ROI processing for visual features extraction in lip-reading

Region of interest (ROI) is the key basis of visual features extraction in lip-reading process. In this paper, we discussed the ROI processing method and explored its impact on recognition accuracy with the comparison of four kinds of processed ROIs obtained by using four basic image processing methods: gray-scale normalization, difference enhancement, edge enhancement and image segmentation. Then recognition tasks for speaker-independent lip-reading were carried out by the aid of continuous hidden Markov model (CHMM). The experimental results show that for discrete cosine transform (DCT) based features, normalized gray-scale image can achieve the best recognition performance among these four ROIs.