Concepts of pattern recognition
Henri H. Arsenault · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2000
This tutorial reviews some concepts of pattern recognition and introduces some recent new ideas. After a brief overview of classical pattern recognition approaches, some unconventional and useful concepts will be introduced and used to show how the input formats of images for both digital and optical pattern recognition techniques influence the choice of methods and of criteria for measuring similarity. The importance of nonimearities and of pre-processing will be demonstrated, and the important differences between cases where the objects of interest may be segmented from the scene and those where they may not will be pointed out. It will be shown that the mathematics underlying conventional classification methods and those using neural networks are not as different as they may first appear, and that some tasks predicted for neural networks involving pattern classification are impossible in principle. Some new approaches for achieving object classification invariant under translation, rotation, illumination and other distortions will be discussed.