Lip contour extraction scheme based on K-means clustering in different color planes

In this paper, an efficient lip contour extraction scheme is proposed. From a given video frame, first the mouth region is extracted by using pixel threshold based binary conversion and detection of largest object in that region. Analysing the variation of pixel intensity pattern of RGB and a weighted colour plane, intensity ratio based lip region detection is performed, which provides accurate estimate of separate upper and lower lip regions. At the first stage, k-means classification is employed after obtaining feature value from green colour plane for the upper part of the lip region. In the second stage, the weighted colour plane with excellent lip distinguishing capability along with k-means classification is employed to detect the lower lip region. Because of critical shape of the upper lip, a piecewise curve fitting process is employed, where smoothing operation is performed on the outer contour by taking four consecutive pixels and fitting them in polynomials of certain order. From extensive experimentation on several real-life images from audio-visual clips, it is found that the proposed method offers high level of accuracy in image.