Unsupervised learning for modulated view-based face detection system

In order to realize an unsupervised learning for view-based face detection system, we introduce an architecture with separation of clustering features from detecting features. In this architecture, the number of detectors for individual "view" of faces is determined by the result of clustering autonomously. For autonomous clustering, the combination of Kohonen's self-organizing feature map (SOM) and a novel cluster determination algorithm is introduced. This cluster determination algorithm allows us to find gaps between clusters by comparing the distance from a particular data to adjacent clusters. Therefore, it can define the similarity of face "view" to represent the view-based features in case of face detection. For feature detection, a non-linear subspace model based on kernel method is used. Our architecture self-organizes feature detectors corresponding to face "view"; the face detection system shows good face detection performance under a wide variety of lighting conditions.