基于内容的概念建模和图像检索重排序

发布时间：2018-03-11 03:30

本文选题：图像检索　切入点：聚类　出处：《复旦大学》2014年硕士论文　论文类型：学位论文

【摘要】：随着互联网的发展,多媒体技术(图像、视频、音频等)在互联网中的广泛应用。近些年来,数字化图像网络资源呈指数级增长。社交网络的出现,如Facebook, Flickr, Twitter, Instagram, Whatsapp,人人网,微博,微信等,更加推动了这一过程的大众化。用户在这个过程中扮演着两种角色,一种是多媒体资源的提供者,另一种是多媒体资源的消费者。用户通过互联网获取自己想要的多媒体资源。在这个过程中,检索技术的出现提高了用户获得资源的效率。对图像检索来说,用户如何快速并准确地检索到自己想要的图像。这个课题成为近些年研究者们关切的热点话题。但是,对于图像搜索引擎返回给用户的图像当中,其中有部分并不是用户期望的图像。本文基于这个出发点,对搜索引擎返回给用户的图像进行重新排序。我们期望通过自己提出的方法,将尽可能多的与用户期望的图像置于重排序列表的前部。最后,将重排序好新的图像列表返回给用户,从而提高检索结果的相关度和用户体验。本文主要做了以下几个方面的研究工作：首先,我们对图像检索重排序进行了定义。并给出了与之相关的图像处理基础知识,对图像的视觉特征也进行了简单的介绍。同时,对最近几年出现的图像检索重排序算法进行了系统的归类。其次,我们提出了一种自适应选择的图像检索重排序方法,对不同类别的图像分别采用不同的图像视觉特征进行重排序。这个过程中,我们采用了图像的各种特征,如颜色,纹理,及局部SIFT特征。这种方法得到了比图像检索更好的效果。再次,我们提出了一种用于图像检索重排序的概念模型。该方法采用初始检索列表中前N幅图像对检索关键字(概念)进行建模,然后运用概念模型对图像检索进行重排序。在这个过程中我们采用了图像的局部SIFT特征。在概念建模阶段,我们采用了增量式学习方法。这个方法也得到了比仅通过分类和聚类方法更好的结果。最后,我们对本方进行了简单的总结,并对将来进一步研究工作做了展望。
[Abstract]:With the development of the Internet, multimedia technology (image, video, audio etc.) is widely used in the Internet. In recent years, digital image network resources have increased exponentially. Social networks, such as Facebook, Flickr, Twitter, Instagram. Whatsapp. Weibo, WeChat, etc., have further promoted the popularity of this process. Users play two roles in this process, one is the provider of multimedia resources, The other is the consumers of multimedia resources. Users obtain the multimedia resources they want through the Internet. In this process, the appearance of retrieval technology improves the efficiency of obtaining resources. How to retrieve the image users want quickly and accurately has become a hot topic for researchers in recent years. However, for the images returned to users by image search engines, Based on this, we reorder the images returned by the search engine to the user. Put as many images as you want in front of the reorder list. Finally, the reordered new image list is returned to the user. In order to improve the relevance of retrieval results and user experience, this paper mainly does the following research work: firstly, we define the image retrieval reordering, and give the basic knowledge of image processing. The visual features of images are also introduced briefly. At the same time, the image retrieval reordering algorithms that have appeared in recent years are systematically classified. Secondly, we propose a self-adaptive image retrieval reorder method. Different types of images are reordered with different visual features. In this process, we use various features of the image, such as color, texture, And local SIFT features. This method has better results than image retrieval. Again, We propose a conceptual model for image retrieval reordering, in which the first N images in the initial retrieval list are used to model the retrieval keywords (concepts). Then we use the concept model to reorder the image retrieval. In this process, we adopt the local SIFT feature of the image. We adopt the incremental learning method, which is better than the classification and clustering methods. Finally, we summarize this method briefly and make a prospect for the future research work.
【学位授予单位】：复旦大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP391.41

【相似文献】