利用稀疏语义结合双层深度卷积神经网络的敏感图像检测方法  

Sensitive image detection method based on deep neural network and sparse semantics

作　　者：如先姑力·阿布都热西提[1] 亚森·艾则孜[1] 孙国梓[2] Ruxianguli·Abudurexiti;Yasen·Aizezi;Sun Guozi(Dept.of Information Security Engineering,Xinjiang Police College,Urumqi 830013,China;Institute of Computer Technology,Nanjing University of Posts&Telecommunications,Nanjing 210003,China)

机构地区：[1]新疆警察学院信息安全工程系,乌鲁木齐830013 [2]南京邮电大学计算机技术研究所,南京210003

出　　处：《计算机应用研究》2020年第5期1557-1560,1565,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61762086);新疆维吾尔自治区高校科研计划面上项目(XJEDU2017M046)。

摘　　要：互联网技术的飞速发展导致敏感内容图像由原先基本隐蔽的内容交换变为海量的数据共享,传统基于图像特征提取的敏感内容检测方法不再适用。针对上述难点,提出基于稀疏语义和双层深度卷积神经网络相结合的敏感内容检测方法。上层网络首先进行训练样本的预处理,并通过构造图像的稀疏语义表示作为神经网络的输入;而下层网络则进一步考虑第三方管控机制(如政府代理等),提出针对特定群体的敏感内容图像检测方法。与现有常用敏感内容图像检测方法相比,该检测方法可有效降低训练样本数量,且检测精度比传统图像检测方法(如基于视觉词袋方法等)提升7%以上。With the rapid development of Internet technology,sensitive content images have changed from basic concealed content exchange to mass data sharing.The traditional method of sensitive content detection based on image feature extraction is no longer applicable.To overcome these difficulties,this paper proposed a sensitive content detection method based on sparse semantics and double-layer deep convolution neural network.In this method,the upper network preprocessed the training samples and constructed sparse semantic representation of the image as the input of the neural network,while the lower network further considered the third-party control mechanism(such as government agents)and proposed a sensitive content image detection method for specific groups.Compared with the existing image detection methods for sensitive content,this method can effectively reduce the number of training samples,and the detection accuracy is more than 7%higher than that of traditional image detection methods(such as visual word bag method).

关 键 词：敏感图像内容检测 双层卷积神经网络 深度学习算法 稀疏语义表示 视觉词袋 皮肤检测器 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]