A Study Of Sparse Coding Based Lip Biometric Representation

Several different kinds of biometric are applied to the field of human identification and authentication,which brings convenience to our daily life.Lip biometric,a newly developed biometric,can provide sufficient information related to speaker's identity.Lip biometric is believed promising because it's easy to collect and user friendly.Some researches show that human lip contains two different type of information: physiological and behavioral.The discriminative power of them is still unknown.Based on these researches,the presentation of lip biometric is discussed in this paper.Two different presentations are proposed to fit different situation and requirement.The first one demands lip outer contour being extracted.The sparse coding is applied to reconstruct the lip region and the reconstruction error is used to identify the speaker.The second one can be used when lip outer contour is hard to get.The image containing lip is coded in sparse coding manner,and max-pooling is adopted to generate feature vector.Support vector machine is then employed to identify the speaker.These two presentations work efficiently and provide enough information for speaker identification and authentication.In the experiment on standard lip database,the first presentation achieved an identification accuracy of 91.37% and while the second one achieved an identification accuracy of 98.57% and an authorization HTER of 2.15%.