Fusing multi-feature representation and PSO-Adaboost based feature selection for reliable frontal face detection

We propose a reliable frontal face detector based on multifeature descriptors and feature selection using PSO-Adaboost. Utilization of multiple heterogeneous feature descriptors enriches the diversity of feature types for face modeling and feature learning. To speed up the training process of face detector, we also propose a PSO-Adaboost algorithm that replaces exhaustive search used in original Adaboost framework with Particle Swarm Optimization (PSO) technique for efficient feature selection. Finally, a three-stage cascade classifier is developed to remove background rapidly. In particular, an initial stage is designed to detect candidate face regions more quickly by using a large size window with a large moving step. Radial Basis Function (RBF) SVM classifiers are used instead of decision stump functions in the last stage to remove those remaining complex non-face patterns that can not be rejected in the previous two stages. Combining these three effective modules, our face detector achieves a detection rate of 96.50% at ten false positives on the CMU+MIT frontal face dataset.