Steac State using For ard
Bertram Emil Shi · 2006
We describe atechnique bywhichthesteady statesolution ofsomeCNN template operations canbe approximated inafewpasses overtheinput image. For digital implementations, thisismuchmoreefficient than numerical integration oftheCNN differential equations. In particular, weshowthatthesteady state solution tolinear 1DCNN filtering operations canbefoundexactly intwo passes overtheinput imageusing theforward-backward filtering technique commonlyusedin digital signal processing toobtain zerophase linear filters. We thenshow thatthistechnique canbemodified tofindapproximate solutions tosomenonlinear CNN operations, using the example ofalocal winner-take-all network. IndexTerms-cellular neural networks, recursive digital filters, image processing, analog anddigital systems