Histograms of Gabor Ordinal Measures for face representation and recognition

This paper proposes a new image representation method named Histograms of Gabor Ordinal Measures (HOGOM) for robust face recognition. First, a novel texture descriptor, Gabor Ordinal Measures (GOM), is developed to inherit the advantages from Gabor features and Ordinal Measures. GOM applies Gabor filters of different orientations and scales on the face image and then computes Ordinal Measures over each Gabor magnitude response. Second, in order to obtain an effective and compact representation, the binary values of each GOM, for different orientations at a given scale, are encoded into a single decimal number and then spatial histograms of non-overlapping rectangular regions are computed. Finally, a nearest-neighbor classifier with the χ2 dissimilarity measure is used for classification. HOGOM has three principal advantages: 1) it succeeds the spatial locality and orientation selectivity from Gabor features; 2) the adopted region-comparison strategy makes it more robust; 3) by applying the binary codification and computing spatial histograms, it becomes more stable and efficient. Extensive experiments on the large-scale FERET database and AR database show the robustness of the proposed descriptor, achieving the state of the art.