Histogram of Log-Gabor Magnitude Patterns for face recognition

The Gabor-based features have achieved excellent performances for face recognition on traditional face databases. However, on the recent LFW (Labeled Faces in the Wild) face database, Gabor-based features attract little attention due to their high computing complexity and feature dimension and poor performance. In this paper, we propose a Gabor-based feature termed Histogram of Gabor Magnitude Patterns (HGMP) which is very simple but effective. HGMP adopts the Bag-of-Words (BoW) image representation framework. It views the Gabor filters as codewords and the Gabor magnitudes of each point as the responses of the point to these codewords. Then the point is coded by the orientation normalization and scale non-maximum suppression of its magnitudes, which are efficient to compute. Moreover, the number of codewords is so small that the feature dimension of HGMP is very low. In addition, we analyze the advantages of log-Gabor filters to Gabor filters to serve as the codewords, and propose to replace Gabor filters with log-Gabor filters in HGMP, which produces the Histogram of Log-Gabor Magnitude Patterns (HLGMP) feature. The experimental results on LFW show that HLGMP outperforms HGMP and it achieves the state-of-the-art performance, although its computing complexity and feature dimension are very low.