| Quantitative phase imaging technology can achieve marker-free,high-precision,noncontact biomedical optical imaging.Among them,the diffractive phase microscopy technology has been widely used in cell biology and material physics due to its stable common-path structure and compact off-axis mode,which has the advantages of fast acquisition rate,high time sensitivity,and strong anti-interference ability.However,there are still some imaging quality problems to be solved in diffraction phase microscopic imaging.Deep learning can autonomously learn data features to complete end-to-end mapping,which not only has high computational efficiency,but also provides high-quality solutions to optical computing imaging problems.To improve the quality of microscopic imaging,this thesis studies the quality optimization of diffraction phase microscopic imaging based on deep learning methods.Different deep neural network models were designed to optimize the quality of the halo-free for white light diffraction phase microscopy and auto-focus for diffraction tomography.By building a novel model architecture,adding optimization algorithms,and adjusting network hyperparameters,etc.,the personalized design of the deep neural network model was realized,thereby quickly and accurately improving the quality of microscopic imaging.The main research contents are as follows:1.According to the basic principle of diffraction phase microscopy technology,a diffraction phase microscopy system was designed and constructed.Through the analysis of the system principle and the calculation of the interference conditions,the parameters of the optical components and the optical path structure of the system were obtained,and then the basic principle of interferogram acquisition was analyzed and the process and method of phase image reconstruction were introduced.2.Aiming at the halo effect in the imaging process of white light diffraction phase microscopy,a deep neural network for halo-free was built to eliminate the halo effect based on the principle of the AutoEncoder.Based on the diffraction phase microscopy system,white light irradiation was used,and standard polystyrene beads and red blood cells were used as the research objects for experimental analysis.A deep neural network model was designed to train of the measured images by the beads samples and their corresponding halofree images,and different samples were used for blind measurement.By comparing the results with the existing methods of halo elimination,it was proved that the deep neural network can achieve rapid and accurate halo elimination and improve the imaging quality.3.Aiming at the defocus problem caused by sample rotation during tomography based on diffraction phase microscopy,a deep neural network based on U-Net structure was built to conduct autofocusing research.A sample rotation device is added to the laser diffraction phase microscopic system to form a tomography system.Through this system,ordinary single-mode fiber phase images of different angles were used as the research objects for experimental analysis.An auto-focus deep neural network model is designed to train the measured focused image and different degrees of defocused images,and different angles of defocused images were used for blind measurement.On the system,180 single-mode fiber phase imagess at multiple angles were re-continuously collected,and the high-quality phase images with focus were calculated using the trained deep neural network model.According to the principle of tomography,the processed two-dimensional phase images were reconstructed into a three-dimensional image,and a standard three-dimensional refractive index distribution map was established according to the simulation of the structural parameters of the sample.By comparing the theoretical three-dimensional refractive index distribution and reconstructed by deep learning method,it was proved that the method of deep learning can significantly improve the quality of three-dimensional imaging. |
PDF Full Text Request