C-MEX for 2D Rolling Median
MATLAB/MEX wrapper for fast rolling median by ashelly
based on min/max heap
see https://gist.github.com/ashelly/5665911 also https://stackoverflow.com/questions/5527437/rolling-median-in-c-turlach-implementation
to compile in MATLAB type "mex -V -O -largeArrayDims rollingMedian.c"
B = rollingMedian(A, R, C) Performs median filtering of the matrix A in two dimensions with minimal edge effects and phase shift.
A : Input Array Dimensions Allowed: (M x N), (M x N x ?), (M x N x ? x ?), ... As long as the leading dimensions of A (M & N) are nonzero, the filter will operate on all trailing dimensions. R : Filter Window Rows (1 < R < M / 2) C : Filter Window Cols (1 < C < N / 2)
B : Output Array with the same dimensions and class as A.
rollingMedian uses a median-heap to compute the rolling median rather than a sorting approach (i.e. sort all elements for each window). The time complexity of a sorting approach (for e.g. quicksort, mergesort) is O(MNRClog(RC)). The time complexity of the median heap approach is O(MNlog(RC)).
The left and right edges (1) are filtered first using successively wider filter windows for all pixels whose col index is less than C/2. The top and bottom edges (2) are filtered second using successively taller filter windows for all pixels whose row index is less than R/2.
The algorithm operates on 4 pointers simultaneously (one for each of the top-left, bottom-left, top-right, and bottom-right of the array) and moves from the edges of the array inward. This creates a south-east phase shift in the top-left quadrant, a north-east phase shift in the bottom-left quadrant, a south-west phase shift for the top-right quadrant, and a north-west phase shift in the bottom right quadrant. This may create distortion at N/2 if C is even, and M/2 if R is even. If M or N is odd, the median windows from both sides are advanced one row or col and the average of both sides is used.
1a: cols 0 to C/2-1, rows 0 to M/2-1
1b: cols 0 to C/2-1, rows M-1 to M-M/2 (reverse)
1c: cols N-1 to N-C/2 (reverse), rows 0 to M/2-1
1d: cols N-1 to N-C/2 (reverse), rows M-1 to M-M/2 (reverse)
1B: if M%2 : (cols 0 to C/2-1, row M/2) & (col N-1 to N-C/2 (reverse), row M/2)
2a: cols C/2 to N/2-1, rows 0 to R/2-1
2b: cols C/2 to N/2-1, rows M-1 to M-R/2 (reverse)
2c: cols N-C/2-1 to N-N/2 (reverse), rows 0 to R/2-1
2d: cols N-C/2-1 to N-N/2 (reverse), rows M-1 to M-R/2 (reverse)
2B: if N%2 : (col N/2, rows 0 to R/2-1) & (cols N/2, rows M-1 to M-R/2 (reverse))
3a: cols C/2 to N/2-1, rows R/2 to M/2-1
3b: cols C/2 to N/2-1, rows M-R/2-1 to M-M/2 (reverse)
3c: cols N-C/2-1 to N-N/2 (reverse), rows M/2 to M/2-1
3d: cols N-C/2-1 to N-N/2 (reverse), rows M-R/2-1 to M-M/2 (reverse)
3B: if N%2 : (col N/2, row R/2 to M/2-1) & (col N/2, rows M-R/2-1 to M-M/2 (reverse))
3C: if M%2 : (cols C/2 to N/2-1, row M/2) & (cols N-C/2-1 to N-N/2 (reverse), row M/2)
3D: if M%2 & N%2: average of 3B & 3C at (col N/2, row M/2)
-> -> -> -> -> -> -> -> -> -> -> -> v <- <- <- <- <- <- <- <- <- <- <- <-
v 1 1 1 2 2 2 2 2 2 2 2 2 2B 2 2 2 2 2 2 2 2 2 1 1 1 v
| 1 1 1 2 2 2 2 2 2 2 2 2 2B 2 2 2 2 2 2 2 2 2 1 1 1 |
v 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 v
| 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 |
v 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 v
| 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 |
v 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 v
| 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 |
v 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 v
| 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 |
v 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 v
-> 1B 1B 1B 3C3C 3C 3C 3C 3C 3C 3C 3C 3D 3C 3C 3C 3C 3C 3C 3C 3C 3C 1B 1B 1B<-
^ 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 ^
| 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 |
^ 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 ^
| 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 |
^ 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 ^
| 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 |
^ 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 ^
| 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 |
^ 1 1 1 3 3 3 3 3 3 3 3 3 3B 3 3 3 3 3 3 3 3 3 1 1 1 ^
| 1 1 1 2 2 2 2 2 2 2 2 2 2B 2 2 2 2 2 2 2 2 2 1 1 1 |
^ 1 1 1 2 2 2 2 2 2 2 2 2 2B 2 2 2 2 2 2 2 2 2 1 1 1 ^
-> -> -> -> -> -> -> -> -> -> -> -> ^ <- <- <- <- <- <- <- <- <- <- <- <-
uint8, int8, uint16, int16, uint32, int32, uint64, int64, float, double
Peter Cook 2019