| In the past few years, content-based visual information retrieval, which is based on visual information understanding, has been received more and more attention and become the key technique of visual information retrieval. The main point is, instead of being manually annotated by text-based keywords, visual information would be indexed by its own visual content, such as color, texture, etc.In this dissertation, we first give an introduction to the state-of-the-art of content-based image retrieval systems from three points: the framework of system, image content description and the key techniques of content-based image retrieval system. Also in the introduction, we present several important sample systems, both commercial and research. Then we focus on relevance feedback technique in image retrieval and propose a novel"query subspace"-based retrieval model, that firstly classifies the positive images provided by user and then uses these classes of samples to form several query subspaces, which have their own query and retrieval model. The ultimate retrieval result is based on the retrieval results of these query subspaces. The method dynamically builds the mapping between low-level visual features and high-level semantics and adapts to different queries. At the same time, we rationally integrate the negative feedback information into our model to improve the efficiency of system.As the first step of video retrieval, video segmentation is one of the most active research areas. In this dissertation, we propose a DC-Image based shot detection algorithm, which can detect cut, fade and wipe. We first extract DC-Images from compressed video data, and then adopt different detection algorithms based on their respective transform models in their own features. For dissolve detection, we implement and compare two mostly used edge-based algorithms (EC-based algorithm & ECR-based algorithm) and find that the EC-based algorithm is more robust than ECR-based one. |
PDF Full Text Request