Skip to Main Content
A discriminative face feature, i.e. multi-scale ICA texture pattern (MITP), is proposed for automatic gender recognition. First, independent component analysis (ICA) filters of various scales are learned using randomly collected face patches from training samples. Each face image is then encoded by sorting the responses of these filters. Finally, a histogram feature is formed based on the non-overlapping subregions of the encoded images. The newly proposed sparse classifiers are adopted for classification. Experiments on two benchmark face databases validate the effectiveness of MITP.