Image processing techniques for cellular neural network hardware

The Cellular Neural Network (CNN) is a continuous-time feedback neural network where the processors (cells) are arranged on a regular grid and interactions are restricted to be local and space-invariant. The CNN topology is well-suited for image processing applications--image pixels can be mapped directly onto the array of cells for massively-parallel analog processing. The CNN Universal Machine (CNNUM) was later invented to allow the outputs of previous CNN operations to be stored on-chip and transferred as inputs to subsequent CNN operations, a requirement of any complex image processing algorithm.; Researchers have found numerous CNN cell interactions (templates) which perform an interesting image processing operation, upon which many CNNUM-based image processing algorithms have been developed. Unfortunately, many of these algorithms use templates with special properties, which are allowed within an extended CNN definition and provided by standard CNN simulators, but are not likely to be included in the general purpose CNNUM circuits built in the near-future.; In this dissertation, the theoretical image processing capabilities of the standard, simple CNN is capable of solving complex image processing tasks. A bottom-up approach is taken. The basic dynamical phenomena of the CNN are first studied, primarily by employing a modal representation, and some elemental CNNUM processing functions are identified. The bulk of the dissertation builds up some useful image processing techniques from these CNN elemental operations. Some conventional methods, e.g. standard quantization, FIR and IIR spatial filtering, binary morphology, and cellular automata, which are known to be useful and have known behaviors are emulated or approximated by using the CNN primitives. In fact, it is shown that both arbitrary FIR convolutions and arbitrary 3 x 3 neighborhood logic functions can be implemented on the CNNUM. In addition, the potential to exploit the inherent pattern-forming capabilities of some CNN templates to develop new image processing methods is considered. Finally, some examples in video-microscopy are given to demonstrate how the developed CNN-based methods can be applied to real-world image processing tasks.