How far we can improve micro features based face recognition systems?

摘要

This paper presents improvements for face recognition methods that use LBP descriptor as a main technique in encoding micro features of face images. Our improvements are focused on the feature extraction and dimension reduction steps. In feature extraction, we use a variant of Local Binary Pattern (LBP) so-called Elliptical Local Binary Pattern (ELBP), which is more efficient than LBP for extracting micro facial features of the human face. ELBP of one pixel is built by thresholding its gray value with its P neighboring pixels on a horizontal ellipse. ELBP operator is applied in Pattern of Oriented Edge Magnitudes (POEM) to build Elliptical POEM (EPOEM) descriptor. The dimension reduction step is conducted by using Singular Value Decomposition (SVD) based Whitened Principal Component Analysis (WPCA). For performance evaluation of our improvements, we compare them with LBP based, POEM based approaches and other popular face recognition systems. The experimental results on state-of-the-art FERET and AR face databases prove the advantages and effectiveness of our improvements.