Multimodel Biometric Personal Identification Based On Data Fusion

In such a high developing information age, how to verify a person's identity correctly and protect information security is a crucial problem. Biometric systems seem to be in place of traditional identification such as key, password and ID cards in the future due to their convenience, security and trustiness. Most depoloyed biometric systems are unimodal which maks use of fingerprints, hand geometry, iris, retina, face or vein alone to verify a person's identity. These systems have been shown to be quite effective in labs. However, their performance easily degrades in a factual environment. Because every biometric system has its own limitations: the sensors always accompany noise; mismatch exists between training and test conditions; some persons lose their biometric traits because of damage, illness or low quality. To combat these limitations, a system which uses more than one biometric at the same time is known as a multimodal identification system can be adopted. It is often comprised of several modality experts and a decision stage. Since a multimodal system uses complimentary discriminative information, lower error rates can be achieved; moreover, such a system can also be more robust, since the contribution of the modality affected by environment conditions can be decreased. Multimodal recognition is therefore acknowledged as a mainstream research area as the next generation of biometric personal recognition.This thesis starts with two unimodal biometric systems based on face images or speech signals. Based on the research before, a very detail introduction about these two technologies is presented; further more, we do some improvement on them. Secondly, several methods used in fusion of speech and face information are proposed. Thirdly, a novel person recognition approach based on lip-movement is introduced. At last, an important application on grid is proposed to create a virtual collaborative environment linking distributed users, models, databases and other hardware and software resouces. The detail research content of this thesis is shown as follows:A new method combines advantages of VQ model and SVM model is proposedto speaker recognition. The hybrid VQ-SVM approach enhances the classifying ability of VQ modal; meanwhile it solves the training problem under large training data of SVM. Experiments show its performance further improved comparing to traditional VQ speaker recognition system. A face detection system is construched using skin color and Gaussian model. It is effective under the condition of simple background face images. A face recongniton system with classic PCA method is also built to verify person's identity.In terms of the fact that speech and face are loosely coupled even uncorrated at all, several methods, for use in fusion of speech and face information at match level, are proposed. These methods include adaptive weight linear combination, D-S evidence theory and artificial neuron network. Experiments show that the fusion system keeps at a fairly high level while the unimodel recognition system degrades quickly under noise conditions.Automatic lip segmentation is an essential stage in lip image analysis. It is also a very hard topic due to lots of difficulties. On the base of research work before, we present a novel segmentation method for lip tracker initialization which adapts fisher transformation in color space to enlarge distinguish between the skin color and lip color. This method is simple and quick. The parameters attained from region of interesting are effective to initialize deformable modal which is used to detect the lip brim. DCT transformation and PCA transformation keep down the main information of lip region in gray image while reduce the dimension rapidly.Lip movement is highly correlated with audio signal and speech content can be revealed througn lip-reading. It is quite natural to assume that lip movement would also characterize the identity of an individual as well as what the individual is speaking. In this dissersion, Audio-visual fusion fearture achieved by connecting acoustic features and lip shape-based feature or intensity feature directly which is derived from deformable model and DCT-PCA transformation coefficients within the lip region at each frame. A semi-continuous density HMM classifier with diagonal covariance matrix gaussian modes associated with each state is performed for the recognition and authenticaton. Experiments on HITLUDB audio-visual database shows an encourage result.Biometric databases, users, models are always diffirent and spread around geographically. Grid technologies are an aggregate of heterogeneous and distribute resources. It shows an extremely high efficiency in the field of large-scale complex applications and hige performance computering. In this thesis, a novel important platform on grid is proposed. It is used to create a virtual collaborative environment linking distributed users, models, databases and other resources. It also provides personal recognition fuctions on single or multiple biometrics and a testbends for the researchers.Personal recognition is an inevitable social problem. Multimodal identification using more than one biometric trait has been a robust and effective way to verify a person's identity. With the deeper of the rearch, this novel technology must attract more and more attention of researchers.