- 期刊
多階段人臉偵測與辨識
Face Detection and Recognition Using Multi-level Classifier
摘要
人臉偵測與辨識的應用日益趨於普及與成熟，未來可能的應用更是值得期待。我們在本研究中介紹我們所提出的兩階段串聯式人臉偵測器以及基於最近特徵空間轉換的人臉辨識方法。在人臉偵測方面，我們採用了捲積式類神經網路（Convolutional Neural Network，CNN）所產生的特徵映像（Feature Map，FM）作為串聯式分類器（Adaboost）的輸入資料。從實驗結果得知，我們提出的方法能利用較少的弱分類器建置出一個偵測速度快的人臉偵測器。在人臉辨識方面，我們提出了最近特徵空間轉換。最近特徵空間轉換法是以點到點、點到線、點到面等最近距離的向量作為共變異數計算的基礎，其所構成的共變異數矩陣具有較佳的特徵空間，所謂較佳，意指此特徵空間較傳統點到點所求得的特徵空間讓投影到其中的樣本點更具有一般性與代表性，最後並介紹相關應用。
並列摘要
In this study, we propose a face detection methodology and a face recognition methodology, respectively. In the proposed face detection method, convolutional neural network (CNN) and Adaboost mechanisms are hybridized to form a more powerful scheme. Based on the feature map S of CNN, a coarse-level cascade classifier is obtained to efficiently filter out most of background regions. However, a few face-like false alarms cannot be removed because the feature map S is too small to keep sufficient information. In consequence, a fine-level cascade classifier based on source image with more detail information but fewer stages is adopted to remove the remaining face-like false alarms. As to the proposed face recognition method, a nearest feature space embedding (called NFS embedding) algorithm is proposed for face recognition to alleviate the effects resulting from facial pose, illumination, and expression (PIE). The distance between a point and the nearest feature line (NFL) or an NFS is embedded in the transformation in the discriminant analysis. Three factors, namely class separability, neighborhood structure preservation, and NFS measurement, are considered to find the most effective and discriminating transformation in eigenspaces. Finally, some applications of face detection and face recognition are introduced to conclude this study.