Robust face detection from complex scene using multi-experts

In this paper, we propose a robust face detection approach by combining multiple experts in both cascade and parallel manner. We design three detection experts which employ different feature representation schemes of local images: 2D Haar wavelet, gradient direction, and Gabor filter. The three features are classified using the same classification model, namely, a polynomial neural network (PNN) on reduced feature subspace. The detection experts are used in multiple stages. At each stage, only when the output similarity of face exceeds a threshold, is the succeeding expert invoked to output a new similarity. To speed up detection, simpler (less time consuming) experts are used in preceding stages and complex experts are used in the succeeding stages. Meanwhile, the output of each expert is combined with the outputs of its preceding experts to improve detection accuracy. The effectiveness of the multi-expert approach has been demonstrated in experiments on a large number of images.