From Avisynth wiki
Jump to: navigation, search
Author Kassandro
Version v0.5
Download RemoveGrainHD___(0.5_-_2011-08-11).7z
Category Denoisers
License GPLv2
Discussion Doom9 Thread



HD stands for High Definition, i.e. for high resolution input. For this kind of video RemoveGrain, which changes a pixel in terms of its eight direct neighbours, may be too narrow, because due to the high resolution noise often consists of pixel clusters rather than isolated pixels. Thus instead of RemoveGrain's 3x3 neighbourhood the filter RemoveGrainHD uses a rather arbitrary symmetric rectangle (preferably but not neccessarily a square) around a pixel to process it. Of course not all the many RemoveGrain modes can be extended to this setting. In fact, the current version of RemoveGrainHD only contains the 4 original modes, with which RemoveGrain was started once. Nevertheless a useful Repair variant, which serves the same purpose as the Repair plugin of the RemoveGrain package, is included as well. There is even a spatio-temporal version called TemporalRemoveGrainHD, which has no counterpart in RemoveGrain at all. While RemoveGrain uses a very high level of parallelism - the SSE2/SSE3 version processes 16 pixels simultaneously - this is unfortunately no more possible for RemoveGrainHD. It is a rather ordinary C/C++ program without any inline assembler code. Nevertheless performance is not that bad thanks to some ideas, which I have taken from Tonny Petersen's MedianBlur filter. However, there are some subtle changes, which should also improve MedianBlur. Thus I have re-implemented MedianBlur in this plugin. This filter is called Quantile, because it not only can calculate the median of an arbitrary surrounding rectangle but also arbitrary quantiles of these rectangles. The median then is just a special quantile. RemoveGrainHD can be emulated with an AviSynth script, which involves two instances of the filter Quantile and two mt_lutxy instances. However, such a script is substantially slower. In version 0.2 of this filter package we have added the filter SmartMedian, which is somehow in the middle between Quantile and RemoveGrainHD. While there is no repair version of this filter, there is also a temporal variant of this filter, called TemporalSmartMedian. In version 0.2 we have also have changed the names of various variables and the inheritance rules to stay compatible with the forthcoming application RemoveGrainImg for still pictures. In version 0.4 the we have added the filter SmartMedian2.




In order to understand RemoveGrainHD and SmartMedian it is quite instructive to discuss the simpler filter Quantile first. All the filters of this plugin treat the three planes, i.e. Y, U and V channel or R, G and B channel completely independent of each other. In the sequel we therefore discuss the operation of the filter on one plane only. A pixel then consists of a single value and not three values. In particular, we can compare pixels and can speak of smallest and largest pixels of certain collections of pixels, usually rectangular boxes.

The simple command Quantile(input) returns a clip, which for each pixel of the clip input calculates the median of the 5x5 box surrounding it, i.e. the 13th smallest pixel, which is also the 13th largest pixel, of this box, which has altogether 25 pixels. For pixels near the border the surrounding box is cut off and like MedianBlur, Quantile handles these pixels properly, although MedianBlur and Quantile yield different results for the pixels near the borders. This may be due to the fact that the median is not unique if the box consists of an even number of pixels, but I have not investigated this problem. The surrounding box of the pixel in one of the four corners consists only of 9 pixels and then the median is the 5th smallest pixel. If you want another radius of the surrounding box you should use the variable "radius". For instance Quantile(input, radius=4) uses 9x9 boxes. If you want to specify the box radius specifically for the Y channel (or the R channel if you use the color spaces RGB or RGB32), use the variable radius_y instead. Similarly you can use the variables radius_u, radius_v for U and V channel (or the G and B channel if an RGB color space is used). If used this variables overwrite the value of the variable "radius" for the specific channels. You may also use a different radius in the vertical direction. For this purpose we have the variables yradius_y, yradius_u, yradius_v. yradius. Again yradius is for all channels, while the other variables are channel specific. For instance, Qunatile(input, yradius_y=4) uses 5x9 boxes for the Y channels and the default 5x5 boxes for the other channels. Quantile(input, radius=6, yradius_u=4) uses 13x13 boxes for the Y and the V channel and 13x9 boxes for the U channel. If radius_y=0 or yradius_y=0 then the Y channel is simply copied. If even radius_y<0 or yradius_y<0 then the Y channel isn't processed at all and the Y channel of the output clip contains only unpredictable, arbitrary values. Similar rules are valid for the U and V channels if radius_u, yradius_u, radius_v, yradius_v have zero or negative values.

Until now Quantile behaves very similar to MedianBlur. However, as already mentioned, you may select other quantiles, not just the median, of the surrounding box. The variables rank, rank_y, rank_u and rank_v are for this purpose. Again the variable rank is for all channels, while the other 3 variables are channel specific. For instance, Quantile(input, rank=3) selects the 3rd smallest pixel of the surrounding 5x5 box for all three color channels. If the box size of the Y channel is not changed by the variables radius_y or yradius_y then rank_y inherits the value of the variable rank otherwise rank_y is set to (2*radius + 1)*(2*radius_y + 1) /2 + 1, i.e. the median of the box. Similar rules are applied for the other color channels. For instance, Quantile(input, rank=1) selects the minimum of the surrounding 5x5 box and Quantile(input, radius=5, rank=121) selects the maximum of the surrounding 11x11 box for all color channels. Quantile(input, radius_y=4, rank=20) selects the median for the surrounding 9x9 box in the Y channel and the 20th smallest pixel of the surrounding 5x5 box in the U and V channels. Quantile(input, radius=1) is the same as RemoveGrain(input, mode=4) except the border pixels, but RemoveGrain is certainly a lot faster.

The full syntax of Quantile is as follows:

Quantile(clip input, int "radius", int "rank", int "radius_y", yradius_y, rank_y, radius_u, yradius_u, rank_u, radius_v, yradius_v, rank_v, planar)

Here "input" is a clip variable and "planar" is a boolean variable. All the other variables are integer variables. The variable "planar" must be set to 'true', if the input clip is a planar YUY2 or a planar RGB clip. Thus the behaviour of this variable is the same as in RemoveGrain. Quantile like all other filters of the RemoveGrainHD and the RemoveGrain package cannot handle ordinary, non-planar YUY2 and RGB clips and the variable "planar" should prevent the user from accidentally applying these filters to non-planar YUY2 or RGB clips. The SSETools plugin (included in the RemoveGrain package) provides tools to convert non-planar YUY2 and RGB clips on to planar ones and reversely.


RemoveGrainHD(input) calculates the 5th smallest pixel min5 and the 5th largest pixel max5 of the 5x5 box around a pixel p of the input clip and then clips the value p with respect to min5 and max5, i.e. p is replaced by min(max5, max(min5,p)). Of course, instead of the 5th smallest and the 5th largest pixel of the box we may also take, say, the 7th smallest and the 7th largest pixel of the surrounding 5x5 box. The variable rank is for this purpose and the task is accomplished by RemoveGrainHD(input, rank=7). If specified the channel specific rank_y, rank_u and rank_v overwrite the value of the variable "rank" for the various color channels. For instance, RemoveGrainHD(input, rank=6, rank_y=4) uses the 4th smallest and the 4th largest value for clipping the Y channel and the 6th smallest and 6th largest value for clipping the other color channels. One may even select different lower and upper clipping values. The channel specific variables urank_y, urank_u and urank_v are for this purpose. For instance RemoveGrainHD(input, rank_y=6, urank_y=4) clips the center pixel in the Y channel with repect to the 6th smallest and the 4th largest value of the surrounding 5x5 box. Of course, RemoveGrainHD is not restricted to boxes with radius 2 - this is only the default value. As in the filter Quantile the variables radius, radius_y, yradius_y, radius_u, yradius_u, radius_v, yradius_v are for this purpose. They are used in the same way as in Quantile. The default value for the variable rank is 2*radius + 1, i.e. 5 for the default radius 2. Why? If there is a line or curve through the center of the 5x5 box, then it consists at least of 5 pixels. Thus the default value 5 assures that such a line or curve is not interrupted in the center by RemoveGrainHD. If the box size of the Y channel is not changed by the variables radius_y or yradius_y then as in the filter Quantile rank_y inherits the value of the variable rank otherwise rank_y is set to min(2*radius_y + 1, 2*yradius_y + 1) to insure the lines and curves through the center of the box are not broken by RemoveGrainHD. Similar rules are applied for the other color channels. From the above thoughts it is clear that square boxes are preferable. Let us discuss some further examples. RemoveGrainHD(input, rank=2, radius=3) clips the center pixel of the surrounding 7x7 box by the minimum and the maximum of these box. where the minimum and maximum is calculated without the center pixel. rank=1 wouldn't touch any pixel at all. Thus with rank=2 RemoveGrainHD extends Undot to arbitrary surrounding boxes. If only the size of the surrounding box but not the rank values are increased, then RemoveGrainHD becomes weaker and weaker, i.e. pixels are changed less and less. RemoveGrainHD(input, lrank_y=13) is the same as Quantile(input) except a few pixels near the border, but Quantile is faster. For the pixels near the border the surrounding box is cut off. Although both, Quantile and RemoveGrainHD, handle pixels near the border by reducing the ranks proportional to the size of the surrounding box, they do not behave identical in the above example.

If xradius_y <= 0 or yradius <= 0 then the Y channel of the input clip is simply copied to the output clip. Note that RemoveGrainHD behaves differently from Quantile if xradius_y < 0 or y_radius < 0. The only reason for this different behavior is the different implementation of both filters. Of course the U an V channel behave analogously.

The Repair Mode of RemoveGrainHD

RemoveGrainHD has one further variable, the clip variable "repair". RemoveGrainHD(input1, repair=input2) calculates for the 5x5 surrounding box of each pixel p1(n, x, y) (n is the frame number and x,y are the coordinates) the 5th smallest pixel min5(n, x, y) and the 5th largest pixel max5(n, x, y) and then clips the corresponding pixel p2(n, x, y) of the clip input2 with respect to these two values, i.e. p2(n, x, y) is replaced by min(max5(n, x, y), max(min5(n, x, y), p2(n, x, y)) in the output clip. The two clips input1 and input2 must have the same size, type etc. otherwise the filter terminates with an error message. Of course the other variables of RemoveGrainHD continue to play the same role as without setting the repair variable. In fact, RemoveGrainHD(input1, repair=input1, ...) is the same as RemoveGrainHD(input1, ...). For n=1,2,3,4 RemoveGrainHD(input1, repair=input2, xradius_y=1) is the same as Repair(input2, input1, mode=n). In this way RemoveGrainHD extends the first four modes of the Repair plugin of the RemoveGrain package. Again Repair is a lot faster than RemoveGrainHD.

The full syntax of RemoveGrainHD is as follows:

RemoveGrainHD(input1, repair, radius, rank, radius_y, yradius_y, rank_y, urank_y, radius_u, yradius_u, rank_u, urank_u, radius_v, yradius_v, rank_v u, rank_v, planar) 

Here "input1" and "repair" are clip variables and "planar" is a boolean variable. All the other variables are integer variables.


TemporalRemoveGrainHD uses the same variables as RemoveGrainHD, but for computing the clipping bounds not only the surrounding box of a pixel in the current frame, but also in the preceding and the succeeding frame is used. Thus three times as much pixels are used and the rank variables rank, rank_y, urank_y, etc. should be tripled to obtain a similar effect as the purely spatial RemoveGrainHD. Thus the default value of rank now is 15 and more generally if radius, radius_y, yradius_y etc. are used, then the default value for rank_y is 3 * max(2 * radius_y + 1, 2 * yradius_y + 1). This value is small enough such that a curve going through the center is not harmed but the filter if it does not move to much. Thus with increasing radius TemporalRemoveGrainHD becomes more and more robust against motion, i.e. motion does not cause artifacts and does not hurt denoising. This is a very nice property.

By default all three clips have equal weight. However, it is only natural to give the current frame, i.e. the frame in the middle, more weight. To this end we have the only additional variable "weight". If set to 2, then the pixels of the current frame are counted twice while the pxiels of the other two frames are still counted only once. Thus we have now four times as much pixels as in RemoveGrainHD and the default value for rank_y has to be adjusted to 4 * max(2 * radius_y + 1, 2 * yradius_y + 1). As the variable weight increases TemporalRemoveGrainHD becomes more and more spatial and behaves more and more like RemoveGrainHD with suitably adjusted rank values. Although this doesn't make much sense, one may set the variable "weight" to 0 and then only the previous and the subsequent frame are used in the computation. weighted = 1 is the default. The various rank variables in TemporalRemoveGrainHD should have about (2 + weight) times the values of the corresponding variables in the RemoveGrainHD filter to achieve a similar amount of cleaning in static areas.

For each frame TemporalRemoveGrainHD requires a preceding and a succeeding frame. Because one of these two frames does not exist for the first and the last frame of a clip, these two frames are not changed by TemporalRemoveGrainHD.

The full syntax of TemporalRemoveGrainHD is as follows:

TemporalRemoveGrainHD(input1, repair, radius, rank, radius_y, yradius_y, rank_y, urank_y, radius_u, yradius_u, rank_u, urank_u, radius_v, yradius_v, rank_v u, rank_v, weighted, planar) 

Here "input1" and "repair" are clip variables and "planar" is a boolean variable. All the other variables are integer variables.


The filters MedianBlur and Quantile are so poor - at least if the surrounding box is large - because all pixel of the surrounding box have equal influence on the end result and unlike RemoveGrainHD (with suitable rank values) the more distant pixels play a more and more dominant role in the calculation of the median as the box radius increases. SmartMedian tries to resolve this problem by computing the median of a subsample of the surrounding box. restricting the sample, over which the median is taken, to pixels with values sufficiently close to the value of the center pixel. The variables radius, radius_y, yradius_y, xradius_u, yradius_u, xradius_v, yradius_v are used by SmartMedian as well and play the same role as in the other filters of this plugin. However, instead of the various rank variables, the variables pixels, pixels_y, pixels_u, pixels_v are used. These variables describe the size of the subsample. The inheritance rules are the same as for the variables rank, rank_y, rank_u, rank_v in the filter Quantile. Obviously, the size of the subsample must be smaller than the size of the box. Thus pixel_y <= (2*radius_y + 1) * (2*yradius_y + 1), pixel_u <= (2*radius_u + 1) * (2*yradius_u + 1) and pixel_v <= (2*radius_v + 1) * (2*yradius_v + 1). Of course these variables must also be positive. If unreasonable values are specified fpr these variables, they are suitably adjusted. To describe the algorithm underlying the SmartMedian filter we restrict ourselves to the Y channel and let A(r) be the pixels from the surrounding box of the center pixel c with distance(p,c) = |p - c| <= r (p,c are the pixel values and not the coordinates). In the first step SmartMedian calculates the smallest R>=0 such that the size |A(R)| of the subsample A(R) exceeds the value of pixels_y, i.e. |A(R)|>= pixels_y. If the pixel c is from an area with unform color, then R=0. If c is a noisy pixel, then R will be quite large, but usually it is quite small, if the value of pixels_y is reasonable. It may very well happen, that |A(R)| > pixels_y. If pixel_y = (2*radius_y + 1) * (2*yradius_y + 1) (the maximal value), then the subsample A(R) is equal to the entire surrounding box. If pixels_y=1 (the minimal value) then R=0 because c ist always contained in A(0). In the second step SmartMedian calculates n-th smallest min(n,R) and the n-th largest pixel max(n,R) of the subsample A(R), where n = pixels_y/2 (with rounding to the upside). Finally c is clipped with respect to min(n,R) and max(n,R), i.e. c is replaced by min(max(n,R), max(min(n,R),c)). If |A(R)| = pixles_y then c is indeed replaced by the median of the subsample A(R), otherwise is only approximately the median. Obviously, if R=0 then min(n,R) = c = max(n,R) and c is left unchanged. The default value for pixels_y is 2 * min(2*radius_y + 1, 2*yradius_y + 1), i.e. twice the default value for rank_y in RemoveGrainHD. Again the default value assures that no thin lines or curves through the center c are interrupted by SmartMedian. Compared with RemoveGrainHD the value of pixels_y should be twice the value of rank_y to have a similar preservation of detail. If the default values for rank_y and pixels_y are chosen, then RemoveGrainHD leaves many pixels unchanged, while SmartMedian changes most pixels at least a little. All in all SmartMedian is more aggressive than RemoveGrainHD and should yield better compression. To illuminate the above description, let us look at some concrete cases. There Quantile, RemoveGrainHD and SmartMedian are used with the default values radius_y = yradius_y = 2, rank_y = urank_y = 5, pixels_y = 10 if not specified otherwise.

61 12 10 13 13
11 62 10 14 11
12 09 65 12 10
15 11 15 62 11
10 12 14 11 61

In the first example we essentially have a diagonal line (from northwest to southeast) through the center with a dark background with only small noise. The 5th smallest pixel of the box is 10 while the 5th largest pixel is 61. Thus RemoveGrainHD clips the center value 65 with respect to 10 and 61. Thus RemoveGrainHD assigns the value 61 to the center pixel. SmartMedian calculates the value R = 52 and the subsample A(52) looks follows:

61 13 13
62 14
15 15 62
14 61

In this case we have |A(52)| = 11 > pixels_y = 10 > 9 = |A (51)|. We have min(5,52) = 15 and max(5,52) = 61. Thus the value 65 of the center pixel is changed to 61 by SmartMedian, i.e. the same result as RemoveGrainHD. The median of the entire box is 12. Thus MedianBlur and Quantile would replace the center pixel with the very destructive value 12. Let us now damage the diagonal line a little bit:

61 12 10 13 13
11 62 10 14 11
12 09 12 65 10
15 11 15 62 11
10 12 14 11 61

We have simply exchanged the center with its neighbour on the right hand side. The 5th smallest and the 5th largest pixel are hence again 10 and 61. Thus the center value 12 is clipped with respect to 10 and 61 and is not changed by RemoveGrainHD. SmartMedian now calculates R = 1 and A(1) looks as follows:

12 13 13
11 11
12 12
11 11
12 11

In this case we have |A(1)| = 11 > pixels_y. The 5th smallest element of A(1) is 11 and 5th largest element is 12. Thus the center pixel is clipped between 11 and 12 and therefore remains unchanged. Neither RemoveGrainHD nor SmartMedian was able to repair the damaged line. If a denoiser would repair the line, it would destroy the dark line from southwest to northeast as well as many other dark curves through the center. Thus any denoiser which would repair the line, would very likely creat artifacts in other situations. Let us next discuss an example, where RemoveGrainHD and SmartMedian differ substantially:

61 12 97 99 98
61 62 10 14 98
62 69 73 12 10
65 61 65 62 11
60 62 64 61 62

In this case the 5th smallest pixel is 12 while the 5th largest pixel is 73. Thus clipping the center pixel 73 with respect to these two pixels doesn't change anything. On the other Hand SmartMedian calculates R = 11. A(11) looks as follows:

62 69 73
65 65 62
62 64 62

We have |A(11)| = 10 = pixels_y, min(5,10) = 62 and max(5,10)=64. Thus SmartMedian replaces the value 73 by the more reasonable value 64. The median of the entire box is 62. Thus in this case SmartMedian unlike RemoveGrainHD delivers similar results as MedianBlur and Quantile. Under rare circumstances SmartMedian may amplify noise:

11 12 90 95 93
11 13 10 94 91
12 09 55 12 90
15 11 15 14 11
13 15 14 11 11

Here we have essentially a sharp edge near the upper right corner with a noisy pixel in the center. SmartMedian calculates R=40 and A(40) looks as follows:

90 95 93
94 91
55 90
15 15

We have |A(40)|=10 and min(5,40) = 90 = max(5,40). Thus SmartMedian replaces the center pixel 55 with the even worse value 90. On the other hand RemoveGrainHD leaves the center pixel unchanged, because the 5th smallest element of the box is 11 and the 5th largest element is 90. The above situation may also happen in the case of very strong noise. In this case pixels_y has to be raised. Also the problem increases with the radius being absent if radius_y=1. However in practice the above noise amplification is rare and hardly visible. If we only make a tiny change it disappears:

11 12 90 95 93
11 13 10 94 91
12 09 55 12 90
14 11 15 14 11
13 15 14 11 11

Now we have R=41 and A(41) looks as follows:

90 95 93
94 91
55 90
14 15 14
15 14

Now we have |A(41)|=12, min(5,41)=15 and max(5,41)=90. Hence the center pixel now remains unchanged.
If pixels_y is raised somewhat beyond its default value, then SmartMedian becomes edge sharpening. To this end we now assume that pixels_y=11 and look at the following example:

72 93 92 95 93
31 73 92 94 91
12 29 71 92 93
14 11 30 74 91
13 15 14 31 73

In this case SmartMedian calculates R=22 and A(22) looks as follows:

72 93 92 93
73 92 91
71 92 93
74 91

Thus |A(22)|=13, min(6,22) = 91, max(6, 22) = 92, whence the center pixel is changed to 91. Note that we have pixels_y/2 = 11/ 2 = 6, because we always round ro the upside. With the default value pixels_y=10, we would have min(5,22) = 73, max(5, 22) = 92 and therefore no significant sharpening. RemoveGrainHD leaves the center pixel unchanged. The median of the entire box is 73. Thus if we raise pixels_y too much we SmartMedian approaches more and more the ordinary median and instead of sharpening we usually get some softening although the median never blurs (MedianBlur is really a misleading filter name). While SmartMedian never produces ugly sharpening halos, it may very well creat aliasing. However, with the default values for pixels_y there shouldn't be much sharpening and virtually no aliasing.

Choosing the right value for pixels_y is a rather subtle tightrope walk and requires a good understanding of the algorithm. For a sharp source only damaged by noise pixels_y should not exceed the default value. For a somewhat blurry source pixels_y should be raised beyond the default value - but not too much. A softer source also justifies a larger value for radius_y, although radius values > 2 are not very damaging for sharp input as well as long pixels_y is not too large.

The full syntax of SmartMedian is as follows:

SmartMedian(input, radius, pixels, radius_y, yradius_y, pixels_y, radius_u, yradius_u, pixels_u, radius_v, yradius_v, pixels_v, planar) 

Final remarks:

  • If pixels_y = 2 * rank_y = 2 * urank_y, then SmartMedian always changes the center pixel more than RemoveGrainHD.
  • SmartMedian is already very interesting if radius_y=1 (with the default value pixels_y=6). It would therefore be very desirable to have a much faster SSE implementation of this special case. This is possible, but due to a shortage of SSE registers the speed is expected to be about three slower than that of RemoveGrain(mode=4) (the median of the 3x3 box). Also the implementation is way more complicated than all the other RemoveGrain modes. At the moment I have simply not the time to do it. Having 16 SSE registers as in the AMD 64 bit mode would also be very advantageous for this purpose.


SmartMedian2 is similar to SmartMedian. However, instead of one input clip it has two. Of course, both clips must have same width, height, color space etc.. In SmartMedian2 there are now two frames at each time instance. Thus A(r) is now the set of all pixels with .distance(p,c) = |p - c| <= r . Here c is the value of the center pixel in the first clip and p are the pixel values (not the coordinates) of the surrounding boxes in both clips. Consequently SmartMedian2 behaves differently if the two clips are exchanged. Because A(r) depends heavily on the center value c, which is from the first clip, the first clip has a stronger influence on the output than the second clip. There is the further variable weight to give pixels from the first clip even more weight. For instance, if weight=2, then the pixels from the first clip are counted twice in A(r). The various pixels variables in SmartMedian2 should have about (1 + weight) times the values of the corresponding variables in the SmartMedian filter to achieve a similar amount of cleaning.

The full syntax of SmartMedian2 is as follows:

SmartMedian2(input1, input2, radius, pixels, radius_y, yradius_y, pixels_y, radius_u, yradius_u, pixels_u, radius_v, yradius_v, pixels_v, weight, planar)

Here input1 and input2 are clip variables, and the other variables are as in the SmartMedian filter.

SmartMedian2(input, input, pixels = 2*p, radius = r) is the same as SmartMedian(input, pixels=p, radius=r). However SmartMedian2 is slower. Firstly SmartMedian2 can be used to melt together two recordings of the same source. Secondly, SmartMedian2 can be used together with the Delay filter from the AvsTimer plugin to built recursive filters.


RankRepair attempts to restore the rank of a pixel within a surrounding rectangular box, which has been skewed by another filter, usually a sharpening filter. To this end let us assume that we have two video frames, the original frame and the filtered frame. Let c the value of a pixel in the original frame and c' be the value of the same pixel in the filtered frame. We denote by rank(c,r) the rank of c within a surrounding rectangular box of radius r. For instance, if rank(c,r)=5, the c is the 5th smallest pixel of the surrounding box. Similarly we denote rank(c',r). Unlike a denoiser a good sharpener should not change rank significantly, i.e. we should have rank(c,r) = rank(c',r). On the hand, if the sharpener creates a sharpening halo, the ranks of this halo pixels change substantially. RankRepair tries to repair such artifacts. The simplest idea is to replace c' by the value d in the surrounding box of c', which has the same rank as c. Thus if rank(c,r)=5, then d is the 5th smallest value of the surrounding box of c'. However, this choice may easily create new artifacts or move artifacts created by the previous filter to another place. To prohibit this behavior we clip the value d with respect c and c'. Thus, if d is outside the interval spanned by c and c', d is replaced by the nearest end point. In this way we assure that the repaired filtered frame is always closer to the original frame than the filtered frame. Unfortunately matters are not as easy as described above. The main complication arises from the fact that the rank of a pixel is not unique. Especially in a large surrounding box there may be other pixels which have the same value as the center pixel. In the following example

11 12 90 95 93
11 13 10 94 91
12 09 11 12 90
15 11 15 14 11
13 15 14 11 11

the center pixel share its value 11 with 6 other pixels. Thus the center pixel is simultaneously the 3rd, the 4th, the 5th, the 6th, the 7th, 8th and 9th smallest pixel. RankRepair resolves this problem in the most conservative way, i.e. minimizing change of the filtered frame. For more details I refer to the source code.

RankRepair is also a discontinuous filter, i.e. a slight change of input values may cause a big change of output values. However, RankRepair is not as discontinuous as many other AviSynth filters, which use thresholds to decide whether a pixel should be processed or not.

The full syntax of RankRepair is as follows:

RankRepair(filtered, original, radius, radius_y, yradius_y, radius_u, yradius_u, radius_v, yradius_v, planar) 

Here "filtered" is the clip variable for the filtered clip and "original" is the clip variable for the original, unfiltered clip. The other variables are the same as in the other filters and describe the size of the surrounding box in the various color channels. The following example is rather typical application of RankRepair:

function ControlledSharpen(clip input, float amount)

    sharpened=sharpen(input, amount)
    return RankRepair(sharpened, input, radius=2)


Version      Date            Changes
v0.5 2011/08/11 - Fix Quantile bug v0.5 2009/01/27 - RankRepair added and documentation update. v0.4 2007/12/18 - SmartMedian2 added and documentation update. v0.3 2007/11/26 - Completely new implementation (hopefully faster) of the RemoveGrainHD and TemporalRemoveGrainHD and a more flexible design with include skeletons of the entire plugin. v0.2 2007/xx/xx - (no public release): SmartMedian and TemporalSmartMedian added, initialization bug of the TemporalRemoveGrainHD filter fixed. v0.1 2007/08/31 - Initial release

Archived Downloads

Version Download Mirror
v0.5 RemoveGrainHD___(0.5_-_2011-08-11).7z RemoveGrainHD___(0.5_-_2011-08-11).7z

External Links

Back to External Filters

Personal tools