【作者】 王颖芳；

【导师】 陈立伟；

【作者基本信息】 哈尔滨工程大学 ， 信号与信息处理， 2010， 硕士

【摘要】 科技的不断进步使网络也随之快速的发展,互联网就像一把双刃剑,人们还在欣喜于网络的强大时许多道德败坏的人也趁机窜了进来,他们上传许多淫秽、色情的图片,严重污染了网络空间。很多研究人员采用封锁黄色网站、对带有敏感字样的文本信息进行过滤等方法来控制色情信息在网络上的传播,但是这些对色情图像的控制都不是很有效,因此基于内容的敏感图像过滤技术应运而生。本文研究基于特定内容的敏感图像过滤技术,研究内容如下。敏感图像的过滤模块主要有肤色检测模块、纹理检测模块、特征提取模块和分类器模块。在肤色检测模块中,通过实验比较本文提出了色度空间模型和统计直方图模型相结合的方法来进行肤色检测,实验结果表明与传统的色度空间模型、高斯混合模型、统计直方图模型相比检测速度提高到2.32(幅/秒)。在纹理检测模块中,本文对五种常用的纹理算法进行了研究,在此基础上提出了一种将基于粒子群模糊聚类算法的边缘检测方法和Gabor函数相结合来进行皮肤纹理检测的方法,取得了较好的实验结果。通过与DCT变换方法的比较可以看出本文的方法在正检率上提高了16%,误检率下降了22%,实验结果很理想。在分类器选取中,本文选用决策树作为本文的分类器。实验结果表明,本文对敏感图像的过滤具有明显的效果,能够很好的分辨非正常图像与正常图像。更多还原

【Abstract】 With the continuous development of technology network progressed rapidly. WEB just like a double-edged sword, many morally corrupt men ran into upload pictures of many obscene, pornographic, causing serious pollution of cyberspace, when we are rejoicing in powerful WEB. Many researchers using blockade pornographic websites, filtering the text information with sensitive words to clean the WEB. But the image control of pornography is not very effective, so the technology of filtering sensitive image that based on content emerged as the times require. In this paper, we study the based on the specific content of the sensitive image filtering techniques. The study included the following.Sensitive image filtering module includes skin detection module、texture detection module、feature extraction module and classification module. In the skin detection module, compared through experiments, this paper proposed the method of combining color space model and statistical histogram model. Compared with color space model, Gaussian mixture model, and statistical histogram model, the experimental results show that the detection speed up to 2.32 images per second. In the texture detection module, in this paper, five kinds of commonly texture algorithms are researched. In the algorithm of skin texture, a fuzzy clustering algorithm which based on particle swarm optimization and the Gabor wavelet transform edge detection is proposed in this article. It has achieved good results. By comparison with the DCT transformation method can be seen that using this method, the correct detection rate increases 16% and the error detection rate decreases 22%.In classifier selection, this article selected tree as classifier. Experimental results show that the filtering of sensitive images have a significant effect, it is able to distinguish between sensitive image and normal image.更多还原