【作者】 熊霞；

【导师】 桑庆兵；

【作者基本信息】 江南大学 ， 计算机应用技术， 2012， 硕士

【摘要】 近年来随着网络技术的日益发展成熟,互联网上各种资源信息成倍的增加,其中网络色情信息的泛滥已经成为危害人们身心健康的全球性公害,基于信息内容的过滤问题逐渐引起人们的重视,为了阻止不健康信息的传播和净化网络环境,很多研究学者包括商业市场都对敏感信息的过滤领域做出了大量的研究,一部分商业过滤软件也有投入使用,但是其主要的过滤方法都局限于网址屏蔽和敏感关键字匹配技术,具有滞后性,误检率和漏检率都较高等诸多缺点,并且目前网络中绝大部分的敏感信息都是通过图片为载体来传递的,因此提出根据图像内容进行过滤才能从根本上解决目前网络安全技术对图像信息过滤与监控能力不足的问题。本文主要在前人研究的基础上,结合自己对于图片内容理解和过滤原理的研究成果,提出一种新的并有效的图片内容检测过滤器。本文的主要研究工作及成果如下:(1)对图片格式的种类进行分析。本文主要介绍了目前互联网上图片格式:BMP、TIFF、GIF和JPEG等格式种类,并针对有损压缩图像中的JPEG图像的形成原理进行了介绍,为后文的基于压缩域的肤色检测提供理论基础。(2)图片颜色空间和肤色检测模型的种类和研究介绍。图片颜色空间包括RGB、YCbCr、YIQ、YUV、YCgCr等,并对颜色空间之间的相互转换公式进行了介绍,以及肤色像素点在不同颜色空间中聚类性的实验研究。本文还介绍了三种肤色检测模型包括阈值化肤色模型、统计直方图肤色模型和混合高斯肤色模型,并对其肤色检测效果通过实验进行了比较,为基于不同光照自适应的肤色检测模型的选择提供实验依据。(3)肤色纹理检测方法和图片分类模型的研究分析,肤色纹理检测方法包括一阶灰度统计法,主要原理是对肤色样本进行灰度统计,得到皮肤纹理的平均统计值与阈值进行比较判断。灰度共生矩阵法是从数学角度去研究图像纹理中像素灰度级的联合分布的统计量来表示纹理的空间相互依赖关系,Gabor滤波法通过对不同的频域信息全面的反映出图像空间的局部特征来进行纹理检测。还对三种敏感图片判定分类方法包括决策树分类法、支持向量机和BP神经网络分类法的判定原理进行了介绍和实验应用。?(4)通过对图片内容中的肤色检测方法的研究为背景知识,提出两种不同肤色检测的方法,分别是提高检测精度的光照自适应的肤色检测方法和提高检测速度的基于模糊认知图的图像压缩域肤色检测方法,并通过实验证明两种肤色检测方法均在其专注领域取得了很好的肤色检测效果,以此肤色检测为基础的不健康图片过滤系统也取得了很好的过滤正确率。通过大量实验数据表明,本文提出的敏感图像检测方法能够有效地过滤不健康图片并对图片内容的肤色检测也取得了很好的检测效果。更多还原

【Abstract】 Along with the mature of network technology, all kinds of the resources have been increased multiples. Especially in picture and video message area. As a result porn message have overflowed in network and become a global hazard. Public have a strong interest in image filter based on its content to forbid this un-healthy message spread in network and rationalize the relations of network. And a lot of scholars which include business market have a deep research in image filter field, of course some commerce image filter software have been sold by company. But there have a bottleneck in filter which based on website and sensitive key works to block the un-healthy message. It brings some shortages like hysteretic, high fallout ration and high miss detection rates, etc. But most of un-healthy information is spread by images. We need raise a research direction on image filter based on content which can solve the root cause of the shortage of network safety technology on image filter and monitor. This paper concentrate on the image filter which based on the former scholars’work, combine with mine insight in image content and filter theory, then raise a new and effective filter tool for image. The research works are list below:(1) Image format analysis. This paper have introduced popular image format at the moment: BMP,TIFF,GIF and JPEG, etc. And concentrate on the loss compression image which has a kind of current popular standard JPEG image. Given a brief introduction in compression and decompression theory which as a theory basement for image filter.(2) Research on images’color space and skin detection model’s choice. Images’color space include RGB、YCbCr、YIQ、YUV、YCgCr, etc. There are also have an analysis on transform methods on different color space and skin pixel distributes compression level in different color space. There are also having introductions on three skin detection models: threshold skin detection model, statistics histogram skin model and mixed gauss skin model. And through a skin pixel test to get the examining rate to compare the three models’merit rating which offer the groundwork on illumination adaptive skin color detection method.(3) Skin texture and image classification model analysis. Skin texture method include first order gray matrix, calculate the skin samples gray, then get the skin texture mean statistics to compare with the threshold value to get the result. Gray level co occurrence matrix in math angle to research images texture gray grade unite distribution for reaction the textures space inter rely on relation ship. Gabor filter method use different frequent region messages to reflect the image texture part of features. Image classifications include support vector machine, decision tree and BP nerve network classification.(4) According to previous background analysis on image content detection, skin pixel verification and image classification. There have a deep insight on image filter based on its’content. And got a conclusion on images’classification. This paper raise two detect methods on image filter: illumination adaptive skin color detection method which concentrate on improve are examine rate, fuzzy cognitive map based skin detection method in compressed domain which focused on detection rate. Follow by an experiment to verify the filter rate, and test result prove two different methods all get a well result on detection. Through the experiment method demonstrate the nude image filter method based on image content which raised in this paper get very well detection result.更多还原