A design framework for hybrid approaches of image noise estimation and its application to noise reduction

Noise estimation is an important process in digital imaging systems. Many noise reduction algorithms require their parameters to be adjusted based on the noise level. Filter-based approaches of image noise estimation usually were more efficient but had difficulty on separating noise from images. Block-based approaches could provide more accurate results but usually required higher computation complexity. In this work, a design framework for combining the strengths of filter-based and block-based approaches is presented. Different homogeneity analyzers for identifying the homogeneous blocks are discussed and their performances are compared. Then, two well-known filters, the bilateral and the non-local mean, are reviewed and their parameter settings are investigated. A new bilateral filter with edge enhancement is proposed. A modified non-local mean filter with much less complexity is also present. Compared to the original non-local mean filter, the complexity is dramatically reduced by 75% and yet the image quality is maintained.