運用人臉動態及靜態資訊執行個人身分辨識

Abstract

本論文內容主要是在探討人臉辨識相關的研究課題，其內容共可分成三大部分。 第一個部份是有關於如何建立正確的人臉資料庫，以作為準確人臉辨識的基礎。因為在以統計為基礎的人臉辨識系統中，應該以人臉五官為辨識結果的依據，而不應該受非臉部分，也就是頭髮、肩膀、背景等影響辨識結果。在本論文中，我們利用統計學的方法，定量地說明所使用的人臉影像若包含頭髮、肩膀、背景等，而非單純的臉部五官影像時，將會影響辨識結果的準確性。經過實驗檢測，發現對絕大多數的合成人臉影像樣本（交換不同人臉影像中間五官的臉部分），都是以非臉部分決定辨識結果。根據這個研究結果，我們建立了一個純臉部五官影像的人臉資料庫，以進行接下來的人臉辨識研究。 本論文的第二個部份是有關於以統計上的線性區別分析為基礎的人臉辨識研究。就辨識能力而言，線性區別分析（Linear Discriminant Analysis）是一種很好的線性投影特徵擷取工具。它的目標是希望找出一組最佳投影軸，使得所有投影後的樣本之間，屬於同一群的樣本距離越近越好，而不同群的樣本距離越遠越好。一般而言，線性區別分析主要是基於對 Fisher's 函數進行最佳化，以求得最佳投影軸。不過，基於 Fisher's 函數而進行的線性區別分析，會發生所謂的小樣本數問題，也就是樣本數數目不夠去分群時，求 Fisher's 函數最佳解的過程中將會遭遇到計算上的困難。於是我們提出了一個定理，說明必須在 Fisher's 函數最佳化的子空間中，再做一次主要成分分析（Principal Component Analysis），才能求得線性區別分析的最佳解。實驗結果顯示我們所提出的系統大大改進了人臉辨識的效能。 由於直接利用人臉影像的亮度資訊來進行人臉辨識時，容易受到環境光線改變的影響而降低其辨識率，同時若只用單張影像灰階值進行辨識，無法分辨是否為非法者把照片放在攝影機之前的真偽。於是本論文的最後一個部份提出一個新的人臉辨識技術，來避免環境光線改變的效應。我們利用光流（Optical Flow）的概念來計算臉部的運動，繼而用臉部運動為特徵來執行人臉辨識的工作。由於需估測綢密的光流場，我們提出了一個新的光流估測技術，利用小波理論（Wavelet Theory）中的小波近似（Wavelet Approximation）來作為光流場及影像相關函數的模型，並運用了小波理論中空間基底（Scaling Function）的微積分可事先計算的優勢，將複雜的解光流場非線性最佳化的問題，簡化為解小波係數線性方程組的問題。根據實驗結果顯示，這個方法能估測得準確、綢密的光流場，在特徵點較少的人臉影像也能得到相當不錯的結果。光流本身是一種偵測運動的方式，人的臉部因表情變化或發聲講話時的扭動將產生光流，我們藉著每個人講話時臉部動作的差異性來作為個人身份辨識的標準。同時這種臉部運動的差異性並不會因為臉部灰階值改變而影響，如化妝、皮膚曬黑等。實驗結果顯示以臉部運動為基礎的人臉辨識在光源有變化下，其辨識結果確實比以灰階為特徵值的辨識效果來得好。

This dissertation discusses three major issues regarding face recognition. The first topic is to discuss what kind of face database is ``appropriate'' for statistics-based face recognition. Since the global data of an image are used to determine the set of decision boundaries in statistics-based face recognition, data which are irrelevant to facial portions should be disregarded. In this thesis, we first use a statistics-based technique to make a quantitative study on how the ``real'' face portion influences the face recognition result. After implementing the PCA plus LDA approach, the experimental results show that the influence of the ``real'' face portion on recognition is much less than that of the nonface portion. This outcome confirms quantitatively that recognition in a statistics-based face recognition system should be based solely on the ``pure'' face portion. The second topic in this dissertation is to develop a new LDA-based face recognition system. The major drawback of applying LDA (Linear discriminant analysis) is that it may encounter the small sample size problem. This problem arises whenever the number of samples is smaller than the dimensionality of the samples. In this thesis, we propose a new LDA-based technique which can solve the small sample size problem. The technique is based on an observation in which we prove that the most expressive vectors derived in the null space of the within-class scatter matrix using Principal Component Analysis (PCA) are equal to the optimal discriminant vectors derived in the original space using LDA.The experimental results show that the new LDA process improves the performance of a face recognition system significantly. The third topic in this dissertation is to address how to estimate facial motion accurately and how to identify a person based on his/her facial motion. In order to compute facial motion accurately, we first propose a new algorithm for optical flow estimation based on discrete wavelet approximation. The proposed method takes advantages of the nature of wavelet theory, which can efficiently and accurately represent ``information,'' to model optical flow vectors and image related functions. Based on wavelet modeling, the proposed method can successfully convert the problem of minimizing a constraint function into that of solving a linear system of a quadratic and convex function of wavelet coefficients. Once all the corresponding coefficients are decided, the flow vectors can be determined accordingly. We then adopt the above wavelet-based technique to estimate flow fields in an image sequence which contains a facial motion. Facial motion is represented by a high-dimensional feature vector that is constructed by concatenating a sequence of flow fields. Then this high-dimensional feature vector, which contains spatiotemporal information of a facial motion, is used in face recognition. The experimental results obtained by applying our method are superior to those obtained by an intensity-based approach in terms of robustness under varying lighting environment.