Improving LBP features for gender classification

Automatic gender classification aims at analyzing the face image to recognize gender with computer, in which feature extraction is one key step. The LBP (local binary pattern) feature has essential applications in face analysis and has been applied in gender recognition. The normally adopted LBP feature will encounter dimension explosion with the increase of sampling density of LBP operator, which could not remarkably improve the performance of gender classification. In this paper, we present two simple methods to improve the common LBP feature, i.e., fusing low-density LBP features and decreasing the dimension of high density LBP feature with PCA (principle component analysis), both of which could drastically lower the feature dimension while preserving the precision. Experiments are performed on FERET upright face database. The results illustrate the drawbacks of general LBP feature and identify the merit of our improved feature extraction algorithms.