Recently, public demands on content-based accessibility are increased. This thesis proposes a method to improve the accessibility by face summarization. Face summarization is the procedure to find one good face image for each person in an input image sequence. The proposed face summarization method includes two algorithms; face-group generation and key face selection.
A face-group for an image sequence is the set of face images from one person. We obtain face-groups by the following procedure. Firstly, based on spatio-temporal connectivity of skin color pixels, we group skin-color pixels according to objects identification. To group skin color pixels, we construct spatio-temporal skin color volumes and modify the skin color volumes on XT(YT)-slices of the spatio-temporal volume. Then, to solve occlusion problems, we separate the modified skin color volumes according to their shape, position, and motion. Finally, we decide that each separated volume contains face images. A face containing skin color volume makes up a face-group.
Key face is the most suitable face image to identify a person in a face group. We propose a method to select the large and frontal face as a key face.
Simulation results show that the proposed algorithm can summarize an image sequence by face images successfully even with frequent appearance or disappearance of people.
