Image denoising with signal dependent noise using block matching and 3D filtering

In this paper, we propose a new method for image denoising. We use block matching 3D filtering (BM3D) to denoise the noisy image, and then denoise the noisy residual and merge this denoised residual into the denoised image. We can perform another BM3D to this merged image if the noise-level is still higher than a threshold. Our method performs similarly as the BM3D for Gaussian white noise, and it outperforms the BM3D, Poisson-Gaussian BM3D (PGBM3D), and Bivariate shrinking (BivShrink) for nearly all cases in our experiments for signal dependent noise. The method does not assume the noise to be Gaussian alone, and it works well for a mixture of Gaussian and signaldependent noise. However, the computational complexity of the new method is twice and at most three-times that of the standard BM3D for image denoising.