Research On Authentication By Fusing With Speaker Recognition And Face Recognition

Multimodal biometric authentication technology has become a hot research topic in today's society because single-mode biometric authentication has its own limitation to satisfy the needs of human beings for nowadays highlight of personal identity information security issues. The thesis realized a speaker recognition and a face recognition firstly. Then the results of both speaker recognition and face recognition are fused in decision stage according to the signal-to-noise rate(SNR) of speech signal and the illumination of image signal to improve the accuracy of authentication so that the needs of the SNR for the speaker recognition and illumination for the face recognition are satisfied. The main work is as follows:Firstly,the thesis realized a speaker recognition by using the Gaussian mixture model-based speaker recognition. The acoustic feature vectors of Mel frequency cepstrum coefficients(MFCC) were extracted by pre-emphasis, framing, windowing and other processes. Then each speaker's acoustic model are trained with acoustic feature vectors extracted from training corpus. All speaker's acoustic model spooled together to obtain a GMM model library. In the recognition phase, acoustic features of MFCC were extracted from input speech signal uttered by the speakers who are going to be identified. The recognition result was obtained according to the maximum matching probability of acoustic feature vectors on all GMM models. The experimental results show that the implemented speaker recognition can achieve 92.8% of recognition rate on clean speech.Secondly, the thesis realized a face recognition by using GMM classifier. In the training stage, face was firstly detected from pre-processed image by using a skin color-based face region detection method. After that, the facial features are extracted by using a principle component analysis(PCA) method from normalized face image. Finally each register's GMM face model was trained with their facial features and spooled to form a GMM face model library. In the recognition stage, face region was firstly detected and facial features are extracted from input image that will be recognized. Recognition result was obtained by matching facial feature vectors on each face model in the GMM face model library according to the classification threshold of system. Experimental results show that implemented face recognition can reach 78.1% of face recognition rate under the condition of high SNR.Finally, the thesis realized a decision stage fusion authentication method by fusing results of both speaker recognition and face recognition. A SNR threshold for speech and a brightness threshold for image are set depend on the experiment. Face recognition result was used for authentication then the speech SNR low than the threshold while speaker recognition result was used for authentication when image brightness low than the image brightness threshold. When speech SNR and image brightness are both great than threshold, the authentication result was obtained by fusing speaker recognition result and face recognition with a set of fusion weights that is obtained by a segmental rule. Experimental results showthat the average recognition rate after fusion is improved 1.55% than the speaker recognition result and is improved 17.41% than the face recognition result in the case of a high speech SNR, and the average recognition rate after fusion is improved 73.22% than the face recognition result in the case at low image brightness.