Efficient off-line feature selection strategies for on-line classifier systems

In this work we discuss feature/signature selection strategies for on-line classifier systems, implemented on a common stand-alone HW/SW vision system. In the chosen computational environment topographic and non-topographic computing can be combined for the targeted task of terrain feature analysis. The topographic front-end of the system is capable of deriving more than thousand different visual signatures per second from a few dozen visual feature channels. These signatures are statistically correlated and might also be computationally inter-dependent. To develop an efficient real-time classification algorithm the dimensionality of the extracted features has to be significantly reduced, i.e. only a few signatures can be selected. In this paper we analyze different signature selection strategies for various classifiers with given performance criteria and system-level time performance constraints. Two blind signature selection techniques are described: variance maximization and a factor-based analysis. We also discuss supervised selection mechanisms: class-based statistical and decision-tree-based approaches.