An Image Retrieval Method Considering Similarity of Impression Based on Color and Image Composition

Abstract

An immense amount of digital image is distributed and stored over the Internet due to the spread of cloud services and smartphones. Under the circumstance, there is an increasing need for image retrieval using not only keyword query but also image query, that is, similarity-based image retrieval. In this study, we propose an image retrieval method considering similarity of impression. The similarity of impression between two images is computed by the eight dimensional emotional vector defined by the emotional words which represent the impression of the images. The emotional vector of an image is determined by the color information and the composition of the image. In particular, we derive emotional vector of an image by using the proportion of 180 emotional words computed by the color information of the image and the proportion of 6 emotional words derived by the location of objects on the image. We verified the validity of the proposed method by the experiments using Caltech 101 database.