Ensemble Of Patterns Of Oriented Edge Magnitudes Descriptors For Face Recognition

In this work we propose an ensemble of descriptors for face recognition. Starting from the base patterns of the oriented edge magnitudes (POEM) descriptor, we developed different ensembles by varying the preprocessing techniques, the parameters for extracting the accumulated magnitude images (AM), and the parameters of the local binary patterns (LBP) applied to AM. Our best proposed ensemble works well regardless of whether dimensionality reduction is performed or not by principal component analysis (PCA) before the matching step. To validate our results we test our approach using the FERET datasets and the Labeled Faces in the Wild (LFW) dataset, and we obtain a very high performance in both datasets. The results are particularly interesting in the FERET datasets, where, to the best of our knowledge, we obtain one of the highest performances reported in the literature. We want to stress that our ensemble obtains outstanding results in both datasets without combining different texture descriptors and without any supervised approach or transform. The main novelties of the proposed work are the following: 1) our experiments show that it is possible to improve considerably a stand-alone descriptor by changing its parameters; 2) we also show that another easy way to boost the performance of a pattern recognition system is to use different enhancement techniques, and 3) some variants of the base POEM are proposed (e.g., using different descriptors applied to AM or to filter the image by Gabor filters before the AM extraction) and are shown to enhance performance. The MATLAB source of our best approach will be freely available.