利用階層式視覺文字和空間關係驗證提升跨來源影像搜尋之效能

Abstract

We proposed a method to solve the general cross-domain image retrieval problem. This method focuses on image content structure but not the detail, so it is particularly useful to matching the images across different visual domains e.g., paintings, hand-drawn sketches and photographs taken in different times. There is a method trying to solve this problem by classification. Although it gets good performance in the experiment, it is too computationally intensive to do the interactive applications. For finding a fast solution, we analyse the problem causing BoW method fails in the cross-domain setting first. Then, According to the observation, our proposed method releases the strict matched constraint and uses a fast spatial verification to describe the content structure similarity. Finally, our method is very faster than the baseline solution and achieves the same performance. Key words: Cross-domain, image retrieval, spatial verification, multi-level matching, image matching, sketch