RELATED WORK

In these subsections, we shall discuss existing methods for the iris segmentation. We cover both classical image processing techniques where algorithms rely on predefined rules and in-depth learning-based solutions that exploit the availability of massive paired databases for iris segmentation. There is no doubt that the results of deep learning-based methodologies surpass those of classical approaches. Recently, data-driven deep learning approaches have proven to give exceptional results in the field of biometrics and beyond it. We provide a comprehensive and comparative study amongst the methods.

Non-Deep Learning-Based Methodologies

In the iris segmentation, the steps followed in general involve as follows: first, the extraction of the Region of Interest (ROI) from the complete image, followed by the approximation of two circles which separate the iris region from the pupil and the sclera [107]. In this section, we have tried to cover all categories of the iris segmentation and boundary detection. Classical approaches take advantage of either the pixel-based features or the boundary-based features [13,59]. The first noticeable work was done by Daugman [28] back in 1993; his work formed the basis of all the work thereafter. In the eye, pupils along with iris are taken as non-concentric circles; an integrodifferential operator localises the boundary to segment out the iris. The method avoided the use of images, which included any type of occlusions like eyelids, eyelashes, and reflection, etc. Overall, we can say that classical methods followed by Daugman focused on extracting the edges of the pupil and iris to localise the to-be segmented area more precisely. Authors relied on the rule-based algorithmic approaches, which in some sense limited their methods to work on vast variations of images of the iris. We can say that those methods could not perform well on non-ideal images of the iris. Realizing the limitation of methods and instead of using simple edge-detection steps, they started to use more statistical approaches [48]. With the advancements in research, researchers started to model the anatomy of the eye realistically, such as taking the boundaries of the pupil and iris as non-circular. While some of the work did handle the problems of obstruction, specular reflection, and eyelash, etc., their limited feature modelling and extraction approaches were not too vast to cover all the variations in the iris images. In this section, we dive deep into the classical work done so far and provide a comparative analysis.

After Daugman, Wilde [106] in 1997 proposed a new approach in which he used an LED point source in addition to a camera for capturing eye images. He identified iris boundaries by gradient-dependent binary edge map in addition to the circular Hough transform. The paper also presented an in-depth comparative study with Daugman’s work. While Wilde’s work is considerably complex than Daugman’s, the segmentation approach proposed by Wilde was better as it detected the eyelids as well as it worked better with noisy images. In Ref. [15], Boles et al. proposed a circular edge-detection method (Figure 12.4).

The authors of [93] proposed an iris localisation technique, namely, circular sector analysis (CSA), before applying rough entropy for segmentation. Their localisation methods decreased the overall uncertainty in the segmentation mask. Another

Some images and their corresponding groud-truth segmentation masks from the UBIRIS-v2 data set [79]

FIGURE 12.4 Some images and their corresponding groud-truth segmentation masks from the UBIRIS-v2 data set [79].

work [107] proposed the iris localisation by assuming that shapes of the pupil and iris are circular wherein they first localised the pupil through eccentricity-dependent bisection approach, and then for iris, a region totally free from noise was obtained with directional segmentation followed by obtaining the gradients of direction lines to localise iris.

It was not until 2001 when Kong and Zhang in their work [57] incorporated the noisy and occluded images for iris segmentation. They used Hough transform to isolate the iris followed by 1-d Gabor filters for eyelids detection and thresholding to identify specular reflection. Their work gave better results for segmentation as well as in the final recognition task. Lim et al. [62] segmented the iris images by the edge- detection method through finding virtual circles where the pupil was detected first by the centrepoint-detection method. They acquired eye images but from a distance, and to reduce the reflections, they used halogen lamps. Their data set consisted of both eyes, with and without lens and glasses. Daugman, in his work [25], proposed the algorithm where he detected eyelid occlusion while segmenting the iris. Huang et al. [46] applied a median filter prior to canny operator for edge detection. Outer boundary was detected using a voting scheme on the maximum circle, and similarly for an inner boundary, it was identified using a rectangular inter interval. Localised iris was then segmented with the help of an integrodifferential operator. They too handled the eyelid occlusion using thresholding of histogram-based Hough transform. Huang et al. [45] again proposed a novel segmentation technique that also eliminated the noise to improve the results. They localised the iris using a simple filtering step with edge detection and Hough transform; occlusion factors were then eliminated using a Gabor filter.

Dorairaj et al. [32] developed an approach to deal with the off-angle iris image. In this work, he used PGA and global ICA for the encoding of off-angle iris images; while applying PCA/ICA, they first estimated the gazing angle by using Hamming distance followed by a simple integrodifferential operator for segmentation. Daugman in Ref. [26] developed an algorithm to tackle off-angle images similar to that of Ref. [32] with the elimination of occlusion caused by eyelashes. Abiyev et al., in their work [5], came up with the neural network-based method for the iris recognition; a rectangular area of size 10x10 was used to identify the pupil region. For the removal of noise, they utilised the standard linear Hough transform for eyelids.

The authors of Ref. [48] proposed a multi-stage technique. First, a moving window of circular shape was used for the pupil estimation, following which the estimation of the pupil was done through the standard-deviation peaks in both x as well as у directions, and after that, a median-filter reduced the eyelash effects. In Ref. [3], the authors proposed AdaBoost for eye detection for further segmentation. Reference [80] presents an unsupervised approach where images were modelled as Markov random field. Graph-cut method extracted the texture region, and for the iris segmentation image, intensities were exploited. Roy et al. [88] proposed a non-ideal iris recognition method, in which they used a Mumford-Shah segmentation method. All these classical approaches claim to handle various noises, distortion, and non-ideal iris images, but all being rule-based feature-driven approaches are limited in handling the variation of a non-ideal iris image. In Table 12.2, some classical approaches are compared based on their novelty and performance.

Methods

Novelty

Strength

Performance

Data set Used

Sardar et al. [93]

CSA and use of rough entropy for Iris localisation

Less computational expensive

El error rate: 0.08%, Acc.: 97.12%

IITD, MMU, CASIA

Parikh et al. [69]

Colour clustering along with curve fitting

Eyelids detected reduces the chances of error in performance

Acc.: 93.3% (UBIRIS), 94.9% (NICE-II), 92.8% (NICE-I)

NICE-I. NICE-II and UBIRIS v2

Khan et al. [5]

Gradient-based approach for localisation of iris with sclera boundary points

Iris borders were identified using novel gradient approach

Acc.: 98.22%(MMU), I00%(CASIA)

MMU, CASIA

Pundlik et al. [80]

Graph-cut approach for iris segmentation

Markov random field removes the eyelashes to boost the performance

None

None

Ibrahim et al. [48]

A two-stage hierarchical approach with moving circular window for iris segmentation

Intensity of pupil makes it easier to separate using probability approximation

Acc.: 98.28% (CASIA IrisV3 Lamp), 99.90% (CASIA IrisVI), 99.77% (MMU)

MMU, CASIA

Hu et al. [43]

Fusion of three models to extract the iris with the help of Daugman’s method

Integral derivative to detect the iris boundary is time-efficient

El error rate: 1.75% (MICHE), 1.30% (UBIRIS v2)

MICHE, UBIRIS v2

Huang et al. [4]

Radial suppression after the thresholding of detected edges

For ideal cases of iris images with the assumption of circular iris boundary, radial suppression gives good performance

El error rate: 0.32%

CASIA

Abate et al. [67]

Novel watershed method in addition to seed selection

Sclera along with eyelash/eyelid are separated using limbus detection approach

None

MICHE

Jeong et al. [3]

Eye detection with Adaboost and moving edge detector of shape circular

Detecting the obstructions in the iris images decreases the chances of error in segmentation map

El, E2 error rates: 2.8%, 14.4%, respectively

UBIRIS.V2

Ibrahim et al. [49]

First derivative-based iris detection with adaptive thresholding

Boundary detection takes lesser time

Acc.: 99.13% (MMU)

MMU vl.O

Patel et al. [70]

Binary-integrated intensity curve with region growing approach

Detection of eyelash and eyelid increases the performance of segmentation task

El error rate: 5.14%

CASIA

 
Source
< Prev   CONTENTS   Source   Next >