Linear representation of intra-class discriminant features for small-sample face recognition

The authors argue that the mean of discriminant features calculated across the samples of a class (intra-class samples) cannot perform well for the classification task. The main reason is that the mean feature ignores intra-class membership's different responses to their own class for a query sample. Meanwhile, they present that the discriminant features of a test sample can also be well-linear approximated by the discriminant features of intra-class memberships. The adaptive weighted intra-class features will be more suitable for the identification ability of a class than original samples via a regression algorithm. To verify this, a new linear representation-based classification method using Fisher discriminant features (LRFC) is suggested. To be more specific, they first extract Fisher discriminant features of all face samples. Second linear regression (LR) algorithm is exploited to obtain weight coefficients of intra-class feature information for the feature representation of a query sample. At last, the weighted intra-class features are re-combined as an agent of each class and the test sample is identified as the class with the maximum similarity. The method is simple but particularly effective. Experimental results on benchmark face databases verify improvements of LRFC over its original methods.