Deep noise estimation and removal for real-world noisy images

Abstract

This paper proposes a joint noise estimation and removal network for real-world noisy images. We observe that the noise variance depends on the image intensity and the variances of neighboring pixels are not similar due to the demosaicing interpolation. Therefore, for noise estimation, we propose to pack each 2 × 2 noisy block into a 4D vector as the input of the noise level estimation network, which makes the noise variances of adjacent pixels to be similar. This strategy not only reduces the computational complexity but also improves the estimation accuracy. For noise removal, we propose to concatenate the estimated noise variance with the noisy image as the input of the noise removal network, which takes advantages of the correlations between noise levels and image intensities. In addition, we propose a RSD (Residual SE D-Conv) block, which explores the spatial and channel correlations of realistic noise. Considering the realistic noisy and clean image pairs are limited for training, we utilize synthetic noisy images to train the network first and then fine tune the network with real images. Experimental results demonstrate that the proposed method outperforms competing methods for both noise estimation and removal tasks.