Learning Symmetry Features for Face Detection Based on Sparse Group Lasso

Face detection is of fundamental importance in face recognition, facial expression recognition and other face biometrics related applications. The core problem of face detection is to select a subset of features from massive local appearance descriptors such as Haar features and LBP. This paper proposes a two stage feature selection method for face detection. Firstly, feature representation of the symmetric characteristics of face pattern is formulated as a structured sparsity problem and sparse group lasso is used to select the most effective local features for face detection. Secondly, minimal redundancy maximal relevance is used to remove the redundant features in group sparsity learning. Experimental results demonstrate that the proposed feature selection method has better generalization ability than Adaboost and Lasso based feature selection methods for face detection problems.