Unsupervised manifold learning through reciprocal kNN graph and Connected Components for image retrieval tasks

Abstract

Performing effective image retrieval tasks, capable of exploiting the underlying structure of datasets still constitutes a challenge research scenario. This paper proposes a novel manifold learning approach that exploits the intrinsic dataset geometry for improving the effectiveness of image retrieval tasks. The underlying dataset manifold is modeled and analyzed in terms of a Reciprocal kNN Graph and its Connected Components. The method computes the new retrieval results on an unsupervised way, without the need of any user intervention. A large experimental evaluation was conducted, considering different image retrieval tasks, various datasets and features. The proposed method yields better effectiveness results than various methods recently proposed, achieving effectiveness gains up to +40.75%.