Local feature based pattern classification: from principle to application

This thesis demonstrates that local feature based approaches are always more stable than global feature based approaches for pattern classification problems. Guided by the original theory that a regional matching approach is more robust than a national matching approach for two-dimensional pattern classification, this thesis examines the applications of the theory in one-dimensional and two-dimensional pattern classifications. We propose two local feature based approaches for two significant applications of pattern classification, namely start codon prediction and content based image classification. For start codon prediction which is considered as a typical one-dimensional pattern classification problem, we have developed a districted neural network that can be taken as a regional voting version of the conventional neural network. Experiments have been performed on the well known translation initiation sites (TIS) data sets and results have shown significant improvement of prediction accuracy. For two-dimensional pattern classification, we propose differential latent semantic index (DLSI) approach for content based image classification. The feasibility of using local features in the DLSI method is also investigated and an extensive experimental study on a real image database has proved its effectiveness.