On the kernel function selection of nonlocal filtering for image denoising

Nonlocal filtering has been proved to yield attractive performance for removing additive Gaussian noise from the image by replacing the intensity value of each pixel via a weighted average of that of the full image. The key challenge of the nonlocal filtering is to establish the kernel function for computing the above-mentioned weighting factors, which control the quality of the denoised image result. In contrast to that the exponential function is used in the conventional nonlocal filtering, several new kernel functions are proposed in this paper to be further incorporated into the conventional nonlocal filtering framework to develop new filters. Extensive experiments are conducted to demonstrate not only that the kernel function is essential to control the performance of the algorithm, but also that the new kernel functions proposed in this paper yield superior performance to that of the conventional nonlocal filtering.