Median Filtering is a classic algorithm. Today, I read the paper and I know that there is also an algorithm called adaptive median filter ramf. The original paper is here.
Ramf processes images in the following two steps.
1. First determine the maximum filtering radius, and then use an appropriate radius R to filter the image. Calculate the Imin, IMAX, and imed of the pixel gray scale of the Current Filter radius, and then judge whether imed is in the middle of [Imin, IMAX]. If imed is in, perform downward, otherwise, the current radius is extended until the R is equal to the maximum filtering radius.
2. If the currently processed pixel IMG (I, j) is between [Imin, IMAX], the current pixel is output; otherwise, the value of the current filtering radius pixel imed is output.
Let's see the effect:
Noisy Image:
Ramf algorithm:
Common 3*3 median filtering:
The Matlab code is as follows:
Clear all; close all1_clc1_img1_mat2gray(imread('lena.jpg '); [m n] = size (IMG); IMG = imnoise (IMG, 'Salt & pepper', 0.1 ); % added the salt and pepper noise imshow (IMG, []); Nmax = 10; % determined the maximum filtering radius %. below is the boundary extension, and Nmax pixels are added to the image from top to bottom to left. Imgn = zeros (m + 2 * Nmax + 1, n + 2 * Nmax + 1); imgn (Nmax + 1: m + Nmax, Nmax + 1: N + Nmax) = IMG; imgn (1: Nmax, Nmax + 1: N + Nmax) = IMG (1: Nmax, 1: N); % extended upper boundary imgn (1: m + Nmax, N + Nmax + 1: n + 2 * Nmax + 1) = imgn (1: m + Nmax, N: N + Nmax ); % extend the right boundary imgn (m + Nmax + 1: m + 2 * Nmax + 1, Nmax + 1: n + 2 * Nmax + 1) = imgn (M: m + Nmax, nmax + 1: n + 2 * Nmax + 1); % extended lower boundary imgn (1: m + 2 * Nmax +: Nmax) = imgn (1: m + 2 * Nmax + 1, Nmax + * Nmax); % expanded left Boundary Re = imgn; for I = Nmax + 1: m + Nmax for J = Nmax + 1: N + Nmax r = 1; % initial filtering radius wh Ile R ~ = Nmax W = imgn (I-r: I + R, J-R: J + r); W = sort (w); Imin = min (w (:)); IMAX = max (w (:)); imed = W (uint8 (2 * r + 1) ^ 2/2 )); if Imin <imed & imed <IMAX % if the median in the current neighborhood is not a noise point, use the break in the current neighborhood; else r = R + 1; % otherwise, expand the window, continue to judge end if Imin