Study Of Blind Image Restoration Methods Of Adaptive Double Regularization And Deep Learning In Fluorescence Microscopy

In fluorescence microscopy imaging,to avoid the inactivation of fluorescent molecules and observe long-term cellular activities,low-dose fluorescent dyes are usually used.Images captured with low-dose fluorescent dyes will also suffer from significant noise.Moreover,PSFs obtained through traditional methods might be inaccurate and not suitable for the deconvolution task of deblurring.We focus on the semi-blind restoration method of regularization and blind restoration method of deep learning techniques without accurate PSFs for image denoising and deblurring.Finally,we successfully apply these proposed methods to the wide-field illumination microscopy and structured illumination microscopy(SIM).In the wide-field illumination microscopy,bead experiments and model fitting are two routine ways to obtain the system PSF.Even if there might be some errors in PSFs obtained from these methods,they can be used as the prior knowledge,within an acceptable error range defined in our work.In this work,we propose to apply the regularization method based on Bayes' theorem,and design an adaptively adjusted prior constraint according to traditionally obtained PSFs.Compared with traditional methods,this adaptive regularization method can restore a more accurate PSF and a higher-quality image,which reflects greater robustness.The SIM imaging relies on the discrete Fourier transform of the PSF,also known as Optical Transfer Function(OTF).OTF mainly affects the frequency band components separated by decoupling in SIM reconstruction.It is also very important to ensure the accuracy of the band information.We also propose the adaptive regularization method for SIM reconstruction,which makes it possible to achieve better reconstruction results in the case of inaccurate OTFs.As deep learning has achieved a remarkable improvement in the field of computer vision,a method based on generative adversarial network is proposed to emphasize the learning of high-frequency components of images in this work.For traditional end-to-end learning methods,it is difficult to learn multiple mappings from degraded images to clear images,resulting in images of poor quality.We emphasize the importance of predicting these high-frequency components,and give larger penalties on them for image restoration.Our proposed method can restore more accurate image details and greatly improve the quality of reconstructed fluorescence images than other end-to-end deep learning methods without accurate PSFs.In this paper,through experiments carried out on corresponding datasets,proposed regularization and deep learning methods have achieved better performance,which illustrates the effectiveness of our methods.