基于多示例学习的图像内容过滤算法研究

【作者】 龚慧超；

【导师】 侯晓霞； 项文波；

【作者基本信息】 南京理工大学 ， 系统工程， 2008， 硕士

【摘要】 本文结合多示例学习算法,研究并实现了基于图像内容的色情图像监控系统,从理论和实践上对多示例学习算法在图像过滤领域的应用进行了一个探索。特征提取是机器获取图像内容的重要手段。针对色情图像的特点,本文选取了颜色、纹理和形状特征作为特征值,由这些特征值组成特征向量交给多示例学习,利用多示例对未知概念包的预测功能来完成图像检测,进而实现图像的过滤。本文的机器学习过程采用多示例学习算法完成。首先将图像过滤中的概念统一到多示例框架下,色情图像特征的求解问题被转化成多示例问题中目标概念的求解问题,在多示例框架下采用EM_DD算法实现目标概念的求解,并用模拟退火算法对其改进,提高了搜索的速度和精度。通过对1500张色情图像和1500张正常图像进行检测,得出本算法的检出率为87.4%,虚警率为12.5%,从检测结果来看,本文提出的基于多示例学习算法的色情图像过滤算法能够有效地识别色情图像和正常图像。最后,本文在Visual C++6.0环境下开发了一个色情图像监控系统,该系统采用面向对象技术完成,具有多种检测方式和对浏览器的实时监控功能,此外该系统还能对色情网址进行记录、汇报、评级等。更多还原

【Abstract】 This paper unifies the Multi-Instance Learning(MIL) algorithm, studys and has realized a pornographic image supervisory system which is based on the content of image. And it has carried on an exploration in theory and practice to the MIL in the image filtration domain’s application.Feature extraction is an important means of the machine to obtain of image content. In view of the pornographic image’s characteristic, this article selectes color, texture and shape features as feature values. Feature vectors are made up from these feature values and learned by MIL. Filtration of image is realized by the prediction of the MIL to the unknown concept bags.In this paper, the machine learning process is completed by MIL. Firstly, the concept of image filtration is unified into the framework of MIL. The solution of pornographic image feature vectors istransformed into the problem of target concept searching, which was realized under the framework by the EM_DD algorithm. With the Improvement of the simulation annealing algorithm, the search speed and the precision was enhanced.Through the detection on 1500 pornographic images and 1500 normal images, it is known that the detection rate is 87.4%, the false alarm rate is 12.5%,. According to detection result, the pornographic image filtration algorithm in this paper can effectively identify pornographic images and normal image.Finally, this article developed a pornographic image Monitoring system, which was developed in the Visual C + + 6.0 environment with OOP. It not only has various detection functions, but also can monitor the web browser in real-time. Besides, the system can also record, report, rate etc. the erotic websites.更多还原