Abstract

Dynamic range compression is an essential feature of sensory neural networks, particularly in vision systems where many decades of mean input luminance must be accepted. Electronic realizations of neural network detector arrays, e.g., electronic retinas or other "smart" pixel arrays, normally employ individual, non-interacting (and therefore non-adaptive) logarithmic equalization elements for each pixel. In the present design, this equalization is provided directly by multiplicative, laterally inhibiting interconnections between the pixels themselves. This construction allows the equalization to adapt itself to mean luminance levels within local sections of the array. A 30-element linear sensory neural-network photodetector array has been constructed as a fully monolithic optoelectronic integrated circuit in GaAs to demonstrate this property. This prototype is based upon only one GaAs FET per unilateral interconnection and an additional three FET's and Schottky diodes for each pixel; thus, the implementation is extremely compact. Modulation of a self-conductance FET for each pixel also allows the threshold at which the inhibition engages to be externally adjustable.

© 1992 Optical Society of America