Multiple Occluded Face Detection Based on Binocular Saliency Map

In this paper, we propose a novel occluded face detection model which can detect multiple occlusions in a stereo image. The biologically motivated face preferable selective attention model localizes candidate regions for human faces in a natural scene, and then the Adaboost based face and eye detection process works for those localized candidate areas to check whether the areas contain a human face. In order to detect each facial area in multiple occluded faces, we use depth information of an eye, which is obtained by a binocular saliency map model. If the facial local features are similar depth information, we use the conventional Adaboost algorithm to localize the face area, and mask off the facial region. Then, we check the next area in a distance point of view whether it has an occlusion by Auto-associative multilayer perceptron (AAMLP) at 3 divided candidate regions. Finally, we implement an efficient model which can detect not only faces in a stereo image but also multiple occluded facial regions in an input scene. Experimental results show that the proposed model successfully localizes multiple occluded faces.