The acquisition of an image from the sensor of a camera introduces a noisy component due to the photon counting process and the electrical and thermal effects in the sensor. Therefore, a digital camera pipeline has to consider also a denoising procedure to remove the noise from the acquired image.
In literature, many denoising approaches has been proposed, exploiting, for instance, the Wiener filtering, the total variation regularization or the bilateral filters. Most of the denoising approaches apply linear transformations that are able to separate low and high frequencies (such as the wavelets, the curvelets or the contourlets). In fact, it is observed that natural images concentrate the energy in the lowest frequencies, while in the high-frequency subbands the energy is localized only in correspondence of the details of the image. On the other hand, if the noise is assumed white, its distribution is constant over all the spectrum. Therefore, many denoising approaches apply a linear transformation (such as the wavelets), then "clean" the high-frequency subbands simply thresholding them. Finally, an inverse transformation of the "cleaned" coefficients is done to obtain the denoised image. This procedure is based on the idea that large coefficients, which are kept, belong to the details of the image, whereas the noise is distributed across small coefficients, which are canceled.