Version 9 (modified by Sam Hocevar, 16 years ago) (diff)

O(1) median filters

<< back to libpipi

Gaussian filtering

The usual way to create a Gaussian kernel is to evaluate a Gaussian function at the center of each cell:

k[i][j] = exp(-(i²+j²)/2σ)

This usually works well, except when the kernel is thin (σ < 1). It gets worse when using our generalised kernel:

k[i][j] = exp(-((i×cosθ-j×sinθ-dx)²+(j×cosθ+i×sinθ-dy)²)/2σ)

TODO interpolate the Gaussian integral on more points, eg. 5×5 (currently libpipi does it at the central point and at 8 additional points in the neighbourhood).

Code

Median filtering

There are several ways to optimise a median filter.

The first idea is an approximation, based upon the assumption that the median of subset medians is close to the median of the whole set. It can already be simulated by applying a 5×0 filter and a 0×5 filter instead of a 5×5 filter, so there is no real point in studying this specific optimisation.

The second idea is to optimise the median selection. The most naive method is to bubble-sort the neighbourhood values and select the middle point, which has complexity O(n²). Improving on the sort algorithm by using eg. heapsort or quicksort, reduces the complexity to O(n.log2(n)). But since we’re only interested in the median and not in ordering the rest of the data, we can use an efficient median selection algorithm. This operation can be done in fewer than 3n tests [1] but the practical implementation is extremely complex.

The third way to optimise a median filter is by accounting for the size of neighbourhood changes in raster scan order: usually only a few values are removed from and added to the list of neighbours. This becomes promising with large kernel sizes, and efficient algorithms exist that handle this case in 1D at least [2].

Methods for O(1) median filters have recently been discovered [3]. However, they rely on histograms, which only works well when the input data consists of 8-bit integers. Since libpipi does most of its computations using 32-bit floats, such techniques are of limited usefulness.

References

Code

Dilate / erode

Erosion and dilation already work, but only for nearest pixels.

TODO each pixel couple test is done twice. There seems to be an opportunity for improvement there.

TODO implement "dilate by X" (Manhattan distance).

TODO implement "dilate by X" (Euclidian distance).

Code

Links