膚色偵測器應用於即時臉部擷取系統

Abstract

人臉偵測系統在人臉辨識、人機介面裝置以及影像監視器的應用上扮演非常重要的角色。本篇論文提出一套針對不同亮度做補償之膚色偵測器，由於膚色極容易受光線影響，因此，必須做一些適當的光線補償。首先，本篇先對輸入影像做光線補償來做亮度校正。隨後，我們利用訓練出的CbCr值來做膚色分割。在訓練過程中，採用「主軸k-means」演算法將膚色訓練資料做分群處理。在特徵擷取部分，利用眼睛、嘴巴，以及比較少見的頭髮等三個特徵來定義人臉區域。實驗結果顯示出，本篇論文所提出的方法已改善光線所產生的問題，而且系統可以即時工作，在膚色偵測率高達95.02％；針對較複雜背景情況的靜態人臉偵測率高達91.67％；動態人臉偵測率高達97.78%。在此，我們主要是針對正臉以及接近正臉的人臉來做實驗，對於背景太複雜的情形下，仍然有偵測不到的現象。

Human face detection plays an important role in applications such as face recognition, human computer interface, and video surveillance. In this paper, we proposed a skin color detector for lighting compensation. Because skin color always suffers from illumination, lighting compensation method is used to improve the skin color detection. First, we use lighting compensation algorithm for lighting calibration and perform skin segmentation using thresholds of Cb and Cr by training. Second, in training phase, we utilize principle axis k-means algorithm to cluster training data. Finally, in feature extraction phase, we use hairs, mouth and eyes to help to locate faces. Experimental results show that proposed method has good performance. In the static skin segmentation experiments, the detection rate is as high as 95.02%, the static face detection rate is as high as 91.67%. For dynamic face detection, the detection rate on front or near front faces can achieve up to 97.78%. However, our model also can’t perform on extremely complex background well.