Color and shape classification with competing paradigms: neural networks versus trainable table classifiers
Robert C. Massen, Thomas Regle, Pia Boettcher · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1990
The pixel-wise classification of CCD colour Images Into previously learned colour classes at video-rate is a demanding vision task, both with regard to the complicated cluster shapes encountered in natui- al scenes and to the required computing power for real-time operation. We discuss two classical solutions based on an algorithmic statistical classifier and on a Neural Network paradigm and propose an alternative simple and low-cost classifier based on approbriately trained look-up-tables. Two different learningrules for the supervised training of this LUT classifier are presented for the colour classification of both synthetic and natural blotechno1ojr scenes. The proposed LUT classifier shows all the positive features of a (simulated) 3-layer perceptron Neural Network, but performs 60.000 times faster with simple, commercially available components.