Understanding neural-network denoisers through an activation function perspective

Images are usually corrupted by noise during acquisition or transmission. Smoothing and thresholding within an adapted domain is a common practice to recover the clean image. Recently, many researchers try to train a denoiser using neural network on a large image training set. Some of them show competitive performance compared to state-of-the-art filter such as BM3D [1][2]. At the same time, neural network's performance has been improved with new activation functions, new solvers, etc., which makes machine learning techniques more reliable. In this work, we leverage cutting-edge techniques in the neural network world to reconsider the problem of image denoising with plain neural network. Instead of seeking blindly a better result, we focus on analysing and understanding the denoising strategies conceived by the neural network. We will provide a perspective through activation functions, and explain how they play a decisive role in the denoising mechanisms.