使用膚色比例前處理之即時性人臉偵測系統

Abstract

在此篇論文中，我們建構出即時性人臉偵測系統，此系統能偵測出-45度至45度的人臉。我們的人臉偵測包含兩部份：第一部份尋找出整張影像中的膚色區域，以獲得可能為臉的區域。由於偵測區域只在膚色區域上，故處理時間可節省20~95%；第二部份為人臉偵訊程序，我們使用積分影像(integral image)快速算出矩形特徵(rectangle feature)，且使用AdaBoost 演算法選出重要的特徵。我們的系統建構三層cascade系統。第一層為0度人臉偵測，第二層為45度人臉偵測，最後一層為-45度人臉偵測。我們也比較與分析以下性能: (1) 使用與非使用AdaBoost 的特徵選取，(2) 兩個不同的系統設計， (3)我們系統使用與非使用膚色前處理的不同階段。我們系統的偵測率能達到88.31%。 在即時性人臉偵側方面，根據不同大小的膚色面積，我們需要20~400ms去處理一張320*240的影像。因此，我們的系統能廣泛地使用於不同的應用程式上。

In the thesis, we construct real-time face detection system which can detect -45~45 degree face. Our face detection system consists of two parts. The first part searches skin color regions over the whole image to segment potential face regions. Detection regions only focus on skin regions so it can save 20~95% processing time. The second part is face detection procedure. We use integral image to compute rectangle features rapidly and AdaBoost algorithm to select important features. Our system constructs three-layer cascade structure. First layer is 0-degree face detection, second layer is 45-degree face detection, and final layer is -45-degree face detection. We also compare and analyze the performance: (1) feature selection with and without AdaBoost, (2) two different systems, and (3) different layers in our system with and without skin ratio preprocessing. The detection rate of our system can achieves 88.31%. In real-time face detection, it requires 20~400ms to process a 320*240 image depending on skin area size. Therefore, our system can be widely used in different application programs.