The weighted submodule LDA for face recognition

Abstract

This paper presents an improved algorithm based on submodule Fisherface recognition (SM-Fisherface), which belongs to the most popular subspace algorithm currently. The method first divides the whole face data into multiple modules and analyzes each module separately, which can not only extract more facial feature data, but also capture more local face information and reduce the loss of information compared with no-module-dividing method. Second, we give weights to submodule data in two stages in order to center the sample data accurately in submodule treating stage and reduce the interference caused by external factors such as facial expressions and illumination image in integration stage. Then the recognition strategy for multi-module is proposed, which integrates multiple modules into a whole, and make use of the overall information to determine the final result. Numerical experiments with the YALE face database and FERET face database are given to compare the proposed method with PCA, Fisherface and unweighted SM-fisherface. The conclusion is that the weighted SM-fisherface method has the best recognition rate in the discussed experimental environment. © 2013 Springer-Verlag.