A two-stage dimensional reduction approach to low-dimensional representation of facial images

We present a two-stage dimensional reduction approach to low-dimensional representation. When facial feature data need to be stored in low capacity storing devices, low-dimensional representation of facial images is very important. Our approach is composed of two consecutive mappings of the input data. The first mapping is concerned with best separation of the input data into classes and the second focuses on the mapping that the distance relationship between data points before and after the map is kept as closely as possible. We claim that if data is well-clustered into classes, features extracted from a topology-preserving map of the data are appropriate for recognition when low-dimensional features are to be used. We have presented two novel methods: FLD (Fisher's Linear Discriminant) combined with SOFM (Self-Organizing Feature Map) method and FLD combined with MDS (Multi-Dimensional Scaling) method. Experimental results using Yale, AT&T and FERET facial image databases show that the recognition performance of our methods degrades gracefully when low-dimensional features are used.