Iris Segmentation in the Wild Using Encoder-Decoder-Based Deep Learning Techniques

INTRODUCTION

In today’s world, where technology has invaded almost every walk of our life, it has become more important now than ever to safeguard confidential information. While passwords were predominantly used for the task, they have become a thing of the past as they are weak and can be cracked using simple-to-sophisticated techniques by accomplished hackers. In comparison, biometrics provides a much more reliable and challenging environment to crack alternative in comparison to passwords. Biometrics can be broadly categorised into two categories: physiological and behavioural. Physiological identification is based on traits such as face, hand geometry, and iris, whereas behavioural identification is made on the basis of characteristics such as signature and gait. According to Ref. [52], any human behavioural or physiological characteristic may be utilised as a biometric given that it satisfies some specific properties:

  • 1. Universality, i.e., it should be possessed by everyone.
  • 2. Uniqueness, i.e., it should be distinct amongst people.
  • 3. Permanence, i.e., it should remain the same throughout time.
  • 4. Collectability, i.e., it should be easily collectible and quantifiable.
  • 5. Performance, i.e., it should be efficient in the identification of the subject.
  • 6. Acceptability, i.e., it should be acceptable to the people in general.
  • 7. Circumvention, i.e., it should not be easy for the system using it to be fooled.

In Ref. [52], the authors have scored the modalities on the above criteria in majorly three categories: Low ([),), Medium (ф), and High (f[).

The survey has extended or represented differently in many studies such as [19,36,51,94,101]. Table 12.1 represents that although Iris doesn’t score ‘High’ in all the criteria, it is still by far the most suitable biometric modality. Owing to its accuracy and reliability [27], it is used in different biometric applications such as forensics [84] and intelligent unlocking [20].

TABLE 12.1

Comparison of Some Commonly Used Biometrics on the discussed Criteria as in Ref. [52]

Biometric

Univer.

Uniq.

Perm.

Collect.

Pert.

Accept.

Circum.

Iris

ft

ft

ft

ft

ft

ft

ft

Fingerprint

ft

ft

ft

ft

ft

ft

ft

Face

ft

ft

ft

ft

ft

ft

ft

Gait

ft

ft

ft

ft

ft

ft

ft

Hand Geometry

ft

ft

ft

ft

ft

ft

ft

Retinal scan

ft

ft

ft

ft

ft

ft

ft

Voice print

ft

ft

ft

ft

ft

ft

ft

Anatomy of human eye

FIGURE 12.1 Anatomy of human eye.

segmentation technique. The next stage is normalisation, wherein the segmented image is transformed into one with a pre-set dimension so as to maintain uniformity in all the segmented images, making it easier to act upon in the further stages. After normalisation, the next stage is feature extraction or biometric template generation. A biometric template is generated by using mathematical functions. The function is defined as such that the template represents the features in the best possible way for an efficient representation. In the last stage, the template is attempted to match with the already existing template of the subject, and authentication is given based on a certain predefined threshold of matching accuracy. Also, if the template is meant to be added into the database against a new subject, then it is added in the ‘enrolment’ mode of the system, while the former is done in the ‘identification’ mode of the biometric system.

However, that is not always the case, which may lead to the inclusion of various artefacts such as off-angle gaze, eyelash/eyelid occlusion, motion blur, and specular- reflections due to less user cooperation and non-ideal environments. All these will subsequently lead to poor segmentation results, and the error will be propagated and compounded when the information is passed through further in the system, ultimately leading to faulty results according to many studies [40,77,81,91]. However, if the segmentation can be done accurately enough, then only the relevant information, although augmented, can be propagated further, making the system more accurate than before. A majority of iris segmentation methodologies also assume the iris shape to be circular, which is deviated from w'hen the eye is partially closed, which further enhances the need for accurate segmentation techniques in non-ideal environments [26]. Moreover, with the rise in demand for the integration of biometric authentication into our daily lives, non-ideal conditions have to be factored in. In further sections, we discuss the various segmentation techniques which involve both ideal and non-ideal environments.

 
Source
< Prev   CONTENTS   Source   Next >