Image compression by orthogonal decomposition and dynamic segmentation using cellular nonlinear network chips

A method is shown using the CNN chip-set hardware architecture for the implementation of a high-speed, low bit-rate image coding system. A simple and fast algorithm is introduced to generate basis functions of 2 dimensional (2D) orthogonal transformations. Using the 2D basis functions of the Hadamard or Cosine functions, the transformation coefficients of the basic block of the image are measured by the CNN. Meanwhile, the CNN can produce the inverse transformation of the measured coefficients and the actual distortion-rate can be computed. If a required distortion-rate is reached, the coding process could be stopped (the use of even more coefficients would increase bit-rate needlessly). Effects of noise and VLSI computing accuracy are also considered to optimise the architecture. We also give a short description of how to join the transform coding method and the object-oriented image model