|Keywords:||Computer science; Artificial intelligence; Face verification; Facial analysis; Multilinear Analysis|
|Full text PDF:||http://www.escholarship.org/uc/item/7jq1x7np|
In the field of computer vision, multilinear (tensor) algebraic approaches to image-based face recognition have attracted interest in recent years. Previously, these methods have operated uniformly over the entire facial image at uniform resolution. In this thesis, we present a multiresolution, region-based multilinear method. By computing multiple multilinear models of various facial features, such as the eyes, nose, and mouth, in appropriate spatially-localized regions, we achieve a representation that, using the same amount of training data, is more discriminative for the purpose of facial verification. Adding a multiresolution image pyramid as well as a weighted signature further improves performance. We report encouraging experimental results on two datasets, one consisting of synthetic images, the other of real-world images.