Object-Independent Image Restoration Based on Deep Learning

The optical imaging system is affected by both internal factors (such as manufacturing and alignment errors) and external factors (such as atmospheric turbulence), preventing it from reaching diffraction-limited imaging. The blind image restoration algorithms typically require lots of iterative computations, leading to poor real-time performance. Image restoration methods based on neural network have limited application scenario. To solve the above mentioned issues, this paper introduces an object-independent image restoration method based on deep convolutional neural network (DCNN). The DCNN utilizes object-independent three-channel image feature as its input. Without estimating the wavefront aberration modes, the blurred point spread function (PSF) is predicted by the DCNN. Subsequently, a deconvolution operation can be applied to restore the image. Not only the blurred PSF, the restored image is also incorporated in the loss function to enable effective image restoration under complex aberrations. It is proved that our method can achieve good image restoration results even for complex aberrations involving 35 Zernike aberration modes. Compared to traditional blind restoration algorithms, it exhibits advantages in restoration time and better SSIM and PSNR.