Accurate face detection by combining multiple classifiers using locally assembled histograms of oriented gradients

Levi et al. [1] have introduced edge orientation histograms (EOH) into accurate face detection, and proved EOH feature to be very discriminative. However, EO-H captures too little spatial information. We propose a novel feature, which uses the same oriented histogram as EOH but contains more spatial information. It is called Locally Assembled Histogram (LAH) of Oriented Gradients. Several neighboring HOG features [2] are assembled to capture their co-occurrence. Then the feature vector is projected into a scalar by Fisher Linear Discrimination. Furthermore, several classifiers are combined during testing to improve the detection rate. One classifier is selected adaptively according to the sliding window size and the training face size. Experiments on CMU+MIT data set demonstrate that our system is better than some well known systems, such as those using EOH [1], Haar [3] or LBP [4].