Deep learning-based spectral extraction algorithm for off-axis digital holographic signal terms

Digital holographic microscopy is widely applied in the study of the dynamic morphology of microscopic objects due to its characteristics of non-contact measurement, high-resolution three dimensional morphology detection, and real-time observation. This paper introduces a novel off-axis digital holographic signal spectrum extraction method based on deep learning, named DL-SEDH. The DL-SEDH algorithm is trained using approximately 1000 signal spectra from a USAF 1951 resolution test target, successfully achieving adaptive extraction of signal spectra and effective suppression of interference components. To validate its effectiveness, experiments are conducted using the USAF 1951 resolution test target and onion epidermal cells as research subjects. The experimental results demonstrate that DL-SEDH not only rapidly and accurately selects signal spectra while suppressing interference frequency components, especially coherent noise distributed near the signal spectra, but also exhibits higher accuracy, robustness, and applicability compared to traditional methods. Validation on holograms not used during training confirms the effectiveness of DL-SEDH in phase reconstruction quality. The proposed DL-SEDH method introduces innovation to the off-axis digital holography field, holding significant practical value and providing an efficient and precise solution for phase reconstruction in digital holographic microscopy.