基于社会媒体信息挖掘的图像标注技术研究

项目名称： 基于社会媒体信息挖掘的图像标注技术研究

项目编号： No.61272214

项目类型： 面上项目

立项/批准年度： 2013

项目学科： 自动化技术、计算机技术

项目作者： 孙福明

作者单位： 辽宁工业大学

项目金额： 76万元

中文摘要： 随着社会媒体网站的兴起，用户共享的图像、文本和视频等多媒体数据呈爆炸式增长。如何利用海量的社会媒体信息，解决大规模训练样本不足，实现图像的自动标注，是当前计算机图像与视频处理领域的一个研究热点。本课题将基于机器学习框架，通过社会媒体信息挖掘，深入研究3个关键问题：弱标注网络图像的标签优化、图像特征的有效表征及概念分类器学习。针对这些问题，本课题拟首先深入研究标签优化技术，基于数据驱动和主动学习构建一个大规模的训练样本库；然后基于该库，通过挖掘图像内容连接和上下文连接所蕴含的语义信息，建立图像特征的鲁棒表示方法；进而研究概念分类器学习技术,提出一种基于上下文信息挖掘的离线学习算法和一种基于群体智慧驱动的在线学习算法；最后，构建一个标注算法验证系统。本课题将有利于推动图像语义理解理论和应用的发展，为新一代网络多媒体服务和海量图像管理提供核心算法。

中文关键词： 社会媒体；图像标注；机器学习；图像语义理解；图像特征

英文摘要： With the popularity of social media websites, multimedia data shared by users, such as image, video and text, are growing explosively. Therefore, exploring large-scale social media information to tackle the issue of deficiency of training data in automatic image annotation has become a hot topic recently and is receiving increasing research interests. In this project, we will focus on three key issues of image annotation based on machine learning framework: (1) the refinement of user-provided noisy tags of images; (2) the effective representation of image content; and (3) the learning of concept classification models. To this end, we plan to build a large training set using data-driven and active learning techniques. Based on this dataset, we will then investigate the effective and robust image representation by mining image content and context. Further,we study concept classification models. As a result, we will propose an offline learning method using context information mining and an online learning algorithm that can learn models based on so called collective intelligence. Finally, we will build a demo system based on the techniques and the data. Our research will push forward the theory and application of image annotation, and will provide supporting algorithms for the next-generation multimedia services an

英文关键词： social media；image annotation；machine learning；image semantic understanding；image feature