分类号：密级：
学号：2009205 单位代码：10407
表格怎么查找
硕士学位论文论文题目: 敏感图像检测算法研究
研究方向模式识别与图像处理
神彩皓月
专业名称控制理论与控制工程
研究生姓名程智林
导师姓名、职称杨国亮副教授
和月有关的诗句2012年05月27日真实灵异
江西风景
亲戚关系表江西·赣州
摘要
互联网和多媒体技术的发展，使得色情图像在网络上传播愈加泛滥，为了净化未成年人的网络环境，给未成年人营造健康而纯洁的网络空间。敏感图像的识别和过滤技术就受到了极大的重视并独立成重要的研究课题。本文对敏感图像识别和过滤算法进行了分析和研究，并给出了相应的实验结果。
具体完成的工作如下：
1、对目前国内外敏感图像检测算法进行了概括和总结，并陈述了敏感图像检测算法目前的技术难题。
2、由于敏感图像有一个共同的特点---有大量裸露皮肤。根据色彩空间和肤色模型的分析，提出了一种基于图像增强算法的肤色检测模型，它是将图像增强算法和双肤色模型结合，弥补了双肤色在复杂环境下检测效果不理想的劣势，提高了检测的准确率。
3、在肤色检测的基础上，引入纹理检测算法。本文利用一阶灰度矩形算法检测纹理，消除了非肤色和误检区域的干扰，为以后提取特征向量奠定了基础。
4、由于人脸图像对敏感图像检测存在干扰，因此本文加入人脸检测算法，可以去除人脸图像的干扰。选择人脸检测准确率较高的算法---Adaboost算法，有效提高了人脸检测率。
5、在得到肤色掩码基础之后，对这些显著区域提取特征向量作为分类器的输入，这些特征向量主要包括肤色占整幅图像的比例、肤色占外接矩形的比例、肤色连通区域的个数、最大连通区域占整个图像的比例、最大肤色连通区域占肤色外接矩形的比例、图像中心肤色区域肤色比例。
6、在提取完特征向量之后，接下来是分类器的选择和设计，分类器的分类能力直接影响到色情图像识别的性能。在选择分类器时，选取对于解决小样本、非线性和高维模式识别、分类问题有较大优点的SVM分类器，并给出了实验结果。
7、为了进一步提高分类的正确率，构造了PSO-SVM模型和GA-SVM模型对SVM的参数进行优化，并对这两种模型进行了大量的Matlab仿真实验，实验结果表明，这两种模型比单纯的使用SVM的检测效果尤其是准确率上有较大提高，达到了我们预期的目的。
关键词：敏感图像肤色检测纹理检测人脸检测特征提取支持向量机
ABSTRAT
With the constant development of Internet and multimedia technology, the spread of pornographic images becomes more and more rious. In order to purify the network environment for minors to minors, create a healthy and pure cyberspace. Sensitive image identification and filtering technology
has been a great deal of attention and parate the important rearch topic. This article analysis and rearch algorithms of Sensitive image recognition and filtering and gives the corresponding experimental results.
The specific work is as follows:
1、reviewed and summarized domestic and foreign nsitive image detection algorithms, and summarizes the technical problem of nsitive image detection algorithm
2、As a result of pornographic images have a common characteristic ----- a large number of bare skin. According to the color space and color model analysis, propod a method bad on image enhancement algorithm for skin color detection model, it is combined to image enhancement algorithm and double skin model, This model make up the double color in complex environment detection effect is not good and improve the detection accuracy.
3、Bad on the texture detection algorithm in color detection, this paper by using first order gray rectangle algorithm to detect texture, no skin and eliminates the error area of interference, extract the feature vector for the future lay the foundation.
用数字说话
4、Due to the face image on the nsitive image detection interference, this paper joins the face detection algorithm, it can remove the face image interference and choo to face higher detection accuracy of the algorithm--- Adaboost algorithm, to effectively improve the rate of face detection.
5、In skin color mask, the extraction of salient region feature vector as the input of the classifier, the feature vectors including color accounted for the whole image ratio、skin for external rectangle ratio、color with the number of connected regions、the largest connected area proportion of the largest connected area、color for skin external rectangular ratio、image center color region color ratio.
6、After extraction of feature vectors, then the classifier lection and design, the classification ability directly affects the performance of eroticism image recognition. In the choice of classifier, choo to solve small sample, nonlinear and high dimensional pattern recognition, classification problem has great advantages of the SVM classifier, and the experimental results are given.
7、In order to further improve the classification accuracy rate, construct PSO-SVM model
and GA-SVM model parameters of SVM were optimized, and the two kind of model of Matlab simulation experiment, the experimental results show that, the models than to simply u a SVM d
批评通报
etection effect especially accuracy is greatly enhanced, reached anticipated goal.
KEYWORDS: Sensitive image；skin detection；texture detection；face detection；feature extraction；support vector machine
