach, and thus may not be suitable for real-time applications.
Active contour models [5], [17], [18] provide flexible contours that allow the iris and eyelid boundaries to be detected at the same time, but the performance of approaches based on active contour models may be downgraded due to reflections, the heterogeneous of the iris and unclear boundaries, and using such models needs also sensibly of many parameters to get a suitable curve, furthermore these models need computation time complexity.

Proposed Iris Localization Algorithm
The iris of the human is an annular part of a pupil that represents the inner border and the white sclera that represents the outer border.An approach of iris localization algorithm on the basis of image arithmetic operations and morphological operators is proposed.Two main stages are featured in the proposed algorithm, namely iris localization (i.e., Detection of the outer boundary of the iris region and pupil localization (i.e., Detection of the inner boundary of the pupil region).

Pupil Localization
For the detection of the inner circle of the iris, the behaviour of the intensity of the image is taken into consideration in each of the pupil and iris parts of the eye.The overall intensity of the pupil's area is relatively smaller than that of the other, from the whole eye image.Next, the most significant dark and packed part that will appear in the image of the eye representes the pupil.The approximate pupil center is discovered by calculating the center of iris region mass in eye image after getting the benefit of these attributes.The following steps are applied to Localize pupil:

Step1: (Apply image addition)
Because of the features that the pupil owns them such as darkness, large connected, and packed area, thus these features are exploited by image adding operation to increase overall image brightness which is eliminated the most details of the image but it still preserved the area of the pupil.Image addition is applied consecutively using the following formula: Where g(x, y) output pixel values, f(x, y) the input pixel values, T the threshold value which is selected using Otsu global image threshold method [19].The Otsu's method chooses the threshold to reduce the intraclass variance of the thresholded black and white pixels.According to this method, the level representing the value of a normalised intensity is located in the range [0,1].The T is computed using the following formula: Figure (1) shows the result of this step.

Step2: (Cut boundaries)
The image of the eye might hold dark, thick eyebrows, thus to avoid these pixels, considered as pupil pixels (the first top rows, last rows, first left columns, and first right columns) are excluded with a specific exclusion rate (ER).For example, when the ER is set to 20, the first 20% of the first rows and the last 20% of the eye image rows are excluded from the domain of the seed point scan.Also, the pixels which fall within 20% of the first left columns and 20% of the last right columns are omitted as depicted in Figure  The resulted image is converted to binary using the following formula: Where g(x, y) is the intensity value at location (x, y), and Binary(x, y) is the converted pixel value.Figure (3) depicts the converted binary image.

Step4: (Removal of Reflection Points)
As notice is that the MMU-1 database includes white points in the region of the pupil.To eliminate existing points of reflection, the region filling process is applied to the binary images to detect these points.
The following substeps are used to reduce reflex points: Step4.1:A is an input binary image and (A) its boundary.
Step4.2:Let X 0 = P // P is a point inside (A) (i.e., a non-boundary pixel) // where B is given a symmetric //structuring element respectively, when scanning horizontally by x = x p and scanning vertically by y = y p .The horizontal and vertical pupil radiuses R h and R v are then calculated as follows: ( ) (9) Figure (6) shows the computed four directions in pupil region.

Iris Localization
The next stage involves isolating the iris of the eye, and this stage included the following consecutive steps: Step1: (Eye Image Enhancement) In order to make the iris boundaries more apparent, initially it will be apply mapping process represented by histogram stretching on the original eye image, by applying the following substeps: Step1.1: Use stretchlim(I) which is returns the Low and High pixel values that specify lower and upper limits that can be used for contrast stretching eye image.
Step2.1: Adjust the contrast in the eye image using Low and High pixel values to set the limits.Figure (7) depicts the result of this step.

Step2: (Smoothing of Image)
The produced enhanced image is smoothed by using mean filter with widow size 9×9.The smoothed image S is depicted in Figure (8).

Step3: (Finding Threshold Value)
The threshold value is selected depending on the relationships between the characteristic of the original image and the image after thresholding with a small value to preserve the dark pixels only and then performed the following substeps: Step3.1: Compute the mean (m 1 ) and standard deviation (σ 1 ) of the original eye image.
Step3.2: Calculate the mean (m 2 ) and standard deviation (σ 2 ) of the eye the image after thresholding with a small value (i.e., 0.15).

Step4: (Multiply Image)
In this step, the smoothed image S has been multiplied by threshold value T 1 .Figure (9) shows the image after multiplying by T 1 = 0.4.

Step5: Convert to Binary
The Resulted Image is converted to binary using the following formula: Where M(x, y) is the intensity value of image M at location (x, y), and B (x, y) is the intensity value of converted image B at location (x, y).The TH is the threshold value which is an experimental value.Figure (10) shows the transformed image with TH=60.

Fig.10: Converted image by TH=60. Step6: Clean image
Removing all connected components (objects) that have fewer than P (i.e.,50) pixels from the binary image, producing another binary image.The default connectivity is 8.Then, erosion operation is performed.

Step7: Compute the Radius
Step7.1: Horizontal scan is applied from the centre of the pupil to the left and right until reaching the first white pixels from two directions, then set R l and R r to the number of pixels that are scanned in these directions.
Step7.2:In some cases, part of pupil is covered by the eyelashes and the arcs of the upper border of left or right eye continue with the iris area so that the scanning will go far away from the border of iris to overcome these cases, the following condition are executed: Step7.3:The iris radius is equal to:

Experimental Results
The proposed system was implemented in Matlab (R2015a) on a windows operating system with Core i5 PC of 2.67 GHz processor and 4 GB DRAM.The proposed iris segmentation method performance was tested using MMU-V1 and CASIA V-1 databases.The CASIA V-1 iris image database [20] includes 756 "non-ideal" images of 108 different users.Each image of the iris is grayscale with 320×280 pixels resolution.The MMU V-1 iris database [21] contains 450 images of 45 people captured using the LGIrisAccess®2200 camera.This semiautomatic camera operates on over 7-25 cm.The produced iris images are belong to 100 volunteers of various ages and nationality.For each eye image of a person has 5 iris images.And a resolution for each image is of 320×240 pixels in grayscale.Figure (11) shows examples of the two famous databases.Examples of the correct pupil and iris localization in which the proposed method yields high-accurate results in the existence of various noises which have been listed in Figure ( 12) and ( 13) respectively.The proposed iris localization algorithm was compared with the well-known iris segmentation algorithms introduced by Daugman [3], Wildes [4] and Libor Masek [9].Further to some of previous iris localization algorithms proposed by Liam and Chekima [22], Guang-zhu XU et al. [23], Weiqi Yuan et.al. [24], Muhammad et al. [25] and Suhad et.al. [26].Daugman's [3], Wildes's [4] is implemented in this work and the code provided by Libor Masek is used [27] to Libor Masek's algorithm [9]; while iris localization results of Chekima's, Guangzhu's, Weiqi Yuan's, Muhammad's, Suhad's algorithms were taken from the publishing works [28], [29], [30].Table 1 and Table 2 show the results of the proposed algorithm accuracy with compared to other algorithms.Where the segmentation accuracy is calculated according to the equation as below: (14) From the results listed in Table 1 and Table 2, it is noted that the proposed method based on image algebra and morphological operators has the best performance.The accuracy rate bellowing to MMU-1 is up to 98.78 %. (i.e., 99.33 for pupil localization and 98.22 for iris localization) and 100% belong to CASIA-1.The failure reason for the proposed pupil detection is that the eyelashes have interfered with the pupil area.Thus, the result of the consecutive addition will eliminate some part of the pupil when the intensity of the eyelashes are different from the pupil as shown in figure (14).The failure reason for the proposed iris detection due to the value of threshold T 1 which is obtained from eq. (10).For Tanwn left eye, the computed T 1 is 0.39 which is led to failing result of iris localization.When the threshold is adjusted to T 1 = 0.42, the iris is localized correctly as shown in Figure (15).Table 3 shows a time comparison among the proposed methodology and the three most commonly used algorithms for localising the iris, [3], [4] and [9].Because that Daugman, Wildes and Masek techniques required a comprehensive search that results in a very high cost of computation, So that, the proposed method employed the arithmetic operators which is very simple and therefore fast.

Conclusions
The image arithmetic operations are elementary and therefore very fast consequently these features are exhausted to design a quick method to localize iris.When the image is added to itself, the resulted image will be bright.Further, if the image continues with addition several times this lead to disappear all details of image accept the parts which are very dark such as pupil, eyelashes and eyebrow.The number of addition must be chosen carefully due to the eye images difference in its brightness.Therefore, the threshold value is selected based on Otsu global image threshold method.Another threshold was required in the proposed method for iris localization, which was selected depending on the statistical relationships between of the original image and the image after thresholding with a small value to preserve only the dark pixels.This value is useful for highlighting the iris area after multiplying the iris image by a threshold value which is wanted to identify.Most of the previous methods depended on techniques using the Hough Transform, active contour models or integro-differential operator.All these methods require a comprehensive search that results in a very high cost of computation.The proposed method was applied on CASIA V-1, and MMU V-1 iris image databases.The results showed that the proposed iris localization method outperforms over some state-of-the-art iris localization methods in both segmentation accuracy and computation time.