Convolutional neural network with median layers for denoising salt-and-pepper contaminations

We propose a deep fully convolutional neural network with new type of layer, named median to restore images contaminated by salt-and-pepper (s&p) noise. A layer simply performs filtering on all feature channels. By adding this kind into some widely used networks, we develop an end-to-end that removes extremely high-level s&p noise without performing any non-trivial preprocessing tasks. Experiments show inserting layers simple fully-convolutional the L2 loss significantly boosts signal-to-noise ratio. Quantitative comparisons testify our outperforms state-of-the-art methods limited amount training data.