Image Denoising Networks with Residual Blocks and RReLUs

Discriminative learning methods have been widely studied in image denoising due to their swift inference and favorable performance. Nonetheless, their application range is greatly restricted by the specialized task (i.e., a specific model is required for each considered noise level), which prompts us to train a single network to tackle the blind image denoising problem. To this end, we take the advantages of state-of-the-art progress in deep learning to construct our denoising networks. Particularly, residual learning is utilized in our deep CNNs (convolutional neural networks) with pre-activation strategy to accelerate the training process. Furthermore, we employ RReLU (randomized leaky rectified linear unit) as the activation rather than the conventional use of ReLU (rectified linear unit). Extensive experiments are conducted to demonstrate that our model enjoys two desirable properties, including: (1) the ability to yield competitive denoising quality in comparison to specifically trained denoisers in several predetermined noise level and (2) the ability to handle a wide scope of noise levels effectively with a single network. The experimental results reveal its efficiency and effectiveness for image denoising tasks.