基于内容的大规模图像检索的设计与实现

【作者】 刘星；

【导师】 胡延琴；

【作者基本信息】 郑州大学 ， 计算机技术（专业学位）， 2016， 硕士

【摘要】 互联网和多媒体技术的快速发展,使得以图像为代表的多媒体数据呈现爆炸性的增长。在一个大规模图像集合中,快速地获取真正感兴趣的特定目标图像,尤其是那些很难用传统文本来清晰表达的图像,已经成为许多应用环境所要面对的日益重要的问题。在此背景下,本文设计并实现了基于内容的大规模图像检索原型系统。其主要研究内容及系统解决方案如下：综合比较研究两个主流的图像检索系统：完整特征-图像检索系统,用提取和压缩的方式,来存储图像数据库中所有图像的特征,并用目标图像的每一个特征精确检索数据库；词袋-图像检索系统,把基于量化的特征转变为视觉词汇,并用一个视觉词汇的直方图来表达图像。本文介绍了目前实现这两个系统所使用的几种方法：Kd-树(Kd-Trees),局部敏感哈希(Locality Sensitive Hashing)、分层K-均值(Hierarchical K-Means)、倒排表(Inverted File)、最小哈希(Min-Hash)。给出了这些方法在进行大规模图像检索时的不同性质及实验结果。实验结果表明,完整特征-图像检索系统占用更大的内存,但有较高的识别率；而词袋-图像检索系统虽然性能相对较低但内存使用情况良好。针对完整特征运行时间长和存储需求高的问题,提出了相应的大规模图像检索系统的解决方案：压缩Kd-Trees图像检索系统。该方法利用压缩的二进制标签来压缩图像局部特征,同时采用Kd-Trees进行快速最邻近搜索,从而减少数据库中存储的特征,大大压缩了特征的信息量。将该系统在4个实际图像数据实验场景中进行了运行实验,实验结果表明,压缩Kd-Trees,能够削减完整的特征表达在存储器中的使用量和运行时间,同时还能够实现较高的检索性能。针对词袋表达方法识别性能较低的问题,提出了相应的大规模图像检索系统的解决方案：多字典词袋-图像检索系统。该方法基于使用不同的独立字典的多视觉单词,而不是增加同一字典多单词的设想。通过在4个实际图像数据实验场景中运行该系统的实验,表明多字典词袋在增加存储和计算量的前提下,能够显著提高词袋的识别性能,并达到完整特征表达方法的水平。更多还原

【Abstract】 The rapid development of Internet and multimedia technologies in order to make an image as the representative of multimedia data showing explosive growth. Quick access to specific target images are really interested in a large set of images, especially those difficult to image using traditional text to articulate has become an increasingly important issue for many application environments to face. In this context, this paper designed and implemented a large-scale prototype image retrieval system based on content. The main research contents and system solutions of this article as follows:Comprehensive Comparative Study of Two mainstream image retrieval system:a complete feature-image retrieval system, characterized by extraction and compressed way to store images in the database of all the images, with each feature of the target image accurate searchable database; bag of words-Image Retrieval system, based on the quantitative characteristics into visual vocabulary, and a histogram of a visual vocabulary to express the image. Describes the current realization of several methods used by the two systems:Kd-Trees, Locality SensitiveHashing, Hierarchical K-Means, Inverted File, Min-Hash. Given the different nature and results of carrying out large-scale image retrieval methods. Experimental results show that the complete feature-image retrieval systems use more memory but has a high recognition rate; and the bag of words-although the performance of image retrieval system memory usage is relatively low but good.For a complete feature high long running time and storage requirements of the problem, we propose a corresponding large-scale image retrieval system solution: compression Kd-Trees image retrieval system. The method uses a compression binary label to compress the image local features, while using Kd-Trees for fast nearest neighbor search, thereby reducing the characteristics stored in the database, the compression characteristics of information. The system runs on four actual image data test scenarios The experimental results show that the compression Kd-Trees can reduce the expression of full-featured memory usage and run-time method, and can achieve higher retrieval performance.For the lower bag of words express way to identify performance problems, we propose the corresponding large-scale image retrieval system solution: multi-dictionary word pocket-image retrieval system. The method is based more visual words using different independent dictionaries, rather than adding the same idea more word dictionary. Runs in four actual image data in the system test scenarios, test results show that the multi-dictionary word bags at an increased storage and computing capacity of the premise, can significantly boost recognition performance bag of words reaches full-featured level of expression methods.更多还原