به نام خدا

Title: Query-Specific Visual Semantic Spaces for Web Image Re-ranking

Authors: Xiaogang Wang,Ke Liu,Xiaoou Tang,

Abstract: Image re-ranking, as an effective way to improve the re-sults of web-based image search, has been adopted by cur-rent commercial search engines. Given a query keyword, apool of images are first retrieved by the search engine basedon textual information. By asking the user to select a queryimage from the pool, the remaining images are re-rankedbased on their visual similarities with the query image. Amajor challenge is that the similarities of visual features donot well correlate with images_ semantic meanings whichinterpret users_ search intention. On the other hand, learn-ing a universal visual semantic space to characterize highlydiverse images from the web is difficult and inefficient.In this paper, we propose a novel image re-rankingframework, which automatically offline learns different vi-sual semantic spaces for different query keywords throughkeyword expansions. The visual features of images are pro-jected into their related visual semantic spaces to get se-mantic signatures. At the online stage, images are re-rankedby comparing their semantic signatures obtained from thevisual semantic space specified by the query keyword. Thenew approach significantly improves both the accuracy andefficiency of image re-ranking. The original visual featuresof thousands of dimensions can be projected to the seman-tic signatures as short as 25 dimensions. Experimental re-sults show that 20% - 35% relative improvement has beenachieved on re-ranking precisions compared with the state-of-the-art methods.   

Publish Year: 2011

Published in: CVPR - IEEE

Number of Pages: 8

موضوع: وب معنایی

مشاهده صفحه اول مقاله

دانلود مقاله

لینک مقاله در سایت ناشر

ایران سای – مرجع مقالات علمی فنی مهندسی

حامی دانش بومی ایرانیان