利用嵌入隱藏式馬可夫模型擷取人臉特徵的人臉辨識系統

Abstract

針對人臉辨識這個問題，本文提出一個從人臉二維影像中擷取人臉特徵方法。此方法即使在頭部轉動變化大時也能夠準確得到人臉特徵的位置。我們利用嵌入隱藏式馬可夫模型來塑造人臉模型，並利用塑造過程的中間產物—狀態分佈序列來擷取長方形的人臉特徵。利用單一嵌入隱藏式馬可夫模型，我們能夠對單一身份固定姿勢下的人臉進行臉部分割、擷取特徵的動作，當頭部轉動變化大時，我們使用人臉影像訓練出多個嵌入隱藏式馬可夫模型，使得我們的方法能夠準確取得 人臉特徵。針對人臉身份識別和人臉身份確認這兩個議題，我們為每一個使用者的每一個特徵建立其外觀子空間，進而利用樣板比對的方法估計特徵相似度。接著再使用Adaboost 演算法將每一個特徵訓練成弱分類器，並合併所有的弱分類器達到結合各特徵的效果。我們利用卡內基美隆大學提供的PIE 人臉資料庫、劍橋大學提供的ORL 人臉資料庫、以及我們實驗室自己拍攝的人臉資料庫上進行人臉辨識的實驗。經由與其他數個方法的比較，我們的方法均獲得較好的辨識率。實驗結果證實在頭部轉動變化大時，我們的方法依然能夠準確的取出人臉特徵，進而在人臉身份比對和人臉身份確認下有較高的準確度。

We propose an algorithm for extracting facial features robustly from images for face recognition even under large pose variation. Rectangular facial features are retrieved via the by-products of an embedded Hidden Markov Model (HMM) which decodes an observed face image into a state sequence. While an HMM is able to segment images into features at a fixed pose, multiple HMMs are trained for each individual to robustly extract features under large pose variation. Using the extracted features of each individual, appearance models based on subspaces are constructed for face identification and verification. Then Adaboost is used for feature combination while each weak classifier compared the distance metric of one facial feature. The effectiveness of the proposed approach is validated through empirical studies against numerous methods using the CMU PIE, ORL and our lab’s database. Our experiments demonstrate that the proposed approach is able to extract facial features robustly, thereby rendering superior results in identification and superior performance in verification under large pose variation.