Research On End-to-end Digital Holographic Reconstruction Technology Based On Deep Learning

Slightly off-axis digital holography can solve both the problem of low bandwidth product of off-axis digital holography and the problem of requiring multiple additional holograms for coaxial digital holography.Also,by combining with the two-step phase shift method,it is widely used because it can preserve the original spatial bandwidth,increase the selectable range of bandpass filter sizes,and the resolution accuracy of reconstructed phase images.However,traditional digital holographic reconstruction algorithms usually require the acquisition of background maps or additional holograms,as well as precise a priori information or complex hardware structures,which makes it impossible to combine the quality and speed of phase reconstruction.Therefore,this thesis combines deep learning techniques with slightly off-axis digital holographic reconstruction techniques to achieve a fast end-to-end high-quality image reconstruction process from out-of-focus holograms to focused reconstructed images,and the proposed algorithm does not require background maps or additional holograms.The reconstruction speed,reconstruction quality,and holographic field of view are also guaranteed while maintaining a high bandwidth product.Only one hologram is required to complete the reconstruction process,and the sensitivity to noise and phase wrapping is low.The details of the study are as follows:Firstly,based on the slightly off-axis digital holography and the basic theory of white light diffraction phase microscopy system,simulated,experimental and double cross-sectional phase type holographic images were acquired and produced,and the holographic reconstruction and autofocusing were realized using angular spectrum method,phase shift method plus cosine fractional self-focusing algorithm,and the reconstructed images obtained were used as markers of the network to rewrite the data set in TFRecords format.Then,by studying the basic theory of VGG and Res Net as well as their advantages and disadvantages,a high quality holographic reconstruction network High Quality Network(Hi Qu-Net)is proposed and successfully applied to the field of digital holographic reconstruction.Hi Qu-Net,in addition to its strong anti-noise capability,also has higher evaluation indexes compared to other algorithms and reconstruction quality.VGG and Res Net structures are both strong feature extraction networks for classification tasks,and the domain is migrated to holographic reconstruction tasks,followed by full validation of the feasibility,practicality and robustness of the algorithm by simulated phase data,experimental phase data and dual-section phase data.Finally,by studying the basic theory of U-Net as well as its advantages and disadvantages,a Fast U-shaped Network(FU-Net),a fast holographic reconstruction network,is proposed for the problem of high reconstruction quality but slow reconstruction speed of Hi Qu-Net.Compared with Hi Qu-Net,FU-Net greatly improves the reconstruction speed with a small loss of performance and has a faster reconstruction speed compared with other algorithms as well as Hi Qu-Net algorithm.The computation time for FU-Net to complete one forward propagation under CPU experiment is only 0.97 s,and the fps is approximately equal to 1,which makes online holographic reconstruction or video holographic reconstruction possible.The feasibility,practicality,and robustness of the algorithm are fully verified by simulated phase data,experimental phase data,and dual-section phase data.