A Static Phase-Bank CNN Demodulator for Rolling-Shutter Optical Camera Communication in Intra-Satellite Links

This work proposes and evaluates a convolutional neural network (CNN) for demodulating LED-transmitted on–off keying signals captured by a rolling-shutter camera, motivated by short-range intra-satellite optical communication. The proposed Static Phase Bank predicts payload bits and their validity directly from camera images, without an external intensity threshold or payload-start detector. A central objective is to investigate training with synthetic frames. Real recordings are used to calibrate the appearance of generated rolling-shutter images, whose payload locations and transmitted bits are known by construction. The CNN is then trained exclusively on these synthetic frames and evaluated on real recordings at 30 lx, without fine-tuning. It achieves a frame-local reference-based BER of 1.79×10−3 and correctly recovers all labelled payloads and slot-validity decisions in 97.6% of test frames. A separate experiment demonstrates that one CNN trained jointly on 42 lx and 4 lx recordings can demodulate both conditions using the same network parameters. Real-data results are measured against reviewed proxy labels and do not establish general robustness or an accuracy advantage over conventional receivers. Nevertheless, the study demonstrates two practical contributions of the CNN approach: training on synthetic frames while retaining useful demodulation performance on real images, and incorporating multiple illumination conditions into a single receiver without condition-dependent model switching or threshold adjustment.