Improving deep image denoising using pseudo-ground-truth images

Deep neural networks, especially convolutional neural networks, have been successfully applied to image denoising tasks. Along with advances in network architectures, many attempts have been made to find an alternative loss function to the widely used L1-loss and L2-loss. However, the perception-distortion tradeoff was recently demonstrated; thus, advanced loss functions such as adversarial loss from generative adversarial networks can only improve the perceptual image quality at the expense of distortion. This Letter shows that distortion can be further decreased when an image denoising network is trained using modified versions of ground-truth (GT) (defined as pseudo-ground-truth (PGT)) images that are obtained by combining the original GT images and initially denoised images. Experimental results show that the proposed denoising network that is trained to predict both PGT and GT images produces denoised images closer to GT images.