Physics-informed and self-supervised multi-image super-resolution reconstruction for digital holography microscopy

Abstract In digital holographic microscopy, the application of a polarization camera with a micro-polarizer array allows simultaneous phase shifting, thus the complex amplitude associated with the surface under test can be obtained by one-shot image acquisition. However, since the four phase-shifted holograms extracted from the polarization camera are sub-sampled, there exist lateral shift between different holograms, which in turn inevitably introduces error in the reconstructed results. To address this problem, a super-resolution reconstruction method is proposed for digital holographic microscopy. A diffraction physics model is combined with a deep learning framework, and the outputs are updated in a self-supervised manner. The proposed algorithm is verified with both simulated and experimental data, and the reconstruction accuracy can be improved by over an order of magnitude.