Robust face detection under partial occlusion

In this paper, a robust face detection method under partial occlusion is proposed. In recent years, the effectiveness of face detection methods using support vector machines (SVM) has been reported, but in conventional algorithms, one kernel is applied to global features extracted from an image. Global features are easily influenced by partial occlusion, and therefore the conventional algorithms appear not to be robust in the presence of occlusion. Good handling of local features is necessary in order to provide robustness to partial occlusion in face detection methods based on SVM. We introduce a local kernel for good handling of local features in SVM and use summation as the integration method. In the experiment, a comparison was made with SVM based on the conventional global kernel and using face images including occlusions and face images including shadows caused by changes in the direction of the light source. The robustness of the proposed method to occlusion was demonstrated. It was also confirmed that faces could be detected from face images including practical occlusions such as sunglasses or scarves. © 2007 Wiley Periodicals, Inc.