Abstract
An image retrieval system that retrieves images such as paintings and illustrations with Kansei words and image-content keywords using a visual user interface was proposed and implemented. Although Kansei image retrieval systems, which retrieve images with a query containing only Kansei words, have been investigated extensively, these systems have a number of disadvantages. For example, the user often wants to retrieve images using not only sensibility terms but also words that expresses the subject of the images. In addition, since the system requires the user to input query in his/her own Kansei words, it may be difficult for the user to express a request using the correct words. Moreover, the user interface displaying the retrieval results should allow the user to review the retrieved images in order to facilitate the determination of which images in the retrieval results satisfy the user's request. Therefore, a system with the following functions was proposed: The system takes a query as a pattern of Kansei words and image-content keywords, and the user can select the Kansei word from the map based on the Color Image Scale; The system shows the retrieval results in a two-dimensional plane allowing the user to know the fitness of the retrieved image at a glance; The system implements the user relevance feedback framework to adapt the retrieval to changes in the user requirements. The system was evaluated subjectively and was found to provide superior accuracy and usability compared to conventional systems.
