On-chip vision
pp. 83-103
in: Virginio Cantoni (ed), Human and machine vision, Berlin, Springer, 1994Abstract
This paper reviews the ever increasing activity in smart sensors, initiated in the 80's by Mead's silicon retina design. The gap between the sensors and the processors in a traditional image processing system can be partially reduced by embedding into the same silicon die both the transduction of incoming illumination and its interpretation through some processing function. If we consider the silicon die a "real estate" on which to build the smart sensor, many possible alternatives arise, which either privilege the sensing function or the processing one. On-chip-vision is the domain of integration that tries to balance the available silicon resources between these two extreme approaches. Analog processing is a viable solution for many tasks, though it leads to implementations that lack a general programmability. Application specific solutions, according to the ASIC paradigm, take a more flexible approach in mixing both analog and digital processing, according to the single task embedded in the silicon die. General purpose digital retinas, which combine an array of sensors with a programmable digital cellular array of simple processors, have also been proposed and built; the area trade-off between the sensing function and the near-neighbour connectivity suited to low level image processing is a major issue for further research.