A Novel Accuracy Enhancement Approach for Single-photon 3D Imaging

Random coding single-photon 3D imaging with single photon avalanche diode (SPAD), which overcomes crosstalk and addresses the range ambiguity, has emerged as a candidate technology for light detection and ranging (LiDAR). However, resulting from low-flux conditions, it is challenged in restoring depth information with limited signal photons. To cope with this, we investigate a neural network method to capture correlations hidden in neighboring or global pixels. Specifically, we exploit a U-shape modulator to collect multi-scale features from fine-grained detail information and coarse-grained semantic information. Additionally, by leveraging the global correlations embodied in non-adjacent areas, we boost the performance by introducing a non-local block. To the best of our knowledge, our method is the first deep learning-based approach for random coding single-photon 3D reconstruction. Extensive experiments on simulated data demonstrate our method significantly reduces the RMSE by 91.3% to 93.1% in various scenarios, even under strict optical flux limits.