A Weighted Modular Principal Component Analysis Approach for Face Authentication

A weighted modular approach for face authentication based on the priorities of different facial regions that change with varying poses, expressions and occlusions is presented in this thesis. This helps in verifying the identity of an individual who claims to be a subject in the database and is unaware of the presence of the face authentication system. A sequence of face images is selected from a video in a particular predefined interval and is used for verification. The face images are divided into different horizontal modules based on the regions representing facial features. A principal component analysis on these modules produces low dimensional representations of the sub images representing the facial feature regions. A weighted comparison of feature regions of the test images with the respective regions of the training images provides the authentication outcome. The performance of the proposed face authentication system is evaluated with several individuals belonging to different ethnicities. It is observed that the weighted modular approach outperforms the state-of-the-art face authentication techniques for input images with varying poses and expressions and occlusions. Research studies are progressing to compute the weights adaptively based on the magnitude of the feature variations due to poses and expressions and occlusions. In addition, application of the techniques based on multiple modalities for face authentication is also being investigated.