A Novel Image Classification Method Based on Double Manifold Learning

To solve the two-class classification problem existing in semantic-based image understanding, a novel classification method based on double manifold learning is proposed, which can transform the classification problem from a high-dimensional data space to a feature space with lower dimensionality. Two manifolds with different intrinsic dimensionalities will be first established separately, according to the significant differences between the positive samples and the negative ones, where globular neighborhood-based locally linear embedding (GNLLE) algorithm is adopted to implement dimensionality reduction and meantime unsupervised clustering. Then the aggregation center of each manifold is calculated, taking into account the grouping characteristics of similar samples. Furthermore, a new classifier is constructed for a double manifold learning model via distance companion. Finally experiments indicate that our method, which can be easily extended to multi-classification manifold learning, will not only reflect the topological structure of the whole data more precisely, but also achieve performance of classification more efficiently.