基于无视觉码本框架的大规模图像检索研究

项目名称： 基于无视觉码本框架的大规模图像检索研究

项目编号： No.61472378

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 计算机科学学科

项目作者： 周文罡

作者单位： 中国科学技术大学

项目金额： 84万元

中文摘要： 现有基于内容的图像检索借鉴文本检索思想，采用视觉词袋模型,其核心技术之一是利用视觉码本将图像局部特征量化为视觉单词。然而，基于码本的量化本质上是一种面向数据压缩的矢量量化，并不完全适用于面向视觉内容识别的大规模图像检索。本项目拟研究的关键科学问题是：面向视觉内容相关度保持的局部视觉特征量化方法。本项目创新性地提出研究无视觉码本的图像检索框架，避免训练视觉码本的约束，可对任意图像数据库自适应地生成量化器，对局部特征进行快速量化和索引，实现可扩展的大规模图像检索。此外，为完善无码本检索框架，本项目还拟研究量化前的视觉特征采样方法和量化后的基于图像上下文分析的检索重排序法方法，以进一步提高检索精度。本项目预期在局部视觉特征的量化理论上有所突破，在基于内容的大规模图像检索技术方法上取得重大创新，尝试探索面向大数据处理与分析的数据压缩理论，为大规模图像检索技术的理论化、实用化奠定基础。

中文关键词： 图像检索；局部特征采样；视觉特征量化；图像索引；检索重排序

英文摘要： Currently, most content-based image retrieval (CBIR) algorithms and techniques adopt the Bag-of-Visual-Words model using the idea from information retrieval, and make use of the visual codebook to perform quantization on image local features. However, such codebook-based feature quantization in many existing CBIR methods is essentially a kind of data-compression-oriented vector quantization, which is not fully applicable to the large-scale image retrieval which targets at visual content identification. With awareness of this problem, this project proposed to research the visual content similarity-preserving quantization on local visual features. We propose a novel codebook-free image retrieval framework, which avoids the constraint of training visual codebook and can be flexibly adapted to any diverse image dataset for efficient quantization and indexing for scalable image search. Moreover, to enhance the codebook-free framework, we propose to research the visual feature sampling method before feature quantization and the search re-ranking method with image context analysis after feature quantization to further boost the retrieval performance. This project is to make a breakthrough in the quantization theory on local visual feature and make significant novelty contributions in techniques and methods for large-scale image retrieval. Besides, it will also explore the data compression theory for large-scale content data processing and analysis to lay the foundation for theoretical and practical large-scale image retrieval techniques.

英文关键词： image retrieval;local feature sampling;visual feature quantization;image indexing;search reranking