Weighted Mse Based Metrics for Characterization of Visual Quality of Image Denoising Methods

We propose modifications of standard MSE and some other metrics related to it for estimation of visual quality of images filtered by different methods. The modified metrics are based on assumption that it is important for a human observer that details and edges in a filtered image are not smeared. We propose to apply a weighted version of MSE as follows. For image fragments where differences of filtered image values with respect to noise-free image are larger than for the corresponding noisy image the weights are set larger than for other image fragments. An experiment on estimation of filtered image visual quality is carried out for a test set containing 72 images. The weight value equal to 5 is recommended. It is shown that such method of calculating MSE can be used for modifying the metrics PSNR, PSNR-HVS, PSNR-HVS-M. This allows increasing their correspondence to human perception and evaluation of filtered images visual quality.