A Robust Non-Blind Deblurring Method Using Deep Denoiser Prior

Yingying Fang, Hao Zhang, Hok Shing Wong, Tieyong Zeng; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2022, pp. 735-744

The existing non-blind deblurring methods are mostly susceptible to noise in the given blurring kernel, which is usually estimated from the observed image. This will produce undesirable ringing artifacts around the recovered edges when the given kernel is not accurate enough. Besides, the noise and outliers in the observed images may also degrade the performance of the deblurring methods seriously. Considering these factors, we designed a robust non-blind deblurring method taking all these noises into account. In this paper, we propose a kernel error term to rectify the given kernel at the time of performing the deconvolution. A residual error term is also introduced to deal with the outliers caused by noise or saturation. A deep learning denoiser prior is adopted to reserve the fine textures in the recovered image. The experiments show clearly that the proposed method achieves remarkable progress in both the visual quality and the numerical results of the recovered images compared to the state-of-the-art deblurring methods.