Noise variances and regularization learning gradient descent network for image deconvolution

Abstract

Existing image deblurring approaches usually assume uniform Additive White Gaussian Noise (AWGN). However, the noise in real-world images is generally non-uniform AWGN and exhibits variations across different images. This work presents a deep learning framework for image deblurring that addresses non-uniform AWGN. We introduce a novel data fitting term within a regularization framework to better handle noise variations. Using gradient descent algorithm, we learn the inverse covariance of the non-uniform AWGN, the gradient of the regularization term, and the gradient adjusting factor from data. To achieve this, we unroll the gradient descent iteration into an end-to-end trainable network, where, these components are parameterized by convolutional neural networks. The proposed model is called the noise variances and regularization learning gradient descent network (NRL-GDN). Its major advantage is that it can automatically deal with both uniform and non-uniform AWGN. Experimental results on synthetic and real-world images demonstrate its superiority over existing baselines.