利用機率圖模型於影片上之人臉辨識研究

Abstract

針對影片上的人臉辨識問題，本論文提出一個機率圖模型來解決並將其公式化。首先，我們將此問題分成兩個部份來探討，分別為相似度之計算與遞移機率，其中相似度之計算可被視作為傳統的單張影像之人臉辨識的結果，在此篇論文中，我們採用二維線性鑑別分析法(2DLDA)摘取特徵，再藉由高斯分佈來估算相似度。而遞移機率則是計算先前時間點的狀態轉移到此時間點的狀態之機率，我們可將遞移機率分成兩個部份來估算，其一為人與人的遞移機率，另一個則為姿勢轉換的遞移機率，希望藉由相鄰影像的時間關係修正錯誤的辨識結果與提升準確率。在本論文的實驗中，我們使用在國際上常採用的 Honda/UCSD 資料庫以及本實驗室自行建立的VIPlab資料庫。實驗證明本研究提出之方法可適用於不同的資料庫，且實驗結果也有90%以上的正確率。

We present a probabilistic graphical model to formulate and deal with video-based face recognition. Our formulation divides the problem into two parts: one for likelihood measure and the other for transition measure. The likelihood measure can be regarded as a traditional task of face recognition within a single image, i.e., to estimate how similar to a specified person this observing face image is. In our work, two-dimensional linear discriminant analysis (2DLDA) is employed for feature extraction, and then we use a Gaussian distribution to assess the likelihood measure. The transition measure is estimated via two terms, person transition and pose transition. The transition terms could fix some incorrect recognition results because of considering the information between adjacent frames. In the face recognition experiments, we adopt two datasets, Honda/UCSD dataset and VIPlab dataset. Finally, it is demonstrated that our proposed approach is robust in different datasets and produces good recognition accuracy which is more than 90%.