基於主動外觀模型及彈性束圖匹配之人臉辨識系統

Abstract

早期的人臉辨識方法常利用簡單的幾何特徵來做辨識，例如顏色、形狀、位置等特徵，而現今的人臉辨識技術已成熟至使用數學的描述及比對方法。人臉辨識技術可用於人臉的偵測與識別，人臉辨識的方法主要可分成兩類：第一種是以特徵為基礎(feature-base)，第二種是以圖像為基礎(view-base)。在眾多的人臉辨識技術中，最廣泛應用的方法為主成分分析(Principal Components Analysis, PCA)的相關演算法，然而PCA基本上對姿勢及表情的變化非常地敏感，因此本論文提出一種創新的feature- base演算法，它不易受姿勢及表情的變化所影響，以此為基礎的人臉辨識系統使用主動外觀模型(Active Appearance Model, AAM)來偵測人臉的特徵點，並且採用彈性束圖匹配(Elastic Bunch Graph Matching, EBGM)來識別偵測到的人臉是否為系統成員，其中我們利用5組不同大小及8組不同方向共40張Gabor filter來描述由AAM所得到的特徵點，並經由分析組內差異(within-class deviation)及組間差異(between-class deviation)的方法從58個特徵點中挑選出最具代表性的14個特徵點來做人臉辨識，此外我們也以不同大小的AR人臉資料庫來分析特徵點的選擇與資料庫大小的關係，最後並將實驗結果與文獻中使用PCA及LDA的實驗結果做比較以檢驗系統的性能。

Early face recognition algorithms used simple geometric features, like feature color, shape, and position etc., but the recognition process has now matured into a mathematical representations and matching processes. Face recognition can be used for both verification and identification. There are two predominant approaches to the face recognition problem: geometric (feature-base) and photometric (view-base). Among many different algorithms, Principal Components Analysis (PCA) related algorithms are the most popular method in face recognition field. However, PCA basically is very sensitive to pose, expression change. In this thesis, we propose a novel feature-base algorithm which is more robust to pose and expression influence. The face recognition system uses Active Appearance Model (AAM) to detect feature points on the face and adopts Elastic Bunch Graph Matching (EBGM) to identify whether the face is a member or not. Firstly, we introduce 40 Gabor filters in 5 different sizes and 8 different directions to describe the feature points obtained from AAM. Besides, we also analyze within-class and between-class deviation of 58 feature points and find out the optimized 14 feature points for face recognition. Furthermore, we use different sizes of AR database to analyze the relationship between feature points selection and database size. Finally, our experiment result is compared with that of PCA and LDA to examine our system performance.