Deep Residual Networks for Impulsive Noise Suppression

We introduce a new method of noise suppression using fully convolutional neural networks for salt and pepper noise. We adopt a well-known residual learning framework to get convergence faster in the training phase than conventional learning. Based on experimental results, our proposed method can outperform the state-of-the-arts method at most levels of noise contamination. In addition, our method is qualitatively proven to maintain the details of the denoised image even at extreme contamination level.

I. Introduction

In recent years, image capturing devices are widely available on the market. However, each device has different sensor quality and some have weaknesses. As a result, the captured image is vulnerable to noise contamination. One of possible noise contamination is salt-and-pepper which a type of impulsive noise contamination [1], This type of noise affects a portion of the image pixels by replacing the value with a very different one. In grayscale image, the very different values resemble sprinkles of salt and pepper which then referred as salt-and-pepper noise. Besides being not visually appealing, noisy images are not eligible to be data sets because their quality has decreased due to some corrupted image information.