Introducing a word vector space model and improve image features to web image-gathering

摘要

For realizing a recognition system for real-world images, recently, many methods by which a large number of images are required as learning images are proposed. It is, however, hard to collect a large number of learning images, so that many conventional studies restricted their target objects to special objects such as human faces and cars. Then, in this paper, we propose a system that can gather a large number of real-world images from the World-Wide Web (WWW) automatically. We realize the system by making three kinds of improvements for the Image Collector, which is an Web-image-gathering system we proposed in the past. (1) We use word vectors of HTML files embedding image files as well as image feature vectors extracted from image files. (2) Instead of simple color histogram, we use Earth Mover's Distance for computing the similarity between images. (3) For gathering much more images, we extract words from all the HTML files embedding final output images, select top ten words about the frequency that words appear in the HTML files, and adds them to query keywords to text-based Web search engines.