Representation and registration in face recognition and medical imaging

This dissertation presents solutions to four problems from face recognition and medical imaging. The first problem identifies an unknown face from a large database of facial images. The algorithm is based on matching pursuit filters, a small set of facial features, and simple geometric model. The set of features consists of the nose and eye regions of the face, and the interior of the face at a reduced scale. The algorithm uses coarse to fine processing to estimate the location of the facial features. Based on the hypothesized locations of the facial features the identification module searches the database for the identity of the unknown face. The identification is made by matching pursuit filters--a self-organizing technique for creating efficient and compact models from data. This technique is based on an adapted wavelet expansion, which is adapted to both the data and the goals of the algorithm. Thus, the filters can automatically find the subtle differences between facial features needed to identify unknown individuals. The algorithm is demonstrated on a database of photographs of 311 individuals and on a database of infrared facial images. The second problem adjusts for illumination differences between two facial images. The algorithm transforms the histogram of pixel values on one face to the histogram of another face. The algorithm, which is computationally efficient, nonlinear, and data-driven, corrects for variations between two different facial images or changes within an image of a face. The third problem uses a sieve algorithm to find the correspondence between pairs of images taken with an electron microscope. A sieve algorithm uses a sequence of approximations to generate increasingly accurate estimates of the correspondence. Initially, the approximations are computationally inexpensive, and at later stages both accuracy and complexity increase. The fourth problem presents an automatic registration algorithm for MR and PET slices of the brain that does not require manual intervention. The algorithm takes an integrated approach and simultaneous segments the brain in both modalities and registers the slices. A sequence of templates from the PET slice is constructed and registered in the MR slice using an energy function. The template with minimum energy gives the final registration.