Biometric fusion methods for adaptive face recognition in computer vision

Abstract

Face recognition is a biometric method that uses different techniques to identify the individuals based on the facial information received from digital image data. The system of face recognition is widely used for security purposes, which has challenging problems. The solutions to some of the most important challenges are proposed in this study. The aim of this thesis is to investigate face recognition across pose problem based on the image parameters of camera calibration. In this thesis, three novel methods have been derived to address the challenges of face recognition and offer solutions to infer the camera parameters from images using a geomtric approach based on perspective projection. The following techniques were used: camera calibration CMT and Face Quadtree Decomposition (FQD), in order to develop the face camera measurement technique (FCMT) for human facial recognition. Facial information from a feature extraction and identity-matching algorithm has been created. The success and efficacy of the proposed algorithm are analysed in terms of robustness to noise, the accuracy of distance measurement, and face recognition. To overcome the intrinsic and extrinsic parameters of camera calibration parameters, a novel technique has been developed based on perspective projection, which uses different geometrical shapes to calibrate the camera. The parameters used in novel measurement technique CMT that enables the system to infer the real distance for regular and irregular objects from the 2-D images. The proposed system of CMT feeds into FQD to measure the distance between the facial points. Quadtree decomposition enhances the representation of edges and other singularities along curves of the face, and thus improves directional features from face detection across face pose. The proposed FCMT system is the new combination of CMT and FQD to recognise the faces in the various pose. The theoretical foundation of the proposed solutions has been thoroughly developed and discussed in detail. The results show that the proposed algorithms outperform existing algorithms in face recognition, with a 2.5% improvement in main error recognition rate compared with recent studies.