人像辨識中人臉偵測之研究

Abstract

本篇論文在彩色影像上進行人臉偵測， 藉由一個描述膚色分佈的模型，標定臉部屬於膚色的區塊， 經過合併與篩選之後，我們可以快速地找出影像中可能的臉部區域， 大幅縮小後續處理的搜尋空間。接著在臉部區域上以雙眼最大距離， 周圍亮度，眼珠近似圓形等條件來精確地標記眼睛的位置， 進而定位出臉部在影像中的位置。本文所採用的資料庫中共有151人， 每人分別就水平旋轉$0^{\circ}, \pm22.5^{\circ}, \pm45^{\circ}$等角度取五張照片，共計755張。 實驗結果雙眼定位誤差均在三個圖素內的佔67.55\%， 誤差均在五個圖素內的佔70.20\%。在AMD-K6-2 300的機器上每張圖平均處理時間為0.83秒。

We propose a method for detecting human faces in a color picture. By applying a skin color model, we label some rectangular regions in the picture as skin. After merging several small skin regions and sifting the proper ones, we can quickly locate the possible face images in a picture. Since the skin regions are relative smaller than the original picture image, the overall computing time can be greatly reduced. In order to precisely locate the eye in the face image, we use some criteria such as the distance between two eyes, the intensity of the neighborhood and a circular pattern to locate both eyes. Finally one can use the coordinates of both eyes to properly extract the facial features for face recognition and/or some other applications. We also perform some experiments on the proposed method. Our face database contains 755 images taken from 151 persons. Each person has five photos taken from five different angles ($0^{\circ}, \pm22.5^{\circ},\pm45^{\circ}$) The correction rate is 67.55\% with fault toleration smaller than 3 pixels, and 70.20\% with fault toleration smaller than 5 pixels. The average overall running time is 0.83 second on a AMD-K6-2 300 machine.