Blurring kernel extraction and super-resolution image reconstruction based on style generative adersarial networks

Abstract

The point spread function (PSF) is the main index used to evaluate the imaging resolution and further improve the quality of an optical image. Its measurement is significant for system development and pattern recognition. However, the precision of current measurement methods is low owing to a complicated modelling process, the pairing of various camera parameters, and disturbances by external factors. In this paper, we propose a method to extract blurring kernels and reconstruct super-resolution images based on style generative adversarial networks (StyleGANs). First, an improved StyleGAN model is introduced and an ideal blurry image generation model based on StyleGAN is trained to obtain a series of ideal Gaussian light-source images with a regular Airy disk; as the intensity distribution in the Airy disk is closer to its theoretical distribution. Second, the blurring kernels are extracted at different depth positions from the generated Gaussian light-source images to replace the PSF. This allows the evaluation of the blurry property of the optical system and effectively avoids the enrolment of noise in parameter identification or curve fitting in PSF representation. Finally, both the blurring kernels are used to deblur the blurry images of the Gaussian light source with a single wavelength and the blurry images of microbeads under visual light illumination at different depths based on the learnable convolutional half-quadratic splitting and convolutional preconditioned Richardson (LCHQS-CPCR) model. Compared to other image deblurring methods, our proposed method achieves high-resolution image reconstruction with blurring kernels from the generated optical images of the Gaussian light source.