Learning-Based Noise Component Map Estimation for Image Denoising

A problem of image denoising, when images are corrupted by a non-stationary noise, is considered in this paper. Since, practice, no priori information on noise available, statistics should be pre-estimated prior to denoising. In paper, deep convolutional neural network (CNN) based method for estimation map local, patch-wise, standard deviations (so-called sigma-map ) proposed. It achieves the state-of-the-art performance accuracy sigma-map case as well variance an additive white Gaussian noise. Extensive experiments denoising using estimated sigma-maps demonstrate that our outperforms recent CNN-based blind methods up 6 dB PSNR, other 0.5 dB, providing, at same time, better usage flexibility. comparison with ideal case, applied ground-truth sigma-map, shows difference corresponding PSNR values most levels within 0.1-0.2 and does not exceed 0.6 dB.