Abstract: For the lip area detection，the method of applying Haar features to train Adaboost classifier can well adapt to kinds of background environment，but can only get the rectangular area，which can not accurately locate the lip area，while the usual lip color separation algorithm can accurately locate the lip area，but has the high dependence on the background environment of images.Based on this，this paper realizes a method that a self-adaptive skin and lip color separation model is established，which can dynamically adjust constant parameters of skin and lip separation algorithm based on the results of applying Haar features to train Adaboost classifier.The model can dynamically acquire distribution range of skin color and lip color，so it can accurately achieve the lip region of kinds of background images，and improve the effectiveness and robustness of lip-reading deletion.Applying the method to deal with 4000 images in GENKI database，it successfully detects lip area and makes curve fit for the edge of the lip region to locate lip.The results show that the method is more effective.

Key words: self-adaptive skin and lip color separation model, Adaboost classifier, matrix transform

摘要： 在唇区检测中，应用Haar特征训练Adaboost分类器方法能很好地适应各种背景环境，但是只能得到包含唇部的矩形区域，并不能准确定位嘴唇部分，而常用的唇色分离算法虽然能准确定位唇部，但是对图像的背景环境要求较高。基于此，提出自适应唇色分离方法，该方法是以Haar特征训练Adaboost分类器作为基础，自适应地调整唇色分离的常量参数，从而能够动态地获得唇色与肤色的分布范围，实现准确地获得各类背景图像的唇部区域，很好地提高了唇区检测的精确性和鲁棒性。利用该方法对GENKI数据库中4 000幅图像进行处理，成功地实现唇区检测，并对唇区域边缘图进行曲线拟合来实现定位，实验结果表明在各种复杂背景下，该方法更有效。

关键词: 自适应唇色分离, Adaboost分类器, 矩阵变换