Spatial-frequency domain joint learning reconstruction for coded aperture temporal compressive femtosecond holographic microscopy

【摘要】 Snapshot compressive imaging(SCI) reconstructs high-speed dynamic scenes from a single compressed measurement.When deployed in digital holography, existing reconstruction algorithms suffer from spectral biases, which cause oversmoothing of high-frequency components, resulting in poor recovery of phase distribution. To address these limitations,we propose a spatial-frequency domain joint learning network(SFD-JLNet) to reconstruct holograms in a coded aperture temporal compressive femtosecond holographic microscopy(CATC-FHM) system. SFD-JLNet constructs a global representation of frequency information by performing a 3D discrete Fourier transform on the extracted features, followed by convolutional operations on the obtained intensity spectrum and phase spectrum. Simultaneously, it models the spatial features of frame sequences using convolutions and a temporal Transformer in the spatial domain. Then, it adaptively fuses features from both domains to establish spatial-frequency correlations, improving the recovery of high-frequency details and enhancing phase fidelity. Simulation and experimental results demonstrate that SFD-JLNet outperforms stateof-the-art methods in CATC-FHM hologram reconstruction, achieving 4.54 dB peak signal-to-noise ratio(PSNR) gain and 47.7% high-frequency error norm(HFEN) reduction. This advancement enables high-speed three-dimensional imaging applications such as laser processing and in vivo cellular imaging.更多还原

【Abstract】 Snapshot compressive imaging(SCI) reconstructs high-speed dynamic scenes from a single compressed measurement.When deployed in digital holography, existing reconstruction algorithms suffer from spectral biases, which cause oversmoothing of high-frequency components, resulting in poor recovery of phase distribution. To address these limitations,we propose a spatial-frequency domain joint learning network(SFD-JLNet) to reconstruct holograms in a coded aperture temporal compressive femtosecond holographic microscopy(CATC-FHM) system. SFD-JLNet constructs a global representation of frequency information by performing a 3D discrete Fourier transform on the extracted features, followed by convolutional operations on the obtained intensity spectrum and phase spectrum. Simultaneously, it models the spatial features of frame sequences using convolutions and a temporal Transformer in the spatial domain. Then, it adaptively fuses features from both domains to establish spatial-frequency correlations, improving the recovery of high-frequency details and enhancing phase fidelity. Simulation and experimental results demonstrate that SFD-JLNet outperforms stateof-the-art methods in CATC-FHM hologram reconstruction, achieving 4.54 dB peak signal-to-noise ratio(PSNR) gain and 47.7% high-frequency error norm(HFEN) reduction. This advancement enables high-speed three-dimensional imaging applications such as laser processing and in vivo cellular imaging.更多还原