An analog CMOS network for Gaussian convolution with embedded image sensing

A circuit that samples an image in two dimensions and uses a resistive network to average it spatially according to a Gaussian weighting function whose width is controlled by an electronically variable resistor is described. A matrix of logarithmic photoreceptors fabricated on the surface of the IC drives currents into the nodes of a mesh consisting of appropriately ratioed positive and negative resistors. The resulting voltage distribution on the network nodes constitutes the smoothed image. No clocking or digital function is required in the action of the network. A prototype network of a 40*45 array of cells along a hexagonal grid is implemented in a standard 2- mu m p-well, double-metal, single-polysilicon CMOS process on a 7.9*9.2-mm die.<>