Keep it sparse: neural spiking demodulator in event optical camera communications

This paper presents a spiking neural network (SNN)-based demodulator for event-camera optical camera communication (EOCC) using pulse-shaped waveform modulation. Information is encoded using four pulse-shaping waveforms, each carrying 2 bits per symbol. The receiver segments the resulting event stream into symbol windows, converts the positive-polarity events into binary spatio-temporal spike tensors, and classifies them with a compact leaky integrate-and-fire SNN using rate-coded output decisions. Experimental evaluation was performed in a controlled indoor static setup at receiver-plane illuminances of 100, 350, and 500 lux and symbol rates of 53, 100, and 150 symbols/s. For the best configurations in the explored design space, no bit errors were observed at 500 lux across all tested symbol rates, nor at 350 lux at 53 and 100 symbols/s for suitable region-of-interest (ROI) radii; at 100 lux the 150 symbols/s condition was limited by incomplete symbol extraction rather than by the demodulator. Evaluated on the same recordings, the same event representation and the same training split, the Gaussian-mixture matching bank previously reported for this modulation format sustains zero observed bit errors only within a two- to three-pixel window of ROI radii at 500 lux, and never falls below 10^-2 at 100 lux, where the proposed receiver reaches 1.8 × 10^-3, more than five times lower. The gap is structural: the template likelihood is a product over received events, giving the matching bank a sharp optimum in ROI radius, whereas the trained network attenuates channels that carry no signal. Learning the demodulator directly from sparse events therefore widens the operating region of the link rather than only matching it at its best point.