不良图像检测系统的设计与实现

【作者】 侯纯；

【导师】 陆建峰；

【作者基本信息】 南京理工大学 ， 计算机应用技术， 2012， 硕士

【摘要】 伴随着信息时代的发展,网络已经成为人们最重要、最便捷的信息传播工具。而在网络带给我们巨大便利的同时,各种不良的信息也随之传播扩散,暴力、色情等不健康信息充斥在网络各个地方。据统计,互联网中色情网站数量要有50多万个,而深受其影响的网民,尤其青少年网民更是数之不尽。所以如何有效的阻止色情信息的传播一直是互联网健康发展的重要课题。因此,本文设计实现一种基于肤色的不良图像检测系统,用来检验图像是否为不良图像。图像中大面积的裸露皮肤是判定该图像是否为色情图像的最重要的一个特征标志,肤色检测是色情图像判定的基础,所以本文主要对肤色检测技术进行研究,并实现一种在YCbCr空间的结合贝叶斯分类准则和阈值分割等技术的肤色提取方法。为了精确肤色提取结果,在此方法上再引入纹理分析等技术,进一步保证肤色提取的准确率。但是单纯的依靠对肤色的判定作为色情图像检测的唯一标准,不能保证色情图像检测的准确率。例如对穿着性感的明星写真或是头像证件照等图像的判定上,由于这类图片也存在大量的裸露皮肤,判定会有很大的误检率。所以,本文又对色情图像进一步分析,实验研究对女性胸部以及私处的判定方式。最终整合肤色提取、人脸检测、人体敏感部位检测等信息,设计实现不良图像检测系统。本文的主要研究工作和取得的成果如下：(1)实现在YCbCr空间的基于贝叶斯的肤色分类模型,并改进模型,设计实现一种结合Cr分量阈值分割和YCbCr空间贝叶斯分类的肤色提取算法。(2)对肤色检测采用纹理分析、形态学变换等方式优化肤色检测结果。(3)实验实现了现有根据色彩分析以及图形几何特征的胸部检测方法。(4)实验设计实现基于adaboost的胸部检测与色彩分析结合的女性胸部检测模型。(5)综合肤色检测、人脸检测以及敏感部位检测等技术,依据决策树的思想整合设计实现色情图像检测系统。更多还原

【Abstract】 With the development of the information, network has become one of the most important and the most convenient tool for the dissemination of information. The Internet brings us great convenience at the same time, all kinds of bad information also subsequently and diffusion. Violence, pornography and other unhealthy information filled in anywhere on the web. According to statistics, the number of Internet pornography sites must have about 500000, and the influenced Internet users, especially teenage netizen is uncountable. So how to effectively prevent the pornographic information transmission has always been an important topic for the development of the Internet health. Based on this, this paper presents the design and Implementation Based on skin color image defect detection system, used to judge whether the image is a bad image.The image with a large bare skin area is a most important feature to determine whether the graphics for the erotic images. Skin detection is the foundation of the judgement of pornographic images, so this paper is mainly on skin detection technology research, we realized a skin detect method in YCbCr space with the Bayesian classification criteria and threshold segmentation technology. In order to accurately extraction results, this method is then introduced to texture analysis techniques, to further ensure that the color extraction accuracy. But simply rely on color judgment as pornographic image detection only standard, it cann’t guarantee of pornographic image determining accuracy. As sexy star portrait, in these images there are lots of skin area too, the judges may have great error rate. So this paper on the erotic images for further analysis, experimental research on women’s breasts and genitals estimation. Finally the integration of skin extraction, face detection, character information such as the sensitive parts of detection, design and achieve a bad image detection system.The main research work and achievements are as follows:(1) In YCbCr color space, implemente a skin classification model which based on Bayesian.And we improved this model, design and implementation of a combination of Cr component threshold segmentation and YCbCr space skin color extraction algorithm of Bayesian classification.(2) On the basis of the above results by using texture analysis, morphological transformation and optimization of skin color detection results, and achieve better effect for skin color detection.(3) Experiments and achieve the chest examination method according to the analysis of image color and the pattern geometry. (4) Experiments design the female chest detection model based on AdaBoost chest examination and color analysis combined.(5) Combined with skin color detection, face detection and sensitive detection technique, based on the decision tree thought integration design of pornographic image detection system.更多还原