Abstract: A video coder automatically identifies the head area from a head against a background scene.Identification involves: storing several frames of an image training sequence (1); taking the difference between adjacent frames (2); for all the picture elements (pels) represented by each set of difference data, thresholding the pels against a luminance threshold (3); applying median filtering to the data sets thus obtained (4); then clustering the data sets (5); determining the minimum rectangle which will contain all the remaining non-zero pels represented by each set (6); generating a border of finite thickness around the rectangles (7); generating a head codebook from the pels of the original frames of the image that correspond to the pels represented by the respective sets of data that fall within the respective borders (8), similarly generating a background codebook; and finally generating a composite codebook.