Asymmetric, Non-unimodal Kernel Regression for Image Processing

Abstract

Kernel regression has been previously proposed as a robust estimator for a wide range of image processing tasks, including image denoising, interpolation and super resolution. In this article we propose a kernel formulation that relaxes the usual symmetric and unimodal properties to effectively exploit the smoothness characteristics of natural images. The proposed method extends the kernel support along similar image characteristics to further increase the robustness of the estimates. Application of the proposed method to image denoising yields significant improvement over the previously reported regression methods and produces results comparable to the state-of the-art denoising techniques.