This paper presents a novel ensemble of descriptors for face recognition derived from the base Patterns of the Oriented Edge Magnitudes (POEM) descriptor. Starting from different texture descriptors recently proposed in the literature, namely, the base patterns of POEM and the Monogenic Binary Coding (MBC), we develop different ensembles by varying the preprocessing techniques, the subspace projections, and some parameters of the system. Our approach is tested on the FERET datasets and the Labeled Faces in the Wild (LFW) dataset. Our system performs well on both datasets, obtaining, to the best of our knowledge, one of the highest performance rates published in the literature on the FERET datasets with an average accuracy of 97.3%. We want to stress that our ensemble obtains outstanding results in both datasets without any supervised approach or transform. The main findings of our proposed system include the following: 1) significant improvement in performance can be obtained by simply varying the parameters of stand-alone descriptors; and 2) performance can be improved by combining different enhancement and feature transform techniques.
Nanni, L., Lumini, A., Brahnam, S. (2017). Ensemble of texture descriptors for face recognition obtained by varying feature transforms and preprocessing approaches. APPLIED SOFT COMPUTING, 61, 8-16 [10.1016/j.asoc.2017.07.057].
Ensemble of texture descriptors for face recognition obtained by varying feature transforms and preprocessing approaches
Lumini, Alessandra;
2017
Abstract
This paper presents a novel ensemble of descriptors for face recognition derived from the base Patterns of the Oriented Edge Magnitudes (POEM) descriptor. Starting from different texture descriptors recently proposed in the literature, namely, the base patterns of POEM and the Monogenic Binary Coding (MBC), we develop different ensembles by varying the preprocessing techniques, the subspace projections, and some parameters of the system. Our approach is tested on the FERET datasets and the Labeled Faces in the Wild (LFW) dataset. Our system performs well on both datasets, obtaining, to the best of our knowledge, one of the highest performance rates published in the literature on the FERET datasets with an average accuracy of 97.3%. We want to stress that our ensemble obtains outstanding results in both datasets without any supervised approach or transform. The main findings of our proposed system include the following: 1) significant improvement in performance can be obtained by simply varying the parameters of stand-alone descriptors; and 2) performance can be improved by combining different enhancement and feature transform techniques.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.