Iterative Locality Preserving Projection for Image Retrieval

Recently there has been considerable interest in subspace learning for efficient multimedia information retrieval. The typical subspace learning for discovering the intrinsic geometrical structure include Principal Component Analysis (PCA) and Locality Preserving Projections (LPP). PCA discover the global Euclidean structure while LPP discovers the local manifold structure. LPP is based on a nearest neighbor graph which models the local geometrical structure of the image manifold. However, such a graph can not always accurately estimate the intrinsic manifold structure. In this paper, we propose a novel algorithm called Iterative Locality Preserving Projections (ILPP). ILPP iteratively updates the nearest neighbor graph, so that it can better model the intrinsic manifold structure. We compared our algorithm with PCA and LPP on the COREL image database. Experimental results show that our algorithm outperforms PCA and LPP for image retrieval.