使用視位與語音生物特徵作即時線上身分辨識; A Real-Time Web-Based Personal Authentication System with Visemes and Acoustic Biometric Features

NCUIR > College of Electrical Engineering & Computer Science > Executive Master of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/8701

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/8701

Title:	使用視位與語音生物特徵作即時線上身分辨識;A Real-Time Web-Based Personal Authentication System with Visemes and Acoustic Biometric Features
Authors:	鄭復榕;Fu-Rong Cheng
Contributors:	資訊工程學系碩士在職專班
Keywords:	視位;biometric information;GMM classifier;visemes
Date:	2005-07-05
Issue Date:	2009-09-22 11:33:13 (UTC+8)
Publisher:	國立中央大學圖書館
Abstract:	音素是語言的基本聲音的要素，而視位是嘴唇對一個字的發音外觀。根據研究顯示[8][9]，發音的特點或方式足以作為個人身分辨識的依據，因此本系統將使用這些生物說話特徵作為個人身分辨識系統之主要特徵。於本研究中，主要的生物說話特徵包含有臉部正面的視覺資訊[1][2][3]、視位[4]與音素[4]。本研究將提出一個以使用中文的生物說話特徵作為身分辨識依據之web-based個人發音身分辨識系統，使用者只需要配備一台web camera即可以進行身分註冊與系統登入。本研究的主要目標是︰比較並分析這些生物說話特徵，使用GMM建立身分辨識分類器，並比較分析這些生物說話特徵應用於分類器的效能，最後建立一個個人身分認證機制的決策分類器。針對所提出的身份辨識系統設計方面，因為考量主從系統的擴充性與網路傳輸效能，本篇所採用的方法可能會犧牲一部分的辨識率；實驗結果顯示，本研究使用的生物特徵、辨識方法、系統設計架構，足以在合理的回應時間內處理320240以上解析度的人臉正面的視訊資料，辨識率可以達到80%以上。 Phonemes are the basic acoustic elements of a language, and the visemes are the lip shapes of a word while speaking. Since the uttering characteristic/manner is unique to each individual [8] [9], a personal authentication system can fully make use of the biometric information. The biometric uttering features include lip visual features [1-3], visemes [4], and phonemes [4] of an individual. In this thesis, we propose an effectively web-based personal authentication system by utilizing filter model with biometric uttering features in Chinese, that is a user equipped with a web camera can register and log in the system later by the registered biometric uttering features. In our work, we compare and analyze these biometric features, then build 2 GMM classifier, and finally fuse the outputs to a personal authentication system. The consideration of system scalability and networking bandwidth may more or less sacrifice the recognition rate. However, experiments results shows that the proposed system can handle data stream of personal frontal view in 320240 pixels with 30*15 pixels of lip in reasonable response time and satisfactory recognition rate above 80% can be achieved.
Appears in Collections:	[Executive Master of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Size	Format

社群 sharing

Loading...