Weighted MSE Based Spatially Adaptive BM3D

Weighted MSE (wMSE), recently introduced modification of MSE, is an image quality metric used to estimate visual quality of filtered images. It provides better than MSE correspondence to a human perception in consideration of distortions introduced by image filters. In this paper, wMSE is used both as a criterion to evaluate filtering efficiency of the modification of BM3D filter with spatially varying parameters, as well as to train a specially designed neural network to predict filters' parameters. Extensive analysis on three image datasets demonstrates that the proposed modification of BM3D provides lower values of wMSE than those of BM3D, both effectively suppressing noise in homogeneous regions as well as preserving fine details and texture.