Hashing with Locally Projections |
|
Go IRIE，Hiroyuki ARAI，Yukinobu TANIGUCHI |
|
|
 Go IRIE |
 Hiroyuki ARAI |
 Yukinobu TANIGUCHI |
Various types of hashing techniques have been proposed, and most of them perform hashing by preserving Euclidean distance closeness in feature space. In general, feature distance will become large in the case of images captured from different viewing angles, even if they captured the same object. That is, appropriate images may not be found when Euclidean distance was used. To overcome this problem, by using the knowledge that similar images can be observed as a non-linear manifold in feature space, this paper proposes a method of hashing-based image retrieval that can deal with the structure of its manifold appropriately. The authors demonstrate that a non-linear manifold can be represented by the combination of locally linear structures and it can be calculated using locally linear sparse decomposition. They also demonstrate that its structure can be preserved in Hamming space by using locally linear projections.
In experiments, the proposed method was compared with eight hashing-based image retrieval methods using three well-known datasets. As shown in the experimental results, the proposed method could improve the performance satisfactorily. In addition, it was demonstrated that the computational complexity of the proposed method is linear against the number of images. Consequently, the proposed method achieved accurate image retrieval using manifold characteristics while maintaining the fast computation that is one of the advantages of hashing-based methods.
From the above points of view, the proposed method is a highly promising approach for effective image retrieval, and this paper is therefore a worthy recipient of the Best Paper Award.
