Image noise level estimation by searching for smooth patches with discrete cosine transform

When denoising an image, noise level is one of the most vital input parameters, because setting wrong noise level affects a result of denoising. Therefore, the noise level must be estimated accurately. In this paper, we propose a new accurate noise level estimation method based on the characteristic of discrete cosine transform (DCT) coefficients. This characteristic is that the high-frequency coefficients of the smooth patches can be assumed to zero. We select smooth patches from a distribution of the standard deviation of high-frequency coefficients in a noisy image, and a distribution function of the standard deviation of high-frequency coefficients in an only noise image. Moreover, we propose a method that the estimated noise level is calculated from high-frequency coefficients in a noisy image. The experiment results with many images demonstrate that the proposed method estimates the noise levels more accurately, in comparison with the conventional methods.