Study On The Microscopic Image Restoration Based On Generative Adversarial Networks

The aberrations induced by the inhomogeneity of the refractive index of the object and the limitations of imaging devices will cause blurred images and low resolution,which seriously hinders the development of microscopic imaging research.Although the method of correcting aberrations by optimizing the optical system of the microscope system can improve the resolution of the imaging system,it cannot easily overcome the physical limitations of the observation target,therefore,it is particularly important to improve the imaging quality of microscopic images by algorithms.Aiming at the problems of low resolution and poor imaging quality in microscopic imaging systems,firstly,this paper proposes a research method for microscopic image restoration based on Generative Adversarial Networks(GAN)by adding a denoising layer between the input layer and the convolutional layer.The denoising layer is used to reduce the influence of the noise of the microscopic image on the prior information,then the GAN restore the images captured by the Charge Coupled Device(CCD)in the actual situation and test the performance of the network and optimize it.The peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)of the images restored by the GAN with the denoising layer are 5% and 2% higher than the GAN without the denoising layer.Secondly,by simulating the influence of spherical aberration,coma,astigmatism,field curvature,defocus and the mixture of these aberrations on the point spread function of microscopic images under different coefficients,the influence of these single aberrations and mixed aberrations on the microscopic image quality are further analyzed,then the corresponding microscopic image dataset is made to provide aberration correction and recovery of subsequent microscopic images.Finally,by using the constructed GAN to train the microscopic images with different aberration coefficients,and use the trained network model to restore the microscopic images with aberrations.The PSNR of the restored microscopic images are increased from about 11 d B to 25 d B,and the SSIM of the restored microscopic images are also increased from about 0.2 to 0.6.In order to verify the universality of the network,the trained network is further used to restore the actually captured microscopic images,the PSNR of the restored microscopic images are increased by about 5d B,and the SSIM increased by 13%,so the quality of the microscopic images restored by GAN have been significantly improved.