基于内容的分布式图像检索

发布时间：2018-02-05 22:39

本文关键词： 对等网络分布式哈希表云计算图像检索局部敏感哈希相关反馈　出处：《北京邮电大学》2015年博士论文　论文类型：学位论文

【摘要】：移动互联网的迅猛发展和移动智能产品的迅速推广,使得新型应用风靡全球,促使互联网中的图像资源呈现指数级爆炸式增长,催生了以“大数据”为特色的网络数据时代。大数据时代以及云计算技术为基于图像检索的应用研究带来了新的机遇和挑战。图像数量的快速增长使得传统的以集中式索引模式构建的数据中心面临低效、扩展性差等问题。近几年来,依据覆盖网络(Overlay network)的思想,借助分布式哈希表(Distributed Hash Table,DHT)技术为数据中心构建一个纯分布式架构成为一种新型解决方案,更加适用于海量资源的检索和管理需求。在基于DHT的结构化分布式系统中,每个结点只负责部分资源索引和小范围路由信息,实现整个分布式网络的寻址和资源快速定位。不同于基于关键字的图像检索,基于内容的图像检索技术能够不借助文本描述找出与查询示例内容相似的图像,更适宜于图像检索,已成为近年来图像检索的主流技术。因此,如何结合DHT检索机制,实现基于内容的图像快速检索和匹配,已成为当今重要的研究课题。本文提出了一种基于内容的大规模分布式图像检索框架,可以针对不同类型的图像生成相应的分布式索引,并采用覆盖网路由协议发布索引和定位查询。该框架基于自组织的云计算环境构建,便于大规模部署,有效提高了基于内容图像检索系统的可扩展性和容错性。实现该框架包括如下关键问题：如何根据图像视觉特征构造分布式索引,并根据用户反馈修改查询；如何在保障负载均衡的情况下,将索引均匀地发布到各个结点上；如何有效降低通信代价并保障查询效率等。围绕上述问题,本文的创新工作概括如下：(一)提出了针对图像纹理特征的大规模分布式检索系统LRFIR (LSH-based Relevance Feedback for Image Retrieval),既支持内容相似性检索也支持语义相似检索。LRFIR运用了多基元纹理直方图进行图像特征提取,利用一组局部敏感哈希函数,将纹理相似的特征以较高的概率映射到同一个哈希值,从而生成了图像的分布式纹理索引。DHT使得查询请求只发送到可能含有该纹理特征的结点,降低了通信代价且保证了查询准确率。(二)采用信息检索中的相关反馈机制克服低级特征和高级特征之间的鸿沟,实现分布式环境下图像内容的语义查询。LRFIR相关反馈允许用户在结果中选择与查询相关和不相关的图像以帮助系统不断更新查询,使查询不断接近用户语义。本文采用分别图像库Corel和Caltech 101 Object Categories进行了验证,结果表明LRFIR能够在较少的跳数内完成检索,并保证系统的负载均衡。(三)提出了针对图像融合特征的高扩展性分布式检索框架LFFIR (LSH-based Fusion Features for Image Retrieval)。LFFIR利用图像融合特征的多特征性,从多个角度捕捉图像内容并将其引入DHT,更适合于检索内容丰富的自然景观图像。LFFIR包含索引构造服务和查询处理服务,前者利用融合特征构造图像的分布式多特征索引,并将相似图像的索引聚集到同一结点；后者负责处理查询图像,发布查询消息,并将消息路由到最可能应答查询的结点。与M-Chord的对比实验证实,LFFIR能以较低的查询代价,完成准确的检索。(四)针对图像的局部特征,提出了基于词袋模型的分布式图像检索框架BVWIR (Bag of Visual Words based Image Retrieval),将词袋模型整合到DHT中,利用局部敏感哈希函数特性为局部特征聚类并自动生成视觉词典,更适合于对物体图像的检索。BVWIR将相似图像块以极高的概率发布到相同的视觉词汇,而不需要任何全局信息。同时,为降低多个局部特征引起的高额查询代价,该框架包含了两种消息传输模式,并行模式和串行模式,以降低查询代价且保证准确率。
[Abstract]:The rapid promotion of the rapid development of mobile Internet and mobile intelligent products, make the new application all over the world, prompted the exponential explosion of image resources in the Internet, spawned a network data in "big data" as the characteristics of the era. The era of big data and cloud computing technology application research based on image retrieval has brought new the opportunities and challenges. The rapid growth of the number of images that to construct the centralized index mode of the traditional data center facing the problem of inefficient, poor scalability and so on. In recent years, based on the overlay network (Overlay network) thought, by means of a distributed hash table (Distributed Hash Table DHT) technology is a pure distributed architecture has become a a new solution for data center construction, retrieval and management needs is more suitable for the massive resources. Based on the structured distributed DHT system, each node only Responsible for part of the resource index and small range of routing information, and realize the distributed network resources addressing and fast positioning. Different from the keyword based image retrieval technology can find text description and query example content similar to the image without the aid of content-based image retrieval is more suitable for image retrieval, has become a mainstream technology in recent years, image retrieval therefore, how to combine the DHT retrieval mechanism, realize content-based image retrieval and matching based on fast, has become an important research topic today. This paper presents a large-scale distributed content image retrieval framework based on, can be for different types of images to generate the corresponding distributed index, and using the overlay routing protocol and query index released positioning. The framework self organizing cloud computing environment based on large-scale deployment, effectively improves the image retrieval based on content The system scalability and fault tolerance. The framework includes the following key problems: how to construct the image visual features according to the distributed index, and according to user feedback on how to modify the query; load balancing security situation, the index released evenly to every node; how to effectively reduce the communication cost and ensure the query efficiency of around. The above problems, the innovation of this work are summarized as follows: (a) is proposed for large-scale distributed image retrieval system LRFIR (LSH-based Relevance Feedback for Image Retrieval), which supports the content similarity retrieval also supports semantic similarity retrieval using.LRFIR multi textons histogram for image texture feature extraction, using a set of local sensitive hash function and texture similar characteristics with high probability mapping to the same hash value, thereby producing a distributed image The texture index.DHT makes the query request only sent to the node may contain texture features, reduce the communication cost and ensure the accurate query. (two) used in information retrieval and relevance feedback mechanism to overcome the gap between low-level features and high-level semantic features, realize the distributed environment image content relevance feedback allows the user to query.LRFIR select the image and query related and unrelated in the results to help the system to continuously update query, the query is close to the user. This paper uses the semantic image library of Corel and Caltech were 101 Object Categories were verified, the results show that LRFIR can achieve retrieval in less hops, and ensure the load balancing system (three.) is proposed for distributed scalable image fusion feature retrieval framework LFFIR (LSH-based Fusion Features for Image Retrieval.LFFIR) using image fusion Features of photosynthetic characteristics, capture the image content from multiple angles and the introduction of DHT, more suitable for retrieval of natural landscape images with rich.LFFIR contains the index construction and query processing service service, the former uses the distributed fusion structure of the image of the characteristics of multi feature index, and the similarity index of the image gathered in the same node; the latter is responsible for processing the query image, issued the query message, and the message is routed to the most likely response query nodes. Confirmed the experimental results compared with the M-Chord, LFFIR to lower the cost of query, complete and accurate retrieval. (four) according to the local characteristics of the image, the image retrieval framework of distributed BVWIR model based on bag of words (Bag of Visual Words based Image Retrieval), the bag of words model integrated with DHT, using the function characteristic of locality sensitive hashing for local feature clustering and automatic generation of visual dictionary is more suitable On the object image retrieval.BVWIR similar image blocks with a high probability of release to the same visual vocabulary, and does not require any global information. At the same time, to reduce the high query cost caused by a number of local features, the framework includes two types of message transmission mode, parallel mode and serial mode, in order to reduce the query cost and to ensure the accuracy.

【学位授予单位】：北京邮电大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP391.41

【相似文献】