This paper discusses a similarity image retrieval method in an image database which corresponds to diversified human similarity sensation. In the proposed similarity image retrieval, a similarity between images is examined by subjective evaluation, and a multidimensional psychological space is constructed using multidimensional scaling. Each of the coordinate axes in the multidimensional psychological space is represented by a linear combination of image feature parameters using multiple regression analysis. The image feature parameters extracted from a retrieval key image are converted into a point in the multidimensional psychological space using part or all of the forementioned linear combination expressions. The image with the minimum distance among the stored images is given as the retrieval output. Two-hundred sixty binary patterns of a butterfly are evaluated by a subjective test, and obtained similarities between patterns are used as the true values. The error is examined between the forementioned result and the value estimated in the multidimensional psychological space calculated from a linear combination expression of image feature parameters. As a result, the optimal dimension of the multidimensional psychological space is determined as 15. Retrieval performance in the space with the optimal dimension is evaluated from viewpoints of retrieval rate and mixture of dissimilar patterns into the retrieval result. Usefulness of the method is verified. Axis implications in the multidimensional psychological space are examined, and the possibility to retrieve an image from an individual viewpoint is indicated.
|ジャーナル||Systems and Computers in Japan|
|出版ステータス||Published - 1993 12月 1|
ASJC Scopus subject areas