A detailed analysis of different CNN implementations for a real-time image processing system

A detailed analysis for different implementations of a real-time CNN signal processing systems is presented. The algorithm for signal reconstruction has been realized both in hardware (analog VLSI multi-FPGA system) and in software (TriMedia VLIW Intel Pentium processor). All implementations are fully functional and embedded in a system environment. Due to the high computational complexity which is needed to solve the nonlinear CNN-equations and the requirements which are different for each application, an efficient implementation has to be tailor-made. In this paper we analyze different realized implementations regarding prototypical pre-requisites.