Image retrieval using dimensionality reduction
Ke Lu,Xiaofei He,Jiazhi Zeng
DOI: https://doi.org/10.1007/978-3-540-30497-5_120
2004-01-01
Abstract:Image representation has been a fundamental problem for many real world applications, such as image database visualization, browsing, retrieval, etc. In this paper, we investigate the use of Laplacian Eigenmap (LE) for image representation and retrieval. Conventional, Principal Component Analysis (PCA) has been considered effective as to discovering the low dimensional structure of the image space. However, PCA can only discover the linear structure. It fails when the images are sampled from a low dimensional nonlinear manifold which is embedded in the high dimensional Euclidean space. By using Laplacian Eigenmap, we first build a nearest neighbor graph which models the local geometrical structure of the image space. A locality preserving mapping is then obtained to respect the graph structure. We compared the PCA and LE based image representations in the context of image retrieval. Experimental results show the effectiveness of the LE based representation.
What problem does this paper attempt to address?