Image Classification Approach Based on Manifold Learning in Web Image Mining

Automatic image classification is a challenging research topic in Web image mining. In this paper, we formulate image classification problem as the calculation of the distance measure between training manifold and test manifold. We propose an improved nonlinear dimensionality reduction algorithm based on neighborhood optimization, not only to decrease feature dimensionality but also to transform the problem from high-dimensional data space into low-dimensional feature space. Considering that the images in most real-world applications have large diversities within category and among categories, we propose a new scheme to construct a set of training manifolds each representing, one semantic category and partition each nonlinear manifold into several linear sub-manifolds via region growing. Moreover, to further reduce computational complexity, each sub-manifold is depicted by aggregation center. Experimental results on two Web image sets demonstrate the feasibility and effectiveness of the proposed approach.