Jennifer Nealy, University of Central Florida


Object Classification using Local Subspace Projection

Abstract: We consider the problem of object classification from image data. Significant challenges are presented when objects can be imaged from different view angles and have different distortions. For example, a vehicle will appear completely different depending on the viewing angle of the sensor but must still be classified as the same vehicle. In regards to face recognition, a person may have a variety of facial expressions and a pattern recognition algorithm would need to account for these distortions. Traditional algorithms such as PCA filters are linear in nature and cannot account for the underlying non-linear structure which characterizes an object. We examine nonlinear manifold techniques applied to the pattern recognition problem. One mathematical construct receiving significant research attention is diffusion maps, whereby the underlying training data are remapped so that Euclidean distance in the mapped data is equivalent to the manifold distance of the original dataset. This technique has been used successfully for applications such as data organization, noise filtering, and anomaly detection with only limited experiments with object classification. For very large datasets (size N), pattern classification with diffusion maps becomes rather onerous as there is a requirement for the eigenvectors of an NxN matrix. We characterize the performance of a 40 person facial recognition problem with standard K-NN classifier, a diffusion distance classifier, and standard PCA. We then develop a local subspace projection algorithm which approximates the diffusion distance without the prohibitive computations and show comparable classification performance.

Mentor: Dr. Robert Muise (University of Central Florida).