Face recognition using Weber local descriptors

This paper presents a method for face recognition using multi-scale Weber local descriptors (WLDs) and multi-level information fusion. Our method introduces the WLD, a novel and robust local descriptor, to describe the facial images and modifies it by a non-linear quantization approach to enhance its discriminative power. Moreover, a multi-scale framework for WLD extraction with multi-level information fusion approaches is provided for face representation and recognition. The proposed method has four main steps: (1) image partition: under given rules, each facial image is uniformly divided into a set of non-overlapped sub-regions; in this way, for a set of facial images, we therefore have a large pool of this type of sub-regions; (2) feature extraction: in this pool of sub-regions, taking one sub-region as a center, a group of similar ones are chosen for extraction of WLD histogram features; (3) features measurement: these WLD histograms are then fused into a single vector – as the feature of the center sub-region. Nearest neighborhood on chi-square is employed for similarity measurement between two sub-regions; and (4) voting: the recognition result of the entire probe (a face in sub-regions) is obtained via a voting function on the recognition result of all its sub-regions. Experimental results demonstrate the effectiveness of the proposed method upon three popular datasets.