Face recognition based on local steerable feature and random subspace LDA

Both local features and holistic features are critical for face recognition and have different contributions. In this paper, we first propose a novel local steerable feature extracted from the face image using steerable filter for face representation. Discriminant information provided by steerable filter is locally stable with respect to scale, noise and brightness changes and it is semi-invariant under common image deformations and distinctive enough to provide useful identity information. We then present a new null space method based on random subspace. Linear Discriminant Analysis (LDA) is a popular holistic feature extraction technique for face recognition. Null Space LDA (NLDA) and Fisherface are adopted to extract global feature in the steerable feature space. Based on random subspaces, multiple NLDA classifiers are constructed under the most suitable situation for the null space. NLDA takes full advantage of the null space, while Fisherface extracts the most discriminant information in the principal subspace. Fisherface classifiers are constructed from the same set of random subspaces for NLDA classifiers. In each random subspace, Fisherface and NLDA share a unique eigen-analysis. There is no redundancy between such two kinds of complementary classifiers. Finally, all of the classifiers are integrated using a fusion rule. Experimental results on different face data sets demonstrate the effectiveness of the proposed method.