Vision chip design is a challenging task. One should consider issues from visual processing algorithms to low level circuit design problems, from phototransduction principles to high-level VLSI architectural issues, from mismatch and digital noise to "readout" techniques, from optics to electronics and optoelectronics, from pure analog to mixed analog/digital to pure digital design problems, from biologically inspired vision models to intuitive models to computational models, ....
A vision chip requires photodetecting elements, image acquisition circuits, analog conditioning and processing circuits, digital processing and interfacing, and image readout circuitry all on the same chip. Many of these components, such as low level analog processing elements, should exist in number the same as photodetectors. In most cases these components should interact with at least their nearest neighbors. The area required for implementing the circuits and routing the information across the chip has put upper bounds on the realization of reliably functional and high resolution vision chips, and in most implementations resolution or functionality has been sacrificed for the other. The design of vision chips can obviously benefit from the high level integration in current VLSI processes, where more than 10 million transistors can be integrated on a single chip. Unfortunately, advanced processes for high level integration are usually tuned and characterized for leading edge digital processors and DRAMs, suffering from sub-micron effects, such as short channel effects, hot-carrier effects, band-to-band tunneling, gate-oxide direct tunneling, gate induced drain leakage (GIDL), drain induced barrier lowering (DIBL), and threshold voltage control [Fienga et al. 94]. Many available processes, on the other hand, do not have any specific photodetecting element, and are not well tuned for analog circuit design. Device mismatch has also severely affected the analog circuit design community, and almost no fabrication processes have been carefully characterized and modeled to account for mismatch.
Design of vision chips has also been affected by the lack of VLSI friendly computer vision algorithms. Most current computational computer vision algorithms are very complex and are even hardly implementable using powerful workstations to run in real time. Many computer vision algorithms are still not reliable enough for application in general uncontrolled environments. Biologically inspired algorithms, on the other hand, rely on the fact that many creatures have developed very efficient visual system. These algorithms, however, are not mature enough and suffer from excessive simplification caused by insufficient understanding of animals visual system.
Despite these facts, the design of single chip VLSI vision sensors, or smart vision sensors is increasingly progressing and many vision chips based on biological or computational algorithms have been developed in the past few years. The complexity of vision chips has also significantly increased, and 2D vision chips with more than 48,000 detectors and processing elements have been designed [Andreou and Boahen 94b].