Automated Gaussian filtering VIA Gaussian scale space and linear diffusion

Image denoising is challenging due to the difficulty to differentiate noise from image fine details. Convolution with a Gaussian mask is a widely used method for denoising. In this paper we propose, based on the relation between linear diffusion and Gaussian scale space, estimators of both the variance and window size of the discrete Gaussian filter applied to image denoising. To achieve content adaptive estimators, we also propose a structure under noise measure based on the median absolute deviation from the image gradient. Our simulations show that the proposed automated filter performs comparable or exceeds non-linear diffusion, while being of significantly lower complexity.