Low-Dimensional Facial Image Representation Using FLD and MDS

Abstract

We present a technique for low-dimensional representation of facial images that achieve graceful degradation of recognition performance. We have observed that if data is well-clustered into classes, features extracted from a topologically continuous transformation of the data are appropriate for recognition when low-dimensional features are to be used. Based on this idea, our technique is composed of two consecutive transformations of the input data. The first transformation is concerned with best separation of the input data into classes and the second focuses on the transformation that the distance relationship between data points before and after the transformation is kept as closely as possible. We employ FLD (Linear Discriminant Analysis) for the first transformation, and classical MDS (Multi-Dimensional Scaling) for the second transformation. We also present a nonlinear extension of the MDS by ‘kernel trick’. We have evaluated the recognition performance of our algorithms: FLD combined with MDS and FLD combined with kernel MDS. Experimental results using FERET facial image database show that the recognition performances degrade gracefully when low-dimensional features are used.