1. INTRODUCTION
From the moment a raw image is acquired until the final JPEG picture is obtained, a complex processing chain is applied. Each of these operations alters the noise model. Indeed, the initial Poisson-Gaussian noise [1] undergoes several operations resulting in a complex noise model. The final stage in most digital images consists in JPEG compression. Indeed, to be stored or transferred in a reasonable amount of time, images must undergo a compression step. Many noise estimation algorithms [2]–[5] suppose that noise can be estimated using only the high frequency coefficients of small patches in an image. However, this is not true for JPEG-compressed images, since the quantization step during the compression process attenuates these high frequencies. Indeed, for JPEG images, noise variance decreases as the frequency increases. As a result, these noise estimation methods yield to an inaccurate estimation of the noise. Noise estimation is a mandatory step of countless image processing tasks such as denoising [6], [7], forgery detection [8], [9], anomaly detection [10], PRNU extraction [11] and steganography [12], just to mention a few. The performance achieved by the methods developed to tackle each of these tasks depend on how accurately they are able to estimate noise.