A Visual and Semantic Image Retrieval Method Based on Similarity Computing with Query-Context Recognition

  • Authors:

  • Xing Chen;Yasushi Kiyoki

  • Affiliations:

  • Department of Information & Computer Sciences, Kanagawa Institute of Technology, 1030 Simo-Ogino, Atsugi-shi, Kanagawa 243-0292, Japan, chen@ic.kanagawa-it.ac.jp;Department of Environmental Information, Keio University, Fujisawa, Kanagawa 252-8520, Japan, kiyoki@mdbl.sfc.keio.ac.jp

  • Venue:
  • Proceedings of the 2007 conference on Information Modelling and Knowledge Bases XVIII
  • Year:
  • 2007

Quantified Score

Hi-index 0.00

Visualization

Abstract

This paper presents an image retrieval method based on visual and semantic similarity computing with a query-context recognition mechanism. The motivation of our work is to solve the problem which can be described as that if only the visual similarity or only the semantic similarity judgment is performed on image retrieval, the retrieval results do not always match the query intentions of users. Our central idea is that similarity computing has to be performed between visual and semantic levels. To understand the relationship between the visual factors and the semantic factors in images, we have performed experimental studies. From our experimental studies, it is found that it is possible to extract semantic factors from the visual factors of images. Furthermore, it is found that users' query intention can be detected from the difference of images in queries. Based on the experimental results, we develop a method to implement both the semantic and visual similarity judgment for image retrieval. In this method, several images are required to be given as the key images in a query for users to indicate their query intentions. Furthermore, an adjusting value is used for users to indicate their query intentions, intending on the visual similarity or the semantic similarity. Both the visual and semantic factors are extracted from the key images and the similarity computation is performed on the extracted factors. The effectiveness of the method is clarified based on our experimental results.