Retinomorphic Vision Systems I: Pixel Design

I present and analyze test results from circuits that perform all four major operations performed by biological reti-nae using neurobiological principles: (1) continuous sensing for detection, (2) local automatic gain control for ampli-cation, (3) spatiotemporal bandpass ltering for prepro-cessing, and (4) adaptive sampling for quantization. In the retinomorphic system that I describe, all these operations are performed at the pixel level, to eliminate redundancy, to reduce power dissipation, and to make eecient use of the capacity of the output channel. The primary diierence between retinomorphic imagers and conventional ones is that retinomorphic imagers perform all operations at the pixel level 1]. The migration of more sophisticated signal processing down to the pixel level is driven by shrinking feature sizes in CMOS technology, allowing higher levels of integration to be achieved 2]. The retinomorphic approach uses the architecture and neuro-circuitry of the nervous system as a blueprint for building low-level vision systems|systems that are retinomorphic in a literal sense 1]. This approach results in integrated systems that ooer enriched functionality, by performing several functions within the same structure, and enhanced system-level performance using minimal-area devices (3L3L), by distributing computation across several pixels. The retinomorphic system described in this paper consists of two chips: a focal-plane image processor and a post-processor with a two-dimensional array of integrators. Both chips are fully functional; speciications and die photos are shown in Table 1 and in Figure 1. I describe the pixel design in Section 2, and the circuits that perform spatiotemporal bandpass ltering, local AGC, temporal integration, and adaptive quantization in Sections 3 through 6. My concluding remarks are in Section 7. The communication channel used to transmit asynchronous pulse streams between these two chips is described in the companion paper 3]. The circuitry in each pixel of the retinomorphic processor is shown in Figure 2. In general terms, the principles of operation are as follows: A CMOS-compatible, vertical bipolar phototransistor performs continuous sensing; its emmiter current is proportional to the incident light intensity 4]. Two current spreading networks 5, 6, 7] diiuse the pho-tocurrent signals over time and space; the rst layer (node V0) excites the second layer (node W0), which reciprocates by inhibiting the rst layer. The result is a spatiotempo