基于统计肤色模型的敏感图像检测

Sensitive image detection based on statistical skin color model

针对敏感图像的特征,提出了一种基于肤色分布统计特征的敏感图像检测算法。首先,扫描由小波变换系数构造的零树得到图像的显著点,选择显著强的点作为初始检测集,根据检测集的邻接区颜色梯度特征直方图,采用最大熵模型检测显著性点邻接区肤色信息,利用置信传播算法估计模型参数检测肤色值。其次,由视觉感知的封闭轮廓获得肤色区域解决肤色特征光照敏感性问题。最后,采用多超球一类支持向量机进行分类。实验表明:算法分类准确率达96.32%,同时具有较快的分类速度,平均每秒处理7幅图像。

According to the analysis of the characters of sensitive images, a skin color distribution statistical based classification method for sensitive image detection was presented. Salient points were detected for skin color initialization sets, and the skin distribution histogram models were computed based on salient point adjacent. Maximum entropy model was written down within the ones that had the feature histograms as observed on the training set. Belief Propagation （BP） was used for estimate of the parameters of the model. Due to the sensitivity to light for skin color features, contour distribution was computed. One Class SVM with multiple hyperspheres was adopted for image classification. The image-based classifier was able to properly identify 96.32 % of the evaluation images at an average processing speed of 7 images per second.