Local IMLDA and its Improvements

Two-dimensional (2D) discrimination analysis using methods such as 2DPCA and IMLDA is more effective in face recognition than one-dimensional (1D) discrimination analysis. The methods before often perform on all classes and do not consider that some training samples not only have no contribution to the classification of one given test sample but also make it hard to classify the test sample correctly. Therefore, a method called local IMLDA (L-IMLDA) which is based on IMLDA is proposed in this paper. This method performs IMLDA one time for every test sample on several selected classes to which the test sample is belong. In addition, we have improved L-IMLDA and then proposed another two approaches IL-IMLDA and IMLDA+ IL-IMLDA. Experiment on ORL database shows that those approaches perform better than IMLDA and 2DPCA.