【作者】 张星；

【导师】 屈志毅；

【作者基本信息】 兰州大学 ， 计算机应用技术， 2009， 硕士

【摘要】 随着21世纪信息技术及互联网技术的飞速发展,网络信息已成为一种人们熟知的便捷信息来源和休闲生活方式,但网络上大量的色情淫秽等不良信息也同时干扰着人们正常的网络生活,严重毒害着青少年的身心健康。作为其技术支持,基于内容的不良图像过滤技术日益引起人们的重视。然而由于不良图像识别本身具有极大的复杂性,目前的不良图像识别系统普遍存在着准确率和检出率同时偏低或着一高一低的问题,而且过滤系统的效率低,耗时大。本文正是在这样的背景下,经过对大量不良图像的研究,总结出了不良图像中的一些人体结构特征,提出了把是否显现了人体某个或某些特定器官或部位作为不良图像的判断依据的新思路,这也是改变目前仅依据低层特征进行不良图像识别的可取方法。正是从这种新的角度出发,本文进行了不良图像识别与检测技术的深入研究,综合人体结构、自适应肤色检测、低通滤波、图像分割、图像过滤、区域合并等关键技术,提出了一种基于内容的不良图像躯干定位算法。本文主要开展了如下研究工作:首先,对肤色检测的原理和相关技术进行了详细介绍。选择了基于YCbCr颜色空间的高斯肤色模型进行自适应肤色检测,并且利用低通滤波对图像进行消噪。实验表明,使用这种方法能够有效地检测出图像的皮肤区域。其次,对图像分割的关键技术进行了深入研究。选择了分割效果较理想的Normalized Cut(Ncut)分割方法对肤色检测图像进行分割,取得了不错的效果。再次,对基于内容的不良图像躯干定位算法框架进行了介绍,其中详细陈述了本算法的每一个实现模块和步骤。最后,进行实验分析和验证。实验结果表明,本文提出的基于内容的不良图像人体躯干定位方法达到了93.3%检测准确率,基本实现了敏感部位的定位,达到了预期的研究目标,同时该课题通过从一个新的角度对不良图像的检测方法进行研究,为打击色情图像的传播提供了新的技术支持。更多还原

【Abstract】 With the rapid development of Internet information technology in 21st century, the network information has become a kind of people’s convenient information source and leisure lifestyle. But the Internet also is flooded with all kinds of pornographic contents, which terribly influences on the people’s normal network life, particularly the health of teenagers. As the monitoring technical support, content-based filtering technology to adverse information has gradually attracted people’s attention. However, the current pornographic image recognition system’s detection accuracy and efficiency are both low, or one high one low.This paper observes many pornographic images and gets some human structure characteristics and considers whether human body’s sensitive organs or parts are shown should be used to judge the pornographic image’s feature. From the new perspective, the paper makes a deep research on pornographic image recognition and detection technology. Combining human body structure characteristics, adaptive skin detection method, low-pass filter, image segmentation, region merging technology, a content-based human torso detection method is proposed.The paper mainly do the following research work. Firstly, the paper introduces the skin detection principle. The paper selects the adaptive Gaussian skin model based the YCbCr color space to detect skin regions. The low-pass filter is used to eliminate the images’ noises. The experimental results show the method can effectively detect the skin regions. Secondly, the paper makes a introduction to the key technology of image segmentation. The paper chooses the Normalized Cut (Ncut) method. The experiment also achieves good results. Thirdly, the paper gives a introduction of the content-based human torso detection algorithm. The algorithm framework and implementation is described in detail. Lastly, the paper do experimental analysis and verification. The experimental results show the algorithm proposed in this paper achieves 93.3% detection rate. And it basically implements the sensitive parts’ positioning of the pornographic image. So it achieves the desired research aim. And the subject from a new perspective provides a new technical support for prohibiting the pornographic images’ flooding.更多还原