A Bio-Inspired Vision Front-End Chip with Spatio-Temporal Processing and Adaptive Image Capture

This paper presents an advanced CMOS imager with concurrent parallel processing for early-vision tasks. The network is arranged in two layers of 32times32 programmable mixed-signal elementary processors with programmable linear feedback and control masks and inter-layer connections for continuous-time cellular neural network dynamics. The ratio of the time constants of these layers is user selectable. There are also feedforward connections to a faster third layer, intended to combine of the outputs of the other two in parallel. We have employed a restricted set of weights, trading flexibility for robustness. It results in a more linear multiplier block and, consequently, a significant reduction of irregularities in the propagation of analog waves ought to asymmetric synapses. Also global and local adaptation to illumination conditions, both throughout the scene and from frame to frame, are implemented on-chip, making use of the available focal-plane processing capabilities. The predicted computing power per area and power consumption is amongst the largest reported, what renders this kind of devices as an efficient front-end for portable applications of artificial vision.