基于颜色和SIFT特征的图像检索技术及其分布式实现

发布时间：2018-07-31 14:37

【摘要】：图像检索已成为获取信息的重要手段之一,如何快速准确地从海量图像中获取所需内容成为图像检索发展的主要瓶颈。因此本文主要研究如何选择图像特征,设计检索算法,构建图像检索系统和提升系统性能。本文的工作可以分为图像特征分析、检索算法设计和基于Hadoop平台的并行化实现三个部分。本文设计实现了一种基于颜色相关图和SIFT特征的图像检索算法,在此基础上利用DBSCAN聚类算法把SIFT特征提取和匹配限定在一定范围内,并借助Hadoop大数据处理框架构建了基于内容的图像检索系统。本文首先梳理了图像检索技术的发展及成果,探讨了基于文本、内容、高层语义的图像检索技术,分析了他们的优缺点和适用场景。其次,为了提高检索的准确度,本文选择融合颜色自相关图和SIFT特征。在此基础上,本文利用基于密度的DBSCAN聚类算法对64维颜色特征进行聚类,找到与样例图像距离最近的类簇,然后在这个类簇范围内进行SIFT特征的提取和匹配以降低算法时间复杂度。考虑到当两张图片的特征点间欧氏距离整体偏大时,传统的基于SIFT特征点匹配比例的相似性度量方式会丢失一定的空间信息,本文使用SIFT特征匹配点的平均欧氏距离作为相似性度量依据。最后本文实现了基于MapReduce的综合特征提取和匹配,并利用AGD-DBSCAN算法中寻找数据集自适应邻域半径和邻域最小点数的思想,实现了 DBSCAN聚类过程的MapReduce化。本文分别对算法的查准率、查全率和Hadoop框架下的系统加速比、效率、扩展率做了评估,验证了本文算法的可用性和可扩展性。
[Abstract]:Image retrieval has become one of the important means to obtain information. How to quickly and accurately obtain the required content from massive images has become the main bottleneck in the development of image retrieval. Therefore, this paper mainly studies how to select image features, design retrieval algorithm, build image retrieval system and improve system performance. The work of this paper can be divided into three parts: image feature analysis, retrieval algorithm design and parallel implementation based on Hadoop platform. In this paper, an image retrieval algorithm based on color correlation graph and SIFT feature is designed and implemented. On this basis, SIFT feature extraction and matching are limited to a certain range by using DBSCAN clustering algorithm. A content-based image retrieval system is constructed with the help of Hadoop big data processing framework. This paper firstly combs the development and achievement of image retrieval technology, discusses the image retrieval technology based on text, content and high-level semantics, and analyzes their advantages and disadvantages and applicable scenarios. Secondly, in order to improve the accuracy of retrieval, this paper selects the fusion of color autocorrelation and SIFT features. On this basis, we use the density-based DBSCAN clustering algorithm to cluster the 64-dimensional color features, and find the cluster closest to the sample image. Then the SIFT features are extracted and matched in this cluster to reduce the time complexity of the algorithm. Considering that when the Euclidean distance between the feature points of two images is relatively large, the traditional similarity measurement method based on the matching ratio of SIFT feature points will lose some spatial information. In this paper, the average Euclidean distance of SIFT feature matching points is used as the basis of similarity measurement. Finally, this paper realizes the synthesis feature extraction and matching based on MapReduce, and makes use of the idea of finding the adaptive neighborhood radius and the minimum number of neighborhood points in the AGD-DBSCAN algorithm to realize the MapReduce of the DBSCAN clustering process. In this paper, the recall rate, recall rate and system speedup, efficiency and expansion rate of the algorithm under Hadoop framework are evaluated, and the availability and extensibility of the algorithm are verified.
【学位授予单位】：东南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】