Color-based Lip Localization Method

This paper is concerned with lip localization for visual speech recognition (VSR) system. We shall present an efficient method for localization human’s lips/mouth in video images. This method is based on using the YCbCr approach to find at least any part of the lip as an initial step. Then we use all the available information about the segmented lip-pixels such as r, g, b, warped hue, etc. to segment the rest of the lip. The mean is calculated for each value, then for each pixel in ROI, Euclidian distance from the mean vector is calculated. Pixels with smaller distances are further clustered as lip pixels. Thus, the rest of the pixels in ROI will be clustered (to lip/non-lip pixel) depending on their distances from the mean vector of the initial segmented lip region. The method is evaluated on a new-recorded database of 780,000 frames; the experiments show that the method localizes the lips efficiently, with high level of accuracy (91.15%) that outperforms existing lip detection approaches.