This thesis is a study on a vector quantizer-based Biometrics that uses voice and face information. The author proposed a classified space vector quantizer which decreased computational complexity of training and recognizing procedure for speaker rec...
This thesis is a study on a vector quantizer-based Biometrics that uses voice and face information. The author proposed a classified space vector quantizer which decreased computational complexity of training and recognizing procedure for speaker recognition. And proposed a principal component analysis vector quantizer which increased the discrimination among the faces in eigenspace for face recognition. And proposed a information fusion method which utilized the vector quantizer for multi-modal biometrics fusing face and voice.
It was hard to make fast and precise recognition system by a discrimination limit of biometric information. For solving the problem, this thesis was devoted to study of three main method. Firstly, author researched the algorithm for reducing computational complexity of vector quantizer utilized for speaker recognition. For reducing the computational complexity, a constrained vector quantizer which was used as classifier or recognizer was proposed.
We named this constrained vector quantizer the space classified vector quantizer. The proposed vector quantizer divides space of data into classified space codebook. And then the produced code vector is the centroid of a classified space. Therefore the proposed vector quantizer is not necessary to compute for iterative learning and decreases the computation complexity for recognition procedure. Secondly, for improving the discriminative ability on the variation of faces, a principal component analysis vector quantizer was proposed. The proposed novel algorithm is the extension of eigenface using vector quantization. This vector quantization make a cluster which represents the distribution of feature vectors in eigenspace, so that the proposed method accepts the variation of faces and make the improvement of recognition rate. Thirdly, the uni-modal biometric methods have the limitation on the stability of system and the ability of recognition. For solving the problem, we proposed multi-modal biometric systems using face and speech, which are based on vector quantizer. The proposed systems are made by the feature vector level information fusion and the decision level information fusion. The feature vector level information fusion has more discriminative ability by the unified feature vector obtained from face and voice than the uni-modal system of face or speech. Also, decision level information fusion recognizes the result of AND, OR decision obtained on the both vector quantizer of face and speech. Therefore, this system has an advantage, which mutually supplement a defect caused by noise of image and sound.
The main results of the thesis are as follows. firstly, we made the experiments on speaker recognition using the classified space vector quantizer and the conventional vector quantizer. The systems had constructed by the speaker-dependent system and the speaker-independent system. When the recognition rate is equal, the classified space vector quantizer reduced the training and recognition time to one tenth compared with that of the conventional vector quantizer. Secondly, we made the experiments on face recognition using the principal component analysis vector quantization method and eigenface method. The experimental results show 10% improvement on the recognition rate for all conditions. Thirdly, we made the experiments on the multi-modal recognition using vector quantizer. In experimental results, the feature vector level information fusion had improvement on the recognition rate for all cases. And the decision level information fusion had the more stability than uni-modal system.
In conclusion, the prosed methods have the reduced computational complexity, the improvements on the recognition rate, the enhanced stability. Therefore, the proposed vector quantizer-based biometric systems are useful in the implementations and the applications.