Image denoising using context quantization and local linear regression

A main challenge for image denoising techniques is the damaging of some specific detailed structures or edges that are useful for image application. That is why edge-preserving filtering has advanced as a prevailing topic in medical image and multimedia processing. Conventional edge-preserving filters have exploited a number of morphological operators and estimated order statistics in a set of variable local windows such that they both enhance the significant edge details and smooth additive and multiplicative noises meanwhile. However, they fail to take into account the importance of weak edges, and therefore treat them as additive or multiplicative noises to be reduced. To overcome this difficulty, we present a efficient image denoising algorithm by using context quantization and local linear regression techniques. The context quantization was conducted according to minimization of conditional entropy of the GAP prediction residual in quantized cells and the local texture features hidden in the contexts. In order to design a robust filter for pixels in each quantized context, the local linear regression technique has been applied. The experimental results validated that the proposed image denoising algorithm outperformed the conventional edge-preserving filters reviewed in this work.