Invited Talk World Congress DSA 2013

Combining Pattern Classification Systems through Similarity Analysis of stable hierarchical Features

Dimitrios A. Karras
Chalkis Institute of Technology, Dept. Automation, Hellas, Greece


A framework is herein presented for combining decisions of pattern classification systems based on hierarchical features similarity analysis. After a critical overview of current trends in combining pattern classification decisions mainly through presentation of state of the art voting schemes, a methodology is developed to analyze similar high order hierarchical patterns extracted from different pattern classification methods. Furthermore, different cases of each such classifier, derived from similar training processes but with different training algorithms/parameters, are investigated in terms of their hierarchical features, through similarity analysis, in order to find out repetitive and stable higher order hierarchical features. Then, all such stable hierarchical features are integrated through a second stage pattern classifier based on artificial neural networks having as inputs suitable similarity features of them. The herein suggested framework for hierarchical pattern classification is investigated experimentally in computer vision and medical decision making tasks with promising results.