Deep learning-driven recovery of oversaturated interferometric signals for continuous 3D morphology measurement of complex surfaces in OFC-TS-DFT systems

To address signal oversaturation caused by the limited dynamic range of photodetectors in optical frequency comb (OFC) time-stretch spectral interferometry, we design a Uformer 1D network that integrates U-Net and Transformer architectures for signal recovery. The model is trained via a three-stage strategy: large-scale ideal-simulation pretraining, real paired-data fine-tuning with frequency-domain and total variation losses, and self-supervised learning on unlabeled real signals to enhance generalization. Experiments show that the recovered signals achieve a mean measurement error of ∼0.2 µm, with final average log-RMSEs of 0.065 (training) and 0.057 (validation), extending the effective dynamic range from 10 dBm to 15 dBm. Real-time surface reconstruction of a one-yuan coin further demonstrates its potential for continuous, real-time, high-dynamic-range 3D morphology measurement.