Abstract: The original batch learning method contains redundant computation on the issue of on-line learning. To improve the learning efficiency, an incremental within-class locality preserving dimension reduction algorithm is proposed in this literature. This algorithm updates the projection matrix immediately after combing the advantages of a dimension reduction algorithm via QR decomposition and the local within-class features preservation kernel fisher discriminant algorithm. At the same time, this algorithm is effective on the data which is multimode or overlapping by considering the local structure and global distribution of the input samples. The experiments on ORL data set and COIL20 data set show this algorithm has a comparative dimension reduction performance with the batch method and a more fast speed.

Key words: On-line learning, Local features preservation, Dimension reduction