| Summary: | 碩士 === 大同大學 === 資訊工程研究所 === 91 === Because the inputs of the speech recognition are voice signal, when the signal is noisy, the speech recognition is influenced by noise. However, the input images of the lip-reading recognition are not influenced by this noise. In addition, when someone pronounces the nasal sounds, such as /m/、/n/、/ng/, the voice information is difficult to distinguish. But when we pronounce these voices, the changes of the mouth shapes are different. In this thesis, we proposed a lip-reading recognition system using motion vectors.
Our approach uses the motion of lips as the feature. We extract predetermined points of lips, and uses block matching approach to find motion vectors. According to the motion variations of those selected points, we can perform the lip-reading recognition. Experimental results show that the propose method using motion vectors is useful and the lip-reading recognition is efficient.
|
|---|
