Noise variance estimation in nonlocal transform domain

We consider the estimation of the variance of an additive white Gaussian noise corrupting an image. In the proposed approach, we exploit the nonlocal self-similarity of images to achieve an improved separation of noise and signal. In particular, we utilize the same adaptive 3-D transform decomposition used in the BM3D (block-matching and 3-D filtering) denoising algorithm, where mutually similar blocks are stacked together and jointly processed. An adaptive-size portion of the high-frequency ends of the 3-D transform spectra is retained and used as input sample for a robust median estimator of the absolute deviation. Experimental analysis demonstrate a state-of-the-art accuracy of the proposed approach.