Realizing a Portal Site for Conceptual Information Retrieval That can Treat Texts and Images in a Unified Way

研究的目的是实现一种新的图像检索模型，该模型可以像通常的文档检索那样接受关键字查询，并根据与查询的概念关联程度输出排序的图像。这是通过在用户和商业Web图像搜索引擎之间建立一个图像检索门户网站来实现的。WWW中的图像经常被不适当和不充分的关键字索引。根据我们的模型，用户被要求对他们认为相关的图像进行合作标注，并将标注后的图像存储在门户网站中，以提高后续检索的有效性。对于图像标注，我们提出了一种帮助用户的想法，即显示图像所属的四种图像类型中的一种，并使用带有四个栏的填表，要求用户为图像输入看似合适的关键字。有时可能会给出一些不合适的关键字，然后通过拟议的概念过滤例程将其过滤掉。为了实现一个真正的图像检索门户网站，需要形式化地定义查询关键词与图像关键词的概念匹配机制，以及门户网站与商业搜索引擎协同工作的机制。已经编制了概念匹配例程，并从检索有效性的角度对其进行了充分检查。至于企业，我们的模式可以与Excite、Yahoo！和Goo图像搜索引擎。

The research aim is to realize a new image retrieval model that can receive a keyword query, as is seen in usual document retrieval, and output the ranked images according to the degrees of the conceptual relatedness to the query. This has been done by settling an image retrieval portal site between users and a commercial Web image search engine.The images in the WWW are often indexed by inappropriate and insufficient keywords. According to our model, users are asked to cooperatively annotate the images that they have decided being relevant, and the annotated images are stored in the portal site to raise the effectiveness of the successive retrieval. As to the image annotation we have proposed an idea of assisting users by showing one of the four image types to which an image belongs and a fill-in-form with four slots in which users are requested to input the seemingly appropriate keywords for the image. Some inappropriate keywords may be occasionally given and then will be filtered out by the proposed conceptual filtering routine. The effects of the conceptual filtering was experimentally examined in comparison with a simple frequency based method on a set of 1,000 images drown via Google image search engine.In order to realize a real image retrieval portal site, a mechanism of conceptually matching the query keywords with the image keywords as well as a mechanism of cooperatively working the portal site with a commercial search engine should be formalized. The conceptual matching routine has been programmed and fully examined in a viewpoint of retrieval effectiveness. As to the corporation, our model can work with Excite, Yahoo! and Goo image search engines.