Selecting Optimal Orientations of Gabor Wavelet Filters for Facial Image Analysis

Gabor wavelet-based methods have been widely used to extract representative features for face analysis. However, the existing methods usually suffer from high computational complexity of Gabor wavelet transform (GWT), and the Gabor parameters are fixed to a few conventional values which are assumed to be the best choice. In this paper we show that, for some facial analysis applications, the conventional GWT could be simplified by selecting the most discriminating Gabor orientations. In the selection process, we analyze the histogram of oriented gradient (HOG) of the average face image in a dataset, and eliminate the less significant orientation combinations. Then we traverse the rest combinations and select the best according to classification performance. We find that the selected orientations match the analysis of HOG well, and are therefore consistent with the intrinsic gradient characteristics of human face images. In order to assess the performance of the selected Gabor filters, we apply the proposed method to two tasks: face recognition and gender classification. The experimental results show that our method improves the accuracy of the classifiers and reduces the computation cost.