Authentication Analysis Of Features Based On Lip-Reading Recognition

Nowadays , there emerge different kinds of audio-recognition system,which have quite high recognition performance.However, audio-recognition system only uses the audio information of speech,and ignores the visual features of it.Under the noise circumstance,the recognition rate declines so much.We learn from the life experience that ,people use not only the ears to feel the information,but also their eyes to see the change of looks,mouth shape of speaker during the talk.So the visual information can do a big favor to audio information.So lip-reading is a good choice to overcome the disadvantage of audio-recognition system under the noise background.Research shows that combining the visual with audio information can improve the recognition rate to some extent , particularly under the noise circumstance.It is necessary to pay much more attention to lip-reading.Research on lip-reading includes two parts: recognition of speech content and person authentication.People pay much more attention to the former research than the latter research.So in this paper, we will conduct corresponding exploratory research on person authentication, and particularly, we will foucs on the analysis of authentication ability of various kinds of features, based on lip-reading recognition, in order to get better visual features to describe the contour of lip model.At the beginning, we introduce the history and development of lip-reading, and then we present the theory background and key technology in details, according to the recognition procedure of classical lip-reading recognition system.The selection and extraction of features are the most improtant part in this paper, and finally, we regard the recognition result of HMM classfication as the standard to judge the authentication ability of different features.The extraction of lip features is a very important part of lip-reading recognition system,and whether selecting the efficient features to describe the lip model or not ,affects the performance of the recognition system to big extent.Because there exist many kinds of features,we need a great number of experiments to accumulate and analyse the data ,in order to get a better choice of features.The selection of features depends on the algorithm of feature extraction , and we use the 14-point ASM(Active Shape Model) in this paper.Accordingly,we select the width ,the height of the model as shape feature,the parametre combination of curve fit as geometry feature, first and second order difference of shape and geometry feature as dynamic feature,and combine them in different ways,to analyse their according authentication ability by experiment.The result shows that the features we select can work well,and we can get a good recognition rate.However,we should do a lot more if we want to catch up with the foreign research.