Novel Periocular Recognition under Non-cooperative Scenarios

Abstract

Components of face, such as iris and ocular region, possess discriminating patterns, making it suitable to be considered as a biometric trait to identify an individual. However, it becomes difficult to recognize a person with a low and degraded quality of images. Sometimes, it fails to authenticate when face is partially occluded (nose, mouth covered). In these situations, periocular region can be used as a biometric trait as opposed to iris which requires user co-operation as well as high resolution images. The periocular region can also be a useful biometric trait to authenticate an individual even if the low-resolution, defocused, off-angle, and only partial face or iris image is available. The thesis first provides an outline of existing ocular biometric datasets and summarizes a guideline on which datasets are useful or can be potentially used for study of periocular biometric. This is essential as there is no dedicated dataset for periocular study. Iris and face datasets are suitably used for periocular study. This thesis further investigates the performance of periocular based biometric system in the following two aspects: (a) Eliminating redundancy through a small and properly selected subset of features obtained using non-overlapping block division approach, and (b) Simulating real-time scenarios by modeling image quality covariates associated with blur, low-resolution and bit-depth on the performance of periocular recognition. The extraction of binned histogram features encoded by an interpolated local binary pattern in a blockwise manner is proposed in the first study. Matching is done using the Canberra distance measure. Experiment results show that the proposed method can reduce feature size without affecting its accuracy. In the second study, four types of image quality covariates such as out-of-focus blurred version (modelled through Gaussian function), camera shake blurred version (modelled through linear motion), low spatial resolution version (modelled through inter-area interpolation) and low bit-depth acquisition (modelled through bit plane slicing) are subjected to recognition and performance is evaluated. A deep learning architecture is used to train models and performance is evaluated using k-fold cross validation. Experimental results show that the performance of periocular recognition is primarily affected by out-of-focus blur. At the same time, it is more robust to camera shake blur. These approaches can be used to speed up the recognition process and evaluate other biometric systems’ performance when surveillance is to be done in unconstrained scenarios.