Extraction of faces of more than one person from natural background for personal identification

Abstract

In general, it is very difficult to extract human faces from a natural background. In this paper, the extracting method of faces from more than one person's facial area with a natural background is presented. The processes are executed by using human skin color and moving information, and consist of the following five stages: (1) detection of moving area; (2) extraction for skin-color areas; (3) evaluation of movement of each skin-color area; (4) extraction of facial areas from each skin-color area when movement is large; and (5) the extraction of facial areas from each skin-color area when movement is small. The moving area is first detected using the difference between two successive frames. The skin-color areas are then extracted based on the modified HSV color system from an input image. The judgment whether the movement of each skin-color area is large or not is done. The extracting process of the facial area is divided into two processes according to the moving quantity: (1) when the movement is large, the facial areas containing a neck area are extracted based on its motion information; and (2) when the movement is small, since the shape of the facial areas does not change between the successive frames, the facial areas are then clipped out based on the facial area that has already been extracted in the previous frame. The skin-color area corresponding to the facial area which is extracted from the previous frame is detected. Computer simulation using 1563 frames shows more than 89% extracting accuracy. From the experimental results, the prospects of using this method for the human face identification, image coding or the recognition of facial expressions are very encouraging.