Date of Graduation
Statler College of Engineering and Mineral Resources
Lane Department of Computer Science and Electrical Engineering
Current face recognition systems capture faces of cooperative individuals in controlled environment as part of the face recognition process. It is therefore possible to control lighting, pose, background, and quality of images. However, in a real world application, we have to deal with both ideal and imperfect data. Performance of current face recognition systems is affected for such non-ideal and challenging cases. This research focuses on designing algorithms to mitigate the effect of covariates in face recognition.;To address the challenge of facial aging, an age transformation algorithm is proposed that registers two face images and minimizes the aging variations. Unlike the conventional method, the gallery face image is transformed with respect to the probe face image and facial features are extracted from the registered gallery and probe face images. The variations due to disguises cause change in visual perception, alter actual data, make pertinent facial information disappear, mask features to varying degrees, or introduce extraneous artifacts in the face image. To recognize face images with variations due to age progression and disguises, a granular face verification approach is designed which uses dynamic feed-forward neural architecture to extract 2D log polar Gabor phase features at different granularity levels. The granular levels provide non-disjoint spatial information which is combined using the proposed likelihood ratio based Support Vector Machine match score fusion algorithm. The face verification algorithm is validated using five face databases including the Notre Dame face database, FG-Net face database and three disguise face databases.;The information in visible spectrum images is compromised due to improper illumination whereas infrared images provide invariance to illumination and expression. A multispectral face image fusion algorithm is proposed to address the variations in illumination. The Support Vector Machine based image fusion algorithm learns the properties of the multispectral face images at different resolution and granularity levels to determine optimal information and combines them to generate a fused image. Experiments on the Equinox and Notre Dame multispectral face databases show that the proposed algorithm outperforms existing algorithms. We next propose a face mosaicing algorithm to address the challenge due to pose variations. The mosaicing algorithm generates a composite face image during enrollment using the evidence provided by frontal and semiprofile face images of an individual. Face mosaicing obviates the need to store multiple face templates representing multiple poses of a users face image. Experiments conducted on three different databases indicate that face mosaicing offers significant benefits by accounting for the pose variations that are commonly observed in face images.;Finally, the concept of online learning is introduced to address the problem of classifier re-training and update. A learning scheme for Support Vector Machine is designed to train the classifier in online mode. This enables the classifier to update the decision hyperplane in order to account for the newly enrolled subjects. On a heterogeneous near infrared face database, the case study using Principal Component Analysis and C2 feature algorithms shows that the proposed online classifier significantly improves the verification performance both in terms of accuracy and computational time.
Singh, Richa, "Mitigating the effect of covariates in face recognition" (2008). Graduate Theses, Dissertations, and Problem Reports. 4420.