【作者】 陈烨；

【导师】 庄越挺； 吴飞；

【作者基本信息】 浙江大学 ， 计算机应用技术， 2010， 硕士

【摘要】 随着网络多媒体技术的快速发展,互联网上的图像等多媒体内容的数量正在以指数级的速度迅猛增长。因此,实现大规模互联网图像的有效管理和检索具有十分重要的现实意义。由于大多数互联网图像标签丢失或标签存在大量噪音,因此研究对这些弱标注(weakly-tagged)互联网图像自动添加标签成为当前热点研究问题。本文首先针对Flickr用户经常会根据上传图像所隐含主题而将其推荐到多个相关社群这一特点,提出了基于社群隐含主题挖掘和多社群信息融合的自动图像标注算法。该算法采用隐Dirichlet分配模型对单个社群中隐含主题进行挖掘,根据候选标签与社群隐含主题之间相关性,对初始候选标注标签进行过滤和排序,最终通过多层次多社群主题信息融合,得到标注结果。对于从Flicrk网站下载的三个社群图像进行实验的结果表明该算法能很大程度提高自动图像标注精度。同时,为了有效辅助用户添加标签,本文提出了结合社群文本、图像和用户上下文信息的个性化标签推荐算法。该算法首先建立由用户、图像、标签组成的三元矩阵,然后在社群中寻找与待标签推荐用户兴趣相近用户以及与待标签推荐图像相似的图像得到用户的个人偏好,最后利用随机游走机制对标签进行排序,得到推荐结果,该方法具有通用性,即只要给出用户、图像和标签中任何一个元素,均可得到标签推荐结果。更多还原

【Abstract】 Nowadays, the number of internet images is growing at an exponential rate. Therefore, how to effectively manage and retrieve large scale Internet images put forth a great challenge. Since a great number of images uploaded onto Internet do not have any labels, or has limited labels with noise, automatic annotation of such "weakly-tagged" Internet images has been a hot topic recently.Since users intend to recommend images to multiple social groups according to semantics of images when they upload images into Flicker, this paper proposes a two-stage approach to automatically annotate weakly-tagged social images. The first stage discovers the latent topics in each group by Latent Dirichlet Allocation(LDA) model, and filters out noisy tags in group level in order to re-rank topic-relevant tags. The second stage discovers the hierarchical topic structure among multiple groups by WordNet, and hierarchically fuses the candidate tags from multiple groups.This paper also proposes an approach to integrate social text, image and user context for tag recommendation. This approach sets up a ternary matrix to represent the relationship among users, images and tags at first; and get a personal preference by discovering users with similar interest and images with similar visual similarity at second, and finally utilizes random walk to recommend tags for unlabeled images. This tag recommend approach is very flexible, since we can get recommendation result once any one of information about a user, an image, or a tag is offered.更多还原