Blind estimation of signal-dependent noise parameters for color image database

Blind estimation of noise characteristics is an important image pre-processing stage for various applications. A model of signal-dependent noise is quite general and adequate for multichannel and color images. Existing techniques are either too complicated or not accurate enough. We propose and intensively test a method consisting of few rather simple stages and operations as pre-segmentation, accumulation of DCT-coefficients in quasi-homogeneous blocks, their normality check and aggregation. Testing is carried out for a recently created database of raw color images of natural scenes. It is demonstrated that the proposed method considerably outperforms existing counterparts and looses only to maximum likelihood estimation technique which is rather slow. Quantitative data proving this are presented.