基于Sparse Coding和DBN的敏感图像检测  

Nude Image Detection based on Sparse Coding and DBN

作　　者：陈亚楠[1] 黄豫蕾 唐麟 王士林[1] 

机构地区：[1]上海交通大学信息安全工程学院,上海200240 [2]上海数据分析与处理技术研究所,上海201112

出　　处：《信息安全与通信保密》2016年第1期113-118,共6页Information Security and Communications Privacy

基　　金：国家自然科学基金(No.61271319)

摘　　要：敏感图像检测,即检测图片是否含有危害青少年健康成长的不良色情信息,对于净化网络环境有重要意义。该文分析了现有的敏感图像检测算法的性能,结合稀疏编码和深度信赖网络,提出了一种改进的敏感图像检测算法。该算法通过稀疏编码来提取特征,将图像切分成标准大小的小图块,然后将其基于字典稀疏表示。接着用max-pooling池化来整合特征,获得最终使用的特征向量。将得到的特征向量输入到DBN网络中进行训练,得到DBN模型。最后将待测图像的特征向量输入到DBN模型中获得分类结果。在文献[10]的数据集上的实验显示,该检测算法较原有算法有较大提升,在以总样本的90%作为训练集时,可获得9.29%的平均错误率。Nude image detection, i.e., detecting whether there is any pornographic information harmful to teenagers in images, is of great significance for purifying network. Performance of existing nude detection algorithms is analyzed, and in combination with sparse coding and deep belief network, a modified nude detection algorithm is proposed in this paper. In this algorithm, feature extraction is done by sparse coding. Specifically, the image is cut into small standard patches, which could be represented sparsely based on a dictionary. Then, max-pooling is used to integrate the features, thus to get the final feature vectors. Next, the obtained feature vectors are input into DBN network for training, thus to get the DBN model. Finally, the feature vectors of the images are input into DBN model to get the classified results. Experiments on the evaluation dataset of article [10] shows that this algorithm enjoys higher performance than that of the original algorithm and achieves an average error rate of 9.29% when 90% of the total samples is used as the training set.

关 键 词：敏感图像 稀疏编码 深度信赖网络 受限玻尔兹曼机 池化 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]