A fast yet reliable noise level estimation algorithm using shallow CNN-based noise separator and BP network

To date, a large number of research works have been conducted on noise level estimation (NLE) that automatically and accurately estimates the unknown noise level for an observed noisy image. Nevertheless, the state-of-the-art NLE algorithms are still limited in efficiency, which will undermine the overall execution performance of the subsequent denoiser. By making full use of the powerful nonlinear modeling capabilities of convolutional neural networks (CNNs), a shallow CNN-based noise separator with high execution efficiency for natural images was proposed to obtain the coarse noise component (difference image) from a single observed noisy image. Based on the fact that the coarse noise component tends to follow a Gaussian-like distribution, we chose to model it with the generalized Gaussian distribution model, whose parameters are strongly sensitive to noise level and can be treated as features to characterize the degradation degree of a given noisy image. As such, the extracted features were instantly mapped to their corresponding noise level via a back-propagation (BP) neural network pre-trained on the representative training samples, resulting in a fast yet reliable NLE algorithm. Experiments demonstrate that our training-based NLE algorithm exploiting the shallow CNN-based noise separator and BP network outperforms the state-of-the-art counterparts on estimating noise level with the least executing time over a wide range of image contents and noise levels, providing a highly effective solution to blind denoising as the preprocessing module of a non-blind denoiser.