結合色彩與紋理特徵之人臉偵測及識別技術

本論文旨在發展整合人臉偵測技術及識別方法，於整合人臉偵測法中結合目前現存的兩大方式，分別為色彩與紋理特徵。本研究的動機為針對居家照護系統中，建立此系統機制，以同時達到高識別率與即時偵測能力。目前主要有兩大獨立發展之人臉偵測技術，分別為色彩基礎法與紋理特徵法。對於以色彩為主之方式，其特色為可用於偵測人體膚色，並進一步做膚色定位藉以找出人臉中嘴唇及眼睛色彩，但其偵測率通常是不令人滿意。於紋理特徵人臉偵測法中，首先將影像轉換到灰階色彩空間，活用此法於影像上搜尋人臉位置，但其盲目搜尋方式較難達成即時應用之需求。於本文中將結合上述兩技術之特性整合成-色彩與特徵之人臉偵測法。本文活用膚色偵測法尋找影像中人體膚色所在位置，以減少後續紋理特徵法之搜尋範圍，免除盲目搜尋問題，完成人臉偵測程序之應用。為了改善人臉偵測率於文中加入可針對特定居家對象建立之高斯膚色模型，藉以達成彩色膚色分割與定位，而後與紋理特徵法結合並使用Haar like 特徵與Adaboost 演算法及串接式分類器找到影像中之人臉位置。人臉識別所用技術為費雪線性鑑別法以建模，與歐式距離來加以判別所屬人臉類別。並採取直方圖等化與小波轉換及原影像直接建模之前處理方式比較識別率與即時能力。最後在實驗結果中加入人臉偵測與識別整合實驗探討錯誤偵測框對識別率的影響。此外，本論文亦發展快速二維經驗模態分解演算法，可用於分解非線性、非穩態影像並找出人臉表情紋理特徵及還原相加成原影像之技術，未來可將其應用於人臉表情特徵萃取與識別中。

關鍵字

人臉偵測； Adaboost 演算法；串接式分類器；費雪線性鑑別法；二維經驗模態分解

並列摘要

The subject in this thesis is developed a combinational technology for face detection and recognition, and this combination is based on two existed face detections, color and texture features. Concurrently achieve the higher recognition rate and real-time detection capability for the digital home-cared systems is the motivation of this research. There are two major and independent face detection technologies, color-based and texture-based. For color-based feature detection, it detects the skin color firstly, and then locates the face based on lip’s or eye’s color, but the accuracy rate of detection is often unacceptable. For texture-based feature detection, it transfers image to gray color firstly, and then locates the face based on the textural features in whole face, but the almost blind searching during the face location is difficult to achieve the requirement of real time applications. In this thesis, the skin color is firstly detected for reducing the possible area there faces included, and then blindingly searches the face features in those bounded area, face recognition process is applied in finally. In order to improve the accuracy rate during face detection, a Gaussian skin color model of family is built. During the searching of face features in color-bounded area, a combination process of Haar likes feature, Adaboost algorithm and cascade classification is applied to locate the position of face. In the face recognition, the Fisher linear discrimination is used to build face models, and Euclidian distance discrimination is used in the face classifications basically. Histogram equalization with wavelet also used to be a preprocess method in additional to improve the accuracy rate of recognition and the ability of real-time. Finally, a real-time bi-dimensional empirical mode decomposition is developed to decompose a photo into some pictures with face texture in different resolution, this achievement is the foundation of the future research about the recognition of facial expression.