基于有向图模型的多模态新闻图像检索研究

A Research on News Images Retrieval Based on Directed Graph Multimodal Characteristics

不同模态特征描述网络新闻数据各具优势。为充分利用不同模态特征的优点,提高网络新闻图像检索准确率,提出了一种基于有向图理论模型的多模态特征融合检索方法。首先对新闻图像进行预处理,提升新闻图像兴趣点对比度,之后对新闻文本提取关键词,然后对新闻图像进行场景辨别,结合图像人物特征,将文本、场景、人物特征依据有向图理论模型进行融合,形成基于有向图理论模型的多模态融合检索。在10万条新闻数据上测试,实验结果表明,本文提出的方法检索准确率达到了69%,查全率达到70%,效果提升了5%。

Different modal characteristics have its own advantages in describing network news data. To make full use of the advantages of different modal characteristics and improve the retrieval accuracy,a method of image retrieval based on the directed graph theory modal with features fusion is put forward. Firstly,the method of image de-noising and image enhancement are used to pretreat the image and the keywords are extracted from the network news text,second the news image scene is identified and the faces in the news image are recognized,finally,the fusion of extracted features form the multimodal retrieval. The method is tested at ten million data and the results show that the retrieval accuracy achieves 69% and the recall achieves 70%,the effects improve 5%.