Feature extraction and classification for image analysis

Abstract

Pattern recognition techniques have been widely used in a variety of scientific disciplines including computer vision, image understanding, biology and so on. Although many methods present satisfactory performances for image analysis, they still have several weak points and thus leave a lot of room for further improvements. For example, the linear discriminant analysis (LDA) algorithm is able to extract discriminative features, but the small sample size (SSS) problem limits its application scope.

In this thesis, several feature extraction and learning algorithms are proposed to improve the classification performance in image analysis. In the first proposal, the multiple Trace feature (MTF) is constructed as a novel pattern representation by integrating several Trace transforms where genetic algorithms (GAs) serve as the information fusion tool. Moreover, a novel fitness function is proposed for GAs by combining the bootstrap aggregating algorithm and the cross-validation scheme. As a result, the GAs-based iterative learning process is able to deal with the overfitting problem by using the new fitness.