基于随机游走的图像多样性搜索重排序算法研究

发布时间：2019-05-15 17:00

【摘要】：近些年来随着互联网的发展，网络上出现的数据呈现出爆炸式的增长，特别是图像数据。图像具有直观、生动的特性，，且与人类生活息息相关，因此如何能够准确地获取所需图像信息是当前多媒体搜索领域需要重点解决的问题。目前主流的搜索引擎主要是依靠图像的标题、描述等元数据信息进行基于文字信息的图像搜索，并没有利用图像的视觉特性，因此搜索结果并不理想。为了使结果更加符合用户的需求，图像搜索重排序算法被提出应用于图像搜索领域。该技术是指在原始文本搜索的基础之上，再次结合图像本身的信息对原始的重排序结果进行重新重排序，以提高用户的搜索体验。目前的大多工作围绕图像的相关性搜索重排序展开，对多样性搜索重排序的研究较少。相关性重排序是指对于返回结果来说，都是将相关的样本排在结果集的前列，但是很少考虑样本相互之间的联系，而多样性重排序是为了提高返回结果的多样性，也就是覆盖主题的多样性，以降低相关重排序中带来的信息冗余问题。本文主要利用随机游走模型实现图像搜索的多样性重排序。具体工作如下： 1）实现了基于图像的搜索相关性重排序算法，能够使产生的结果按照相关性进行重排序。 2）提出了基于DivRank算法的图像多样性重排序算法。DivRank是自定义的随机游走模型，通过引入访问次数随时间增长的马尔科夫数学模型来完成多样性重排，能够有效的降低搜索结果之间的冗余信息。为此，我们将其应用于图像领域并提出相应的GDRID算法。 3）提出了基于DDRank的图像多样性重排序算法。DDRank是基于DivRank算法的改进，它通过修改初始重排序的权值来降低DivRank所带来的永久破坏原始信息网络结构的副作用，能够在保证多样性的基础上兼顾相关性。为此，我们将其应用在图像领域并提出相应的DIR算法。通过在MSRA_V-1.0等数据库上的大规模实验，表明所提GDRID和DIR算法能够在有效保持图像搜索相关性的前提下，提高多样性性能。
[Abstract]:In recent years, with the development of the Internet, the data on the network presents explosive growth, especially image data. The image has the characteristics of being intuitive and vivid, and is closely related to human life, so how to accurately acquire the required image information is a problem that needs to be solved in the current multimedia search field. The present mainstream search engine mainly relies on the title, description and other metadata information of the image to carry out the image search based on the character information, and does not utilize the visual characteristic of the image, so the search result is not ideal. In order to make the results more consistent with the user's needs, the image search re-ordering algorithm is proposed to be applied in the field of image search. The technology is based on the original text search, and the original reordering result is re-ordered again in combination with the information of the image itself to improve the user's search experience. Most of the current work is about the relevance search and re-ordering of the images, and the research on the re-ordering of the diversity search Less. The relevance reordering means that for the return results, the relevant samples are at the top of the result set, but the contact between the samples is rarely taken into account, and the diversity reordering is to improve the diversity of the return results, that is, the diversity of the coverage topics Sex in order to reduce the information redundancy in the related re-ordering In this paper, the diversity of image search is realized by using random walk model. Reorder. Specific as follows:1) a search correlation re-ordering algorithm based on an image is realized, and the generated result can be matched with each other according to the correlation, The reordering is carried out.2) An image is proposed based on the DivRank algorithm. The sample reordering algorithm. DivRank is a self-defined random walk model, which can effectively reduce the search result by introducing the number of access times and the time-increasing Markov model to complete the diversity rearrangement. The redundant information between them. For this, we apply it to the image field and propose the appropriate The GDRID algorithm is presented in this paper. DRank is an improvement based on the DivRank algorithm, which can reduce the side effect of the original information network structure caused by DivRank by modifying the weight of the initial reordering, and can guarantee the diversity. On the basis of the correlation, we apply it in the image field and mention it The corresponding DIR algorithm is presented. The results show that the proposed GDRID and DIR algorithm can keep the image search correlation effectively by the large scale experiment on the database of MSRRA _ V-1.0
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP391.41

【共引文献】