Selecting object of interest contours on variable background

Abstract

A method of selecting the boundary of the object of interest on a variable background, focused on the automatic analysis of human facial expressions, has been developed. The contours of brows, eyes and lips are extracted from the halftone photographic image. At the first stage, the image pixels are classified according to three brightness levels. It is assumed that the pixels belonging to one of them correspond to the skin, the other to the hair, while the remaining pixels make up the third level. To determine the brightness of these levels, an algorithm developed by the author earlier is used, the essence of which is as follows. The image is divided into rectangular fragments. For each of them, the hypothesis is tested that the brightness of its pixels is a sample from the distribution that describes the random noise present in the image (formulas were obtained for the normal noise distribution). Assuming that all pixels of any fragment for which the hypothesis was confirmed are included in only one of the specified groups, the brightness is determined by arithmetic averaging of the hypothetical mathematical expectations of the brightness of pixels belonging to each group. Then, the boundaries of areas with constant brightness are determined by means of line-by-line processing. The method was tested on images from an open database created at the University of Georgia (USA). The results showed that the extracted contours visually correspond well to the boundaries of the objects of interest and there is reason to believe that they can be used to solve the problem of automatic determination of the emotional state of a person from his portrait photo. In addition, the stability of the algorithm with respect to random noise was demonstrated, the value of the standard deviation of which does not exceed 25 % of the average value of the useful signal.