本文使用了一套基於資料分群概念的系統SVM訓練來做人臉辨識系統。辨識人臉首先必須先偵測到人臉影像,利用一種迭代演算法AdaBoost來偵測眼睛,得到眼睛中心坐標後,將兩眼坐標帶入設計好的橢圓遮罩去切割出人臉部位,由於原圖的五官對比不夠明顯,為了增加人臉五官的對比強度,使用了Retinex這個演算法,過去此演算法常用於數位相機自動白平衡中,它除了具有色彩恆常性,亦包含了強化影像的效果,對於光線對影像所造成的影響,具有良好的處理效果,接著就是使用擷取可辨識的臉部特徵,我們使用對紋理特徵有良好效果的LBP 特徵,只在灰階上的運算, 速度較快可應用於即時系統上,最後使用SVM 分類器去訓練LBP所擷取的特徵來做辨識。
Face identification for security systems has become an important research subject. This thesis proposes a face identification system for application in area access control systems. Support vector machine (SVM) was employed to conduct face identification based on a data clustering method. Initially, the face was detected using the AdaBoost algorithm. An elliptical mask was then used to remove the non-face area of the image. If the contrast was insufficient to produce a full-featured image of the face, the image was enhanced using the Retinex algorithm. This algorithm corrects lighting condition and maintains color constancy. A local binary pattern (LBP) was used to capture the facial features because it positively affects the characteristics of the texture. Employing an LBP is simple and fast; therefore it can be appropriately applied in real-time systems. An SVM classifier was used to train LBP features to accomplish identification. The proposed system is proven to be applicable for access control with satisfactory correct identification accuracy.