Sensor-aware physical modelling for deep learning-based single-photon imaging

Deep learning-based single-photon imaging commonly relies on synthetic time-resolved histograms with incompletely specified sensing physics. We present a sensor-aware framework that represents the source-receiver response, spatially varying background, detector artifacts, and first-detection acquisition within a family of photon measurement operators. A controlled cross-condition protocol independently varies the sensing operators used for training and testing, enabling systematic evaluation of sensing-model mismatch across four representative reconstruction networks. An operator-induced risk formulation connects sensing perturbations to changes in reconstruction error and model ranking, while Monte Carlo verification, parameter-sensitivity analysis, and comparison with measured single-photon data provide complementary numerical and physical validation. Under C4 testing, matched C4 training reduces the architecture-averaged mean absolute error by 22.3% relative to C0 training. This effect persists in the same-dataset NYUv2 control, although with reduced magnitude. Experiments on measured WISC photon data further show pronounced architecture-dependent transfer across training sensing conditions. These results establish a reproducible framework for quantifying how sensing-model fidelity and reconstruction architecture jointly determine single-photon imaging performance.