They don't even need to go full smartphone processor and OS, adding something GPGPU-esque to offload the computations to would work. Deep Neural Nets brought us many "edge TPUs", I don't see why they wouldn't equally work to process images quickly and they're reprogrammable.