Noise Blind Deep Residual Wiener Deconvolution network for image deblurring

Abstract

The Deep Wiener Deconvolution Network (DWDN) provides a simple and effective approach for non-blind image deblurring by performing the classical Wiener filtering in deep feature domain. However, it needs estimation of signal-to-noise ratio (SNR), which is obtained under the uniform Gaussian noise assumption. This paper presents the Residual Wiener Deconvolution (RWD) network, which reformulates Wiener deconvolution into two successive operations: deconvolution and denoising. To avoid explicit estimation of SNR, the denoising operation is parameterized by a network, in which the SNR is estimated. The RWD network is then combined with the encoding/decoding network of DWDN+, resulting in an end-to-end trainable model called Noise Blind Deep Residual Wiener Deconvolution (NBDRWD) network. Experimental results show that, the proposed NBDRWD significantly outperforms related baselines in deblurring images corrupted by uniform Gaussian noise, non-uniform Gaussian noise, JPEG compression artifacts, and real blur.