New tools for classification and evaluation of filtering errors in color image denoising

Performance analysis of color image denoising filters requires accurate measurements of many different effects produced during noise removal. Metrics in the literature consider only a subset of the filtering features that should be taken into account to address this issue. The novel set of metrics described in this paper aims at providing the necessary tools for analyzing the quality of a filtered picture from the point of view of residual noise, detail blur and color distortion. The approach adopts the YCbCr color space and performs the decomposition of the mean squared error (MSE) into six different components. Each MSE component focuses on a different class of filtering errors affecting the luminance or chroma channels of the filtered image. In order to validate the approach, the exact values of the MSE components are theoretically evaluated for some important nonlinear filters and used for a comparison. Computer simulations dealing with color pictures corrupted by Gaussian and impulse noiseshow that the results are in very good agreement with the theoretical values and that the method can represent a useful resource for analyzing the behavior of a denoising algorithm.