Visualization System Displaying Retrieved Images in 2-D Semantic Space

We have developed a similarity-based image retrieval system that represents retrieved images as a scatter diagram in a semantic space. An axis of the space shows the suitabilities of a keyword assigned to the images. The suitabilities are estimated by a linear transformation of the image features of retrieved data, and the coefficients of the transformation are learned by a multiple regression analysis. The system can store multiple key-images and can retrieve images by using the center of gravity of the key-image feature vectors. The system can effectively assist a user in retrieving images through the use of semantic visualization and this center of gravity retrieval based on similarity. The system performs as follows. First, a user retrieves images. Then, the system presents the user with some keywords that provide large variances of the estimated suitabilities from among the retrieved results. Next, the user can choose two axes from the given keywords. Finally, the system displays the retrieved images in the semantic space that is spanned by the axes. We examined our method quantitatively by using a large number of samples, each of which was a pair made up of an image and its assigned keywords.