Gabor Orientation Histogram for Face Representation and Recognition

Abstract

Recently the Gabor-based features have been successfully used for face representation and recognition. In these methods, the face image is filtered with the multiscale multiorientation Gabor filter bank to generate multiple Gabor magnitude images (GMIs), and then the down-sampled GMIs or the LBP (local binary pattern) histograms of GMIs are stacked to form the feature. The stacking procedure makes the dimensions of these features very high, which causes extreme computing and storage load. In this paper, a novel Gabor-based feature termed Gabor orientation histogram (GOH) is proposed, which greatly reduces the feature dimension. Unlike stacking, GOH takes the structure underlying different GMIs into account by regarding the GMIs of different orientations at the same point as a whole, namely orientation vector, to represent the point. Moreover, GOH takes the structure of local region into account by calculating the orientation histogram based on the orientation vectors of points in the local region to describe the region, which is robust to local deformation and noises. The experimental results on the FERET and FRGC databases show that the proposed GOH reduces the feature extraction and recognition time significantly while retains the high recognition performance, which makes a progress toward the practical applications of Gabor-based features for face representation and recognition.