基於Fisherface演算法之人臉辨識系統實現

Abstract

本論文中提出一套有效率的人臉辨識系統，此系統主要分為人臉的偵測與定位以及人臉辨識兩部分.在人臉的偵測與定位方面，主要是建立膚色模型的方法來完成；當找到可能為人臉的候選區域後，接下來的工作就是要確定其是否為可辨識之人臉。首先用侵蝕及擴張的方法偵測人臉特徵的存在，然後利用基於不同角度之線性區別分析的方法去估測人臉的角度並且利用多重觀察的方法在可辨識之人臉區域中選取”純”人臉的部分去做辨識。在辨識的方面主要是採用Fisherface的演算方法並加入投票法則來決定其身份。 在實現此人臉辨識系統時並不需要精確的人臉特徵萃取，並且當應用於一些不同的條件之下如人臉的角度及表情變化，有無戴眼鏡或處於不同光度的環境之下，都可以達到很好的辨識效能。

In this thesis, an efficient face recognition system is proposed. The system mainly consists of two stages: one is face detection and location and another is face recognition. In face detection and location stage, a skin color model is used. The recognizable face is determined when the face candidates are obtained. First, the dilation and erosion functions are used for facial features detection and then, the view-based LDA method is used to determine the face orientation. When obtaining a recognizable face, the “pure” face portion extracted by the method of multiple observations is used for face recognition. In the face recognition stage, the fisherface algorithm is used, and the voting rules are applied to determine the face identification. In the face recognition system, the precisely facial features extraction is not required, and a satisfactory performance on this system is obtained even if the face image is under some uncontrolled conditions such as face orientation and expression change, wear (i.e. glasses) or not, and different lighting environment.