佐藤優伍, 福里司, 森島繁生
This paper presents a novel image retrieval system for finding a particular image from a database using the implicit similarity of impression. Previous content-based image retrieval systems depend on the low-level visual features of images (e.g., luminance or RGB color), so it is difficult to compute semantic image similarity. Additionally, if a user does not have much knowledge about a target image, these systems may be difficult to use because they require some image or text queries. To solve these problems, we integrate the deep convolutional neural network and online learning based on relevance feedback to interactively estimate the image the user is searching for. We ran user studies with 10 subjects on a public database consisting of 597 face images, and confirmed that the proposed system is effective for retrieving face images easily and quickly.