High-density CCD neurocomputer chip for accurate real-time segmentation of noisy images

摘要

Neurocomputer chips hold the promise of providing the beyond-supercomputer processing power required for solving a variety of problems in machine vision. In particular, it is proposed that a high-density neurocomputer-chip architecture can be designed and fabricated using CCD technology that has the potential for accurately segmenting noisy images at speeds much faster than the video frame rate. The design is based on the line-process model of Geman and Geman (1984), and the estimated chip computational performance is expected to be significantly better in speed and pixel density than previous designs. The CCD design uses charge sharing to implement a relaxation process and threshold-dependent charge sharing to implement line processes. Parameter selection and simulation comparisons are presented.