Sparse Similarity-Based Fisherfaces

In this work, the effect of introducing Sparse Principal Component Analysis within the Similarity-based Fisherfaces algorithm is examined. The technique aims at mimicking the human ability to discriminate faces by projecting the faces in a highly discriminative and easy interpretative way. Pixel intensities are used by Sparse Principal Component Analysis and Fisher Linear Discriminant Analysis to assign a one dimensional subspace projection to each person belonging to a reference data set. Experimental results performed in the AR dataset show that Similarity-based Fisherfaces in a sparse version can obtain the same recognition results as the technique in a dense version using only a fraction of the input data. Furthermore, the presented results suggest that using SPCA in the technique offers robustness to occlusions.