基于特征向量的敏感图像识别技术  

Feature Vector-Based Nude Picture Detection

作　　者：彭强[1] 张晓飞[1] 

机构地区：[1]西南交通大学信息科学与技术学院,四川成都610031

出　　处：《西南交通大学学报》2007年第1期13-18,共6页Journal of Southwest Jiaotong University

摘　　要：针对现有色情敏感图像识别技术的缺陷,提出了基于特征向量的敏感图像识别技术.用YIQ彩色坐标系的色度和YUV彩色坐标系的色调作为皮肤的肤色模型,以5×5象素窗口的SSD(sum of square difference)作为皮肤的纹理模型.从敏感图像的皮肤掩码图像中提取与肤色区域数量和比例等有关的6个特征.将提取的特征向量输入到SVM(support vector machine)分类机,根据输出结果判断是否为敏感图像.对含457张敏感图像和1 101张非敏感图像的实验表明,该算法分类的精确度平均达到89.3%,查全率达到85.8%,误检率为4.4%.To overcome the shortcomings in the existing nude picture filtering systems and image recognition techniques, a feature vector-based nude picture detection algorithm was proposed. In the algorithm, ehrominanee value of the YIQ color coordinate and hue value of the YUV color coordinate were taken to build the skin color model, and the SSD （sum of square difference） on a 5×5 pixel window is used as the skin texture model. Then features representing the number and proportion of skin areas are extracted from skin mask images and are consequently inputted into an SVM （support vector machine）. The experimental results on databases containing 457 nude pictures and 1 101 non-nude pictures show that the proposed algorithm averagely achieves 89.3% in detection accuracy, 85.8% in retrieval rate, and 4.4% in error rate.

关 键 词：敏感图像识别 图像过滤 肤色模型 纹理模型 支持向量机 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]