灰階人臉辨識之研究

Abstract

本論文提出一套不受背景、燈光、稍微傾斜、平移和大小等因素影響 的自動化人臉辨識系統，此系統主要分成下列三部份：人臉之定位、臉部 特徵之擷取和辨識。人臉定位的部份，主要是利用臉部特徵左右對稱的特 性，頭髮、額頭、眼睛、鼻子與嘴巴等彼此之間的相對位置以及其灰階值 變化的特性，而找出眼睛和嘴巴的初略位置，並以眼睛和嘴巴的鉛直距離 ，於背景複雜的影像中，將初步的人臉切割出來。第二階段，則利用第一 階段所得之結果，做更進一步的處理，以求取更精確的特徵位置，並將人 臉轉正，而於轉正後的人臉影像中，擷取出眼睛和嘴巴的特徵影像，並將 其大小標準化，以便於辨識。第三階段，則將第二階段所擷取的特徵影像 （眼睛、嘴巴），採用特徵臉的方法加以辨識，為了降低燈光對特徵臉方 法的影 響，我們將所擷取的特徵影像於辨識前做均等化（histogram equalization）之處理。實驗結果顯示本系統能於不同的背 景、燈光下 ，和人臉可平移、稍微傾斜以及大小不固定的條件下，仍具有很高的辨識 率。 An automatic face recognition system is proposed in the thesis. The system mainly consists of three phases : location, facial feature extraction and identification. In the face location phase, we base on the following properties : the symmetry of human face, the relative positions of hair, forehead, eyes, nose and mouth in a face, and the characteristics of the variation of the gray values of those facial features, to locate the positions of eyes and mouth. Then, the vertical distance between eyes and mouth is used to clip out the rough face region. In the second phase, the more accurate positions of eyeballs and mouth are located. The rotation effect is adjusted by aligning the center points of left and right eyes into the equal height. Then, the facial features are extracted for recognition. In the last phase, we use the eigenface approach to do recognition. The histogram equalization is locally used to reduce the effect of illumination. Experimental results show that the system has a good recognition rate under complex background, and allowing translation, small tilt, lighting and scaling changing.