多媒体文档检索研究.pdf

摘要 随着数字媒体技术和社交网络的发展与普及，越来越多的网络信息以包含多 种模态数据的多媒体文档的形式出现。传统的文档检索已经不能满足人们的需求， 如何有效的从这些海量多媒体文档中快速、准确的检索到用户需求的信息是高效 利用互联网资源迫切需要解决的问题。 本文首先对现有的多媒体信息检索的方法进行了研究，在其基础上，针对利 用单一模态对多媒体文档进行检索时表达的语义含混不清具有歧义的问题，提出 了一种多模态融合中采用线性加权和的方法对多媒体文档进行检索。其中文本采 取基于Lucene 的全文检索，图像利用基于内容的纹理、颜色等检索的LIRE 。权重 的分配按照各个模态检索结果的正确率做比例来划分，此方法经实验验证检索准 确率较高。本文提出的另外一个方法是利用融合的词袋方法 BOW 和特征袋方法 BOF 对多媒体文档进行检索。其中应用的词袋方法是经过对单词分类后的改进的 词袋方法，改进的方法大大降低了码书的尺寸，使得检索效率更高。实验证明， 融合的BOW 和BOF 综合考虑了文本和图像对多媒体文档检索的影响，效率要高 于只使用BOW 或BOF 对多媒体文档的检索。 关键词：多媒体文档检索 Lucene BOW BOF Abstract With the development and popularization of digital media technology and social networking, network is full of huge amounts of multimedia documents which contain various modalities. Traditional document retrieval can ’t satisfy people ’s needs. How to find information that users need quickly and accurately from huge amounts of multimedia documents is an urgent problem to solve. The paper first makes a research on existing methods of multimedia information retrieval. Based on various modalities in multimedia document, we present a new method that apply linear weighted sum of multimodal fusion to retrieval multimedia documents. With this method, text retrieval resorts to full text search technology which bases on Lucene. To image retrieval, it resorts to LIRE which applied visual features such as texture and color. And hierarchy indexing was adopted to retrieval video. Weight is assigned to each modality in accordance with scale factor that results from precision rate of retrieval outcomes using single modality only. This method was proved to have high efficiency but need much more time. Merging Bag-of-words