Web Image Re-Ranked Specified From User Query Keywords

Abstract

A web image re-ranking is the process of image retrieval from user’s interest. Given a query keyword, a pool of images is first retrieved by the search engine based on textual information. By asking the user to select a query image from the pool, the remaining images are re-ranked based on their visual similarities with the query text. An earlier projects they use universal visual semantic space is characterized to difficult and inefficient. In this project proposed a novel image re-ranking, which automatically offline learns different visual semantic spaces using different query keywords. At the online stage, images are re-ranked by comparing their semantic signatures obtained from visual semantic spaces specified from query keyword. Image Search engines mostly use keywords and they rely on surrounding text for searching images. Ambiguity of query images is hard to describe accurately by using keywords. E.g.: Apple is query keyword then categories can be “red apple”, “apple laptop” etc. Another challenge is without online training low level features may not well co-relate with high level semantic meanings. Low-level features are sometimes inconsistent with visual perception.
Key Words: K-Means Clustering, Semantic Signatures