Font Size: a A A

Research On Content Based Image Retrieval In The Distributed System

Posted on:2016-11-04Degree:DoctorType:Dissertation
Country:ChinaCandidate:D YangFull Text:PDF
GTID:1108330482957866Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
With the further development of the mobile internet, smart products widely used promote new applications which are popular all over the world. The number of images exponentially growths, thanks to these new applications. Challenges and opportunities raised by "big data" and Cloud computing affect new research on images.The traditional centrialized index mode is not effective and scale, due to the huge number of images. Recently, the decentrialized aritechture of the datacenter becomes a new solution, which is built on Distributed Hash Table (DHT) with the help of the idea of Overlay Network. The aritechture is suitable for the massive information retrieval and management. The node is only responsible for a small part of resource indexes and routing information, in order to find the address and locate the resource in the structured distributed system based on DHT. Different from the Text-Based Image Retrieval, Content-based Image Retrieval (CBIR), known as a query by the sample image, analyzes the contents of the image rather than keywords. Thus, it is an important subject to locate and match images in the DHT network.In this dissertation, we proposed a scalable CBIR framework which generates different types of the distributed index for various images. It adopts overlay routing protocol to publish indexes and locate queries. The CBIR framework is convenient for the large-scale deployment, and capable of self-organizing in the cloud environment with the least effort. It has the advantage of robustness and scalability. To implement this framework includes the following problems. How to establish the distributed visual indexes according to the feedback results; How to evenly assign the index to nodes while meet the demand of load balance; How to ensure the efficiency of the query route with low communication cost. Our work focuses on these problems and presents the following contributions:1. The framework of LSH-based and Relevance Feedback for Image Retrieval (LRFIR) for the large scale P2P datacenter is proposed. LRFIR efficiently support content similarity search and semantic search in the distributed environment, which is different from the traditional location approach. In the index construction sevice, LRFIR leverages multi-texton histogram to represent the image content. Its key idea is to integrate image feature into DHTs by exploiting the property of locality sensitive hashing (LSH). Thus, images with similar content are most likely gathered into the same node without the knowledge of any global information. The query is only sent to the nodes which are more likely to answer it.2. To support search semantically close images, the relevant feedback is adopted in our system to overcome the gap between low-level features and high-level features. This approach allows users to interact with LRFIR to refine the query vector, which moves the query point towards positive samples. Database Corel and the subset of Catltech 101 Object Categories are used in our experiments. We show that our approach yields high recall rate with good load balance and only requires a few number of hops.3. LSH-based and Fusion Features for Image Retrieval (LFFIR) for the distributed datacenter is presented. The basic idea is to adopt the peer-to-peer paradigm and relies on fusing multiple image features. The fusion features to represent image content with different aspects, instead of a single feature. LFFIR consists of index construction service and query process service. The former one constructs multiple index replicas through exploiting the LSH property. Thus, the indexes of similar images are probabilistically gathered into the same node. The latter one is responsible to publish query messages and route the query to the node most likely response the query.4. A novel CBIR system called Bag of Visual Words based Image Retrieval (BVWIR) is proposed. BVWIR efficiently incorporates the bag-of-visual-word model into DHTs. Its key idea is to establish visual words for local image features by exploiting the merit of LSH, so that similar image patches are most likely gathered into the same nodes without the knowledge of any global information. Furthermore, two message transmission modes are proposed in order to reduce the query cost generated by numerous local features.
Keywords/Search Tags:Peer-to-peer, DHT, Cloud computing, Content based image retrieval, Locality sensitive hashing, Relevant feedback
PDF Full Text Request
Related items