應用3D電腦圖學技術重建人臉正面影像

Abstract

人臉辨識近年來已被廣泛的研究，而許多辨識理論也隨之發展。然而大部分的人臉辨識理論對於輸入的人臉圖片在姿態上有著嚴格的限制。這些辨識理論只允許輸入有少許偏差的正面人臉。在本論文中,我們提出一各結合電腦圖學與電腦視覺的方法來解決上述的問題。此系統容許輸入非正面姿態的人臉，並將輸入的圖片合成為一各虛擬的人臉正面影像。輸出的影像可以輸入到大部分需要輸入正面人臉影像的辨識系統。易言之，本論文所提出的系統可以與其他的人臉辨識系統整合，並使其對於輸入的人臉在姿態上更具彈性。此外,本論文亦提出一各有效率的人臉姿態預估方法。人臉姿態預估可應用於許多電腦視覺相關的領域。例如人臉追蹤系統或是用於偵測汽車駕駛者狀態的人機介面。

Face recognition has been studied extensively in last decade and many techniques have been developed. However, most of presented face recognition system has strict restriction in pose variation. They only allow input images to be frontal-view or images with little pose variation. In this thesis, an approach that combines both computer graphics and computer vision is presented to handle this problem. The proposed system does not restrict the input image to be frontal-view and synthesis the input image into a virtual frontal-view face image. The output image can be the input of most face recognition systems, which require input images have no pose variation. In other wolds, the proposed system can be integrated to other face recognition system to make face recognition system avoid influence of head pose variation. In addition, an effective human head pose estimation method is also described in this thesis. The pose estimation method is useful in many computer vision tasks such as head tracking and man-machine interface used in detecting state of vehicle driver.