Multi-template training for image processing with cellular neural networks

Cellular neural networks (CNN) are often considered as massive parallel computing arrays for high speed image processing. In order to find appropriate CNN templates, optimization methods are necessary in many cases. We consider the optimization method Iterative Annealing directly using the output of a hardware realization of a CNN-UM Chip. The procedure presented in this contribution generates highly adapted sets of templates for complex image processing tasks. With this approach it is also possible to tune existing CNN programs to compensate inaccuracies of analog CNN hardware leading to noise reduction and more robust behaviour. Finally, an application of practical interest has been developed, by using the introduced method. We achieved the tracing of a certain selected object out of an image sequence showing many moving objects.