Submitted to CVPR ' 99 Discriminant Analysis based Feature ExtractionW

We propose a new feature extraction scheme called Discriminant Component Analysis. The new scheme decomposes a signal into orthonormal bases such that for each base there is an eigenvalue representing the discriminatory power of projection in that direction. The bases and eigenvalues are obtained based on certain classiication criterion. For simplicity, a criterion used in Fisher's Discriminant Analysis (DA) is chosen and is applied iteratively to implement the scheme. We illustrate the motivation of this new scheme and show how it can be used to construct new distance metrics. We then argue that these new distance metrics are more robust than DA based metrics. Finally, very good classiication performance on simulation data and real face images are demonstrated using these new distance metrics.