Improved Blind Image Denoising with DnCNN.

Denoising convolutional neural networks (DnCNN) has been proposed recently for image denoising for additive white Gaussian noise with both blind and non-blind versions. For blind DnCNN, the networks are trained with noise levels from 0 to 55, which is not perfect for other noise levels. In this paper, we train the DnCNN with three noise ranges [0, 40], [40, 80], and [80, 120] separately to obtain three network models so that better denoising results can be achieved. The training of our new models can be done in parallel by taking advantages of GPUs. We choose the suitable network model according to the estimated noise level from the noisy images. Experimental results demonstrate that our proposed method outperforms the standard DnCNN for image denoising for almost all testing cases with all six testing images.