Feature selection for spatially enhanced LBP: application to face recognition

Block-based local binary patterns a.k.a. enhanced local binary patterns (ELBPs) have proven to be a highly discriminative descriptor for face recognition and image retrieval. Since this descriptor is mainly composed by histograms, little work (if any) has been done for selecting its relevant features (either the bins or the blocks). In this paper, we address feature selection for both the classic ELBP representation and the recently proposed color quaternionic LBP (QLBP). We introduce a filter method for the automatic weighting of attributes or blocks using an improved version of the margin-based iterative search Simba algorithm. This new improved version introduces two main modifications: (i) the hypothesis margin of a given instance is computed by taking into account the K-nearest neighboring examples within the same class as well as the K-nearest neighboring examples with a different label; (ii) the distances between samples and their nearest neighbors are computed using the weighted χ2 distance instead of the Euclidean one. This algorithm has been compared favorably with several competing feature selection algorithms including the Euclidean-based Simba as well as variance and Fisher score algorithms giving higher performances. The proposed method is useful for other descriptors that are formed by histograms. Experimental results show that the QLBP descriptor allows an improvement of the accuracy in discriminating faces compared with the ELBP. They also show that the obtained selection (attributes or blocks) can either improve recognition performance or maintain it with a significant reduction in the descriptor size. © 2017, Springer International Publishing AG, part of Springer Nature.