摘要
目前多数敏感图像过滤方法对皮肤裸露较多或类肤色区域较多的图像容易产生误检。为降低对这类图像的误检率,提出一种基于人体关键部位检测的敏感图像过滤方法。该方法提取肤色特征、表征局部对象外观和形状的HOG(Histograms of Orien-ted Gradient)特征、空间分布特征及描述区域灰度分布的Haar-like等特征,利用Adaboost学习算法,训练得到人体关键部位的分类器,通过此分类器实现敏感图像的过滤。实验表明,该方法能够准确地检测关键部位,可以有效地降低敏感图像的误检率。
At present,many non-pornographic images containing larger exposure of skin area or approximate skin-colour area are often prone to be detected as the pornographic images by most of the pornographic image filtering methods.In order to decrease the false detection rate,a new pornographic image filtering method based on erotogenic-zone detection is proposed in the paper.The method extracts main features,including skin-colour features,HOG features which describe the shape and appearance of local objects,spatial distribution based features and Haar-like features which describe local grayscale distribution,trains and obtains the classifier of erotogenic-zone recognition with Adaboost learning algorithm,and achieves the pornographic image filtering through the classifier.Results gained from the experiments confirmed that this method can precisely detect erotogenic-zone in an image,and can effectively reduce the fault detection rate against non-pornographic images.
出处
《计算机应用与软件》
CSCD
2011年第4期154-158,共5页
Computer Applications and Software
基金
浙江省科技厅重大专项项目(2010C11049)