Reduced complexity face recognition using advanced correlation filters and Fourier subspace methods for biometric applications

This thesis presents recent research in the design of various Fourier based methods; ranging from Advanced Correlation Filters to Frequency Subspace Methods for face recognition in Biometric Applications. In particular we investigate how to perform reliable face verification for biometrics with particular focus on achieving illumination tolerance. Most Biometric applications can assume a partially co-operative user, therefore pose variations seen during test phase are assumed to be representative samples seen during the training phase. Illumination conditions of the subject can not be controlled as verification can occur anywhere where the user might not have control of their surrounding illumination conditions. Development of reduced complexity algorithms for performing face verification on limited resource computing platforms such as System-on-Chip implementations or PDA platforms are presented. This work shows that we can simplify our algorithms to produce biometric filters that only require 2 bits/per/frequency of storage, leading to templates requiring only 127 bytes of storage.; Spatial subspace methods are among the most common face recognition algorithms; however their draw back includes sensitivity to shifts of the test image, have difficulty or fail in the presence of extreme illumination and can not handle occlusions. While certain subspace algorithms are designed to handle one of these issues, they face difficulties trying to build tolerance to all of these distortions. This thesis present novel advancements in linking advanced correlation filters and subspace methods to provide hybrid PCA-correlation filters which are referred to as 'Corefaces'. These type of filters union the advantages of both worlds; efficient subspace representation of range of distortions and the shift-invariance, illumination tolerance and ability to handle partial test images featured by proposed advanced correlation filter approaches. Furthermore this research further examines how Principal Component Linear Subspaces are synthesized and presents reformulations that use Fourier transforms to achieve fast classification in a shift-invariant manner.; This thesis also details other proposed advanced correlation filters designs for increased efficiency towards building a real-time face verification system and addresses reduced complexity implementations of various parts of the system. We also address the issue of cancellability of biometric templates, proposing a novel encryption method where verification is performed directly in the encrypted domain while still preserving attractive properties such as linearity and shift invariance. Several new correlation filters are also proposed during the course of this research which offer merit improvement in speed, design or performance over current designs. While the focus of this thesis research is on face biometrics, the research and methods covered can be applied to any other biometrics such as fingerprint and iris.