基于多视角的网络暴力敏感图像识别技术研究

The network behaviors under the Internet environment have become the focus of our society, especially the spread of network violent information has drawn great attentions of government and society. Among all network violent information, violent sensitive images on the network have been the main transmission carrier. Because violent sensitive images lack the sequential movement characteristics of violent videos while the contents of images have stronger pertinence, the researches on violent sensitive images face great challenges. Now, research on violent sensitive images recognition is not sufficient. The deep analysis on images with complex contents is difficult to conduct. This project aims to study the feature extraction, dimension reduction and multi-view classification on violent sensitive images under the Internet environment. First, we describe the consistent regions using connected granules and extract the global features of violent sensitive images. Some other feature extraction methods are combined with our method to construct multi-view features database. Then, this project constructs a subspace using sparse relationships between features and obtains the low-dimensional representations of multi-view features. Finally, we integrate classification model with multi-view learning to obtain a multi-view classification model which can achieve multi-view image recognition on violent sensitive images. Meanwhile, this project aims to construct a multi-view violent sensitive images recognition system and provides a feasible method to identify the violent sensitive images under the Internet environment, which is conducive to eliminate the spread of network violent information.

互联网环境下的网络行为已成为社会关注的焦点问题，尤其是网络暴力信息的传播已引起政府和社会的高度重视。网络中暴力敏感图像是网络暴力信息传播的主要形式。暴力敏感图像缺乏暴力视频的时序运动特征，且图像内容针对性较强，这成为暴力敏感图像研究的难点。目前，针对暴力敏感图像的识别研究工作尚不充分，难以对复杂图像进行深入分析。本项目在互联网背景下研究网络中暴力敏感图像的特征提取、降维及多视角分类问题，最终实现网络中暴力敏感图像的识别和筛选。首先，利用连通粒描述图像的一致性区域，对网络图像进行全局特征分析和提取，并联合其他图像特征提取技术构建多视角特征库。然后，利用数据的稀疏结构实现图像多个特征的降维，得到图像特征的低维表示。最后，利用多视角思想，结合分类模型实现暴力敏感图像识别。本项目拟搭建基于多视角的网络暴力敏感图像识别系统，提供网络中暴力敏感图像识别的可行方案，以清除互联网暴力信息传播。

互联网环境下的网络行为已成为社会关注的焦点问题，尤其是网络暴力信息的传播已引起政府和社会的高度重视。网络中暴力敏感图像是网络暴力信息传播的主要形式。暴力敏感图像缺乏暴力视频的时序运动特征，且图像内容针对性较强，这成为暴力敏感图像研究的难点。目前，针对暴力敏感图像的识别研究工作尚不充分，难以对复杂图像进行深入分析。四年以来，本项目在互联网背景下研究网络中暴力敏感图像的特征提取、降维及多视角分类问题，最终实现网络中暴力敏感图像的识别和筛选。首先，利用连通粒描述图像的一致性区域，对网络图像进行全局特征分析和提取，并联合其他图像特征提取技术构建多视角特征库。然后，利用数据的稀疏结构实现图像多个特征的降维，得到图像特征的低维表示。最后，利用多视角思想，结合分类模型实现暴力敏感图像识别。本项目拟搭建基于多视角的网络暴力敏感图像识别系统，提供网络中暴力敏感图像识别的可行方案，以清除互联网暴力信息传播。本项目共发表学术论文28篇，其中23篇被SCI索引，2篇被SSCI索引。项目在暴力敏感图像处理的各个方面均提供了切实可行的方法，促进了当下大数据环境下网络暴力敏感的图像识别工作。