CAM2-universal machine: A DTCNN implementation for real-time image processing

This paper is first report of implementation of CNN on CAM². CAM² is a highly-parallel two-dimensional cellular automata architecture. Due to digital circuitry restriction, CNN on CAM² operates as a discrete time CNN (DT-CNN), and each cell has quantized output function. In spite of such a restriction, we experimentally confirm that DT-CNN on CAM² works nearly the same way of conventional CT-CNN model. The special feature of the CNN universal machine is flexibility and programability. To show the advantage, as an example, an implementation of nonlinear template is discussed. In our experiments an implementation of hole filling is demonstrated, and two nonlinear templates are evaluated by computer simulation. The experimental results shows ability of real-time image processing of CAM².