AccessGrid Node
Arts & Sciences
A & M University
Alabama A & M University Logo

The Alabama A & M University
Department of Physics
Optics and Information Processing

Image Processing with Pulse Coupled Neural Networks

M. P. Schamschula, R. Inguva, and J. L. Johnson

The Second International Forum on Multimedia and Image Processing,
World Automation Congress 2000, Wailea, Maui, HI (2000)


The Pulse Coupled Neural Network (PCNN) differs from conventional Artificial Neural Networks by having time dependent output and not requiring training (supervised or unsupervised). The basic PCNN algorithm consists of five coupled equations, one of which is non-linear. In an image processing application, the PCNN output takes the form of a series of pulses that are a function of the input intensity. The brighter the pixel, the more frequently it pulses. Two of the PCNN functions interconnect with neighboring cells, in one case with a constant weight in the other with various convolution kernels. By adjusting these interconnection weights (choosing an appropriate kernel), PCNNs can be used for a number of image processing functions. For example we may use the PCNN to resample an image into a smaller intensity lookup table (intensity compression) or use it to filter out intensity dependent high spatial frequency noise. The pulse train for a given image repeats over time, however, the time a pulse is triggered at each pixel changes. We can perform statistical analysis on a series of pulse trains. A statistical average of these pulse trains can be used to reconstruct a cleaned up image. This average image also represents a 3D icon of the image. Other examples from current research will be given.