基于社交圈的移动设备个性化图像标注
发布时间：2019-05-06 08:48
【摘要】：近年来,随着用户移动设备中照片数量的爆炸性增长,用户组织、管理、检索这些个人照片变得十分困难。自动为这些图像提供标签,是解决该问题的一个有效途径。不同于传统基于内容的图像标注算法,对于用户移动设备上的图像标注,用户更关注于图像的上下文信息。另一方面,随着社交网络和wifi/4G网络的发展,大量的用户图像被上传到社交平台上。这些图像基本都是拍摄于用户的日常生活中,因此用户的社交圈信息可以为理解图像的上下文信息提供有用的线索。在本文中,我们提出了一种基于用户社交圈的移动设备上的图像个性化标注框架。但是,由于社交网络上的信息具有稀疏性和不准确性,使得直接利用社交网络进行图像标注的效果并不好。为了解决这个问题,我们基于对社交网络特性的研究,在为用户个人照片提供标签之前,先为社交网络上的图像标签进行修正,产生可靠的标签。考虑到事件是人们生活和记忆中最重要的一种组织方式,用户上传到社交网络中的大多数照片都是拍摄于特定的事件,而相同的事件具有同样的事件属性标签,因此我们提出在社交网络中发现事件,为图像产生事件标签。我们采用基于多模态的层次聚类算法来发现社交网络中的事件。同时,我们创新性地引入了“相册”的概念,并将其作为聚类的基本单元,而不是通常的以单张图像作为聚类的基本单元。最后,我们基于相似的图像更有可能有相同的标签这一思想,采用加权的K近邻模型将修正后社交网络中图像的标签传播给用户移动设备上的待标注图像,从而为用户的个人照片自动产生个性化的标注。为了验证我们对于社交网络特性的观察和分析,我们在来源于Flickr的公开数据集ReSEED和我们自己在人人网上抓取的真实数据集两个数据集上对社交网络各方面的特性进行研究分析,结果证实了我们的观察和假设。同时,我们也在数据集上对我们提出的算法进行了测试,并与其它算法进行比较,实验结果显示我们的算法有着较好的性能。
[Abstract]:In recent years, with the explosive increase in the number of photos in user mobile devices, it is very difficult for users to organize, manage and retrieve these personal photos. Automatically providing labels for these images is an effective way to solve this problem. Different from the traditional content-based image tagging algorithm, users pay more attention to the context information of the image than the traditional content-based image tagging algorithm. On the other hand, with the development of social networks and wifi/4G networks, a large number of user images are uploaded to social platforms. These images are basically photographed in the daily life of the user, so the social circle information of the user can provide a useful clue to understand the context information of the image. In this paper, we propose a personalized image annotation framework for mobile devices based on user social circle. However, because of the sparsity and inaccuracy of the information on the social network, it is not good to use the social network to annotate the image directly. In order to solve this problem, based on the study of the characteristics of social networks, we first modify the image tags on social networks to produce reliable tags before providing tags for users' personal photos. Considering that events are one of the most important forms of organization in people's lives and memories, most of the photos that users upload to social networks are taken on specific events, and the same events have the same event attribute labels, Therefore, we propose to discover events in social networks and generate event tags for images. We use multi-modal hierarchical clustering algorithm to discover events in social networks. At the same time, we introduce the concept of "photo album" creatively, and regard it as the basic unit of clustering, instead of taking a single image as the basic unit of clustering. Finally, based on the idea that similar images are more likely to have the same tags, we adopt a weighted K-nearest neighbor model to propagate the tags of the images in the modified social network to the images to be labeled on the user's mobile devices. Thus for the user's personal photos automatically generate personalized tagging. To validate our observations and analysis of social networking features, we studied and analyzed various aspects of social networking features on two datasets, ReSEED, an open data set from Flickr, and a real dataset we ourselves captured on Renren. The results confirm our observations and assumptions. At the same time, we also test the proposed algorithm on the data set and compare it with other algorithms. The experimental results show that our algorithm has a good performance.
【学位授予单位】：清华大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41
本文编号：2470041
[Abstract]:In recent years, with the explosive increase in the number of photos in user mobile devices, it is very difficult for users to organize, manage and retrieve these personal photos. Automatically providing labels for these images is an effective way to solve this problem. Different from the traditional content-based image tagging algorithm, users pay more attention to the context information of the image than the traditional content-based image tagging algorithm. On the other hand, with the development of social networks and wifi/4G networks, a large number of user images are uploaded to social platforms. These images are basically photographed in the daily life of the user, so the social circle information of the user can provide a useful clue to understand the context information of the image. In this paper, we propose a personalized image annotation framework for mobile devices based on user social circle. However, because of the sparsity and inaccuracy of the information on the social network, it is not good to use the social network to annotate the image directly. In order to solve this problem, based on the study of the characteristics of social networks, we first modify the image tags on social networks to produce reliable tags before providing tags for users' personal photos. Considering that events are one of the most important forms of organization in people's lives and memories, most of the photos that users upload to social networks are taken on specific events, and the same events have the same event attribute labels, Therefore, we propose to discover events in social networks and generate event tags for images. We use multi-modal hierarchical clustering algorithm to discover events in social networks. At the same time, we introduce the concept of "photo album" creatively, and regard it as the basic unit of clustering, instead of taking a single image as the basic unit of clustering. Finally, based on the idea that similar images are more likely to have the same tags, we adopt a weighted K-nearest neighbor model to propagate the tags of the images in the modified social network to the images to be labeled on the user's mobile devices. Thus for the user's personal photos automatically generate personalized tagging. To validate our observations and analysis of social networking features, we studied and analyzed various aspects of social networking features on two datasets, ReSEED, an open data set from Flickr, and a real dataset we ourselves captured on Renren. The results confirm our observations and assumptions. At the same time, we also test the proposed algorithm on the data set and compare it with other algorithms. The experimental results show that our algorithm has a good performance.
【学位授予单位】：清华大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41
【参考文献】
相关期刊论文 前1条
1 路晶;马少平;;使用基于多例学习的启发式SVM算法的图像自动标注[J];计算机研究与发展;2009年05期
，本文编号：2470041
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2470041.html
