Median filtering is a widely used non-linear noise-filtering algorithm, which can
efficiently remove salt and pepper noise while it preserves the edges of the objects.
Unlike linear filters, which use multiply-and-accumulate operation, median filter
sorts the input elements and selects the median of them. This makes it computationally
more intensive and less straightforward to implement. This paper describes several
algorithms which could be used on parallel architectures and propose a histogram based
algorithm which can be efficiently executed on GPUs, resulting in the fastest known
algorithm for medium sized filter windows. The paper also presents an optimized sorting
network based implementation, which outperforms previous solutions for smaller filter
window sizes.
