A Low-Power Imager and Compression Algorithms for a Brain-Machine Visual Prosthesis for the Blind

We present a synchronous time-based dual-threshold imager that experimentally achieves 95.5 dB dynamic range, while consuming 1.79 nJ/pixel/frame, making it one of the most wide-dynamic-range energy-efficient imagers reported. The imager has 150x256 pixels, with a pixel pitch of 12.5μm x 12.5μm and a fill factor of 42.7%. The imager is intended for use in a brain-machine visual prosthesis for the blind where energy efficiency and power are of paramount importance. Such prostheses will also need to convey visual information to patients with relatively few electrodes and in a manner that minimizes electrode interactions, just as cochlear implants have accomplished for deaf subjects. To achieve these goals, we present a strategy that compresses visual information into the basis coefficients of a few image kernels that encode enough information to provide reasonably good image reconstruction with 60 electrodes. The strategy also uses time-multiplexed stimulation of electrodes to minimize channel interactions like the continuous interleaved sampling (CIS) strategy used in cochlear implants. Some of the image kernels that we employ are similar to the receptive fields observed in biology and may thus be natural to learn, just as cochlear-implant subjects have learned to reconstruct sound from a few filter basis coefficients.