Feature extraction using class-oriented regression embedding

Based on linear regression techniques, we present a new supervised learning algorithm called Class-oriented Regression Embedding (CRE) for feature extraction. By minimizing the intra-class reconstruction error, CRE finds a low-dimensional subspace in which samples can be best represented as a combination of their intra-class samples. This characteristic can significantly strengthen the performance of the newly proposed classifier called linear regression-based classification (LRC). The experimental results on the extended-YALE Face Database B (YaleB) and CENPARMI handwritten numeral database show the effectiveness and robustness of CRE plus LRC.