Feature extraction using supervised constrained maximum variance mapping

Constrained maximum variance mapping (CMVM) based on the multi-manifold learning is an efficiency method for feature extraction. CMVM preserves the local manifold structure by keep the sum of the distances of samples unchanged, but ignores the local label information of the samples, which is very important to the recognition. To tackle the shortage, we propose a new method called supervised constrained maximum variance mapping (SCMVM), which projects the local structure into feature space by a linear map. SCMVM combines the Euclidean distance with the label information in local structure and maximizing the distance of samples with different classes. Because consider the local label information, the efficiency of recognition enhances clearly. In this paper, we take experiments on Yale face database and USPS handwriting database using CMVM and SCMVM, and compare the efficiency.