RemoveDirt

Abstract

author: Rainer Wittmann (aka kassandro)
version: 0.6.1
download:
category: Temporal Smoothers
requirements:

Overview

RemoveDirt is a so called temporal cleaner for Avisynth 2.5x. If you don't want to go through the documentation, use it without setting any variables. For an average movie this should take out about half the dirt without creating visible artifacts. Though the default setting is quite conservative, very fast motion bears a moderate artifact risk. However, because postprocessing should remove all non-smooth artifacts, these rare artifacts are usually smooth and very short in time, whence hard to catch for the human eye. Clips are very different and tuning the variables, which control the behavior of RemoveDirt, can make a big difference. Two different debugging tools can be used for this purpose. The variables themself can be explained best by discussing the underlying algorithm. If you don't understand a variable, don't change its default value and post a question in the RemoveDirt thread of the doom9 forum rather than sending me email.

No automatic cleaner can catch all dirt without destroying the movie by artifacts. The amount of cleaning depends heavily on the amount of motion. To avoid artifacts cleaning must be suppressed in the moving areas of a clip. A moving camera is clearly the worst case. The first and the last frame of a clip are never cleaned and there will hardly be any cleaning at a sharp scene switch, because RemoveDirt requires that the previous and the subsequent frame are existent and fit together well. Because RemoveDirt- unlike other cleaners - tries to remove dirt but never blurs dirt, it is adviceable to use it together with a slight softener for the remaining dirt, in order to obtain better compression (such a filter should be after RemoveDirt). If you use divx 5.1.1 the builtin preprocessing filter is a good choice, but based on my personal tests I would recommend only light preprocessing. In its default configuration RemoveDirtcan hardly remove any big pieces of dirt, but it usually removes a lot of small dirt and dust and after being processed by RemoveDirtthe clip should look significantly cleaner.

RemoveDirt supports the two most important color spaces YV12 and YUY2. Support for other YUV color spaces can be added, but the design makes it impossible to support RGB color spaces (convert to YUY2 with Avisynth's internal filter ConvertToYUY2 if you stumble over such color spaces). The width and the height of the clip must be a multiple of 8.

As can be seen from the source code and the descriptions below, RemoveDirtis quite sophisticated. To achieve reasonable performance it was necessary to program all low level routines in assembly language using extensively Intel's integer SSE instructions. Thus a modern CPU with integer SSE capability is necessary for using this plugin.

Installation

Beginning with version 0.5 the binary package of RemoveDirt contains two versions of the plugin. RemoveDirt.dll (dynamically linked, hence small) and RemoveDirtS.dll (staticly linked, hence big). Try first the small dll and copy it to Avisynth's plugin directory. If it doesn't work, this is probably due to missing msvcr70.dll library. Either you install this library in C:\windows\system32 or you delete RemoveDirt.dll and replace it by RemoveDirtS.dll. Please do not put both dlls into the plugin directory. If you fill the plugin directory with all kinds of superfluous dlls, then you only slow down the start of any application which uses Avisynth. RemoveDirt is safe with respect to multithreading and may be used arbitrarily often within a single Avisynth script. There should be no conflict with other filters.

Dirt Detection

1. A dirty spot or scratch occurs only on one frame and not on a whole sequence of frames. This basic property is exploited by any temporal cleaner. Thus scratches caused by bad film projectors at the same position on a series of frames cannot be corrected by a temporal cleaner. In our program design the basic property is made precise as follows: a pixel is potentially dirty if cf[x,y] is not between min(pf[x,y],nf[x,y]) and max(pf[x,y],nf[x,y]), where cf[x,y] denotes the grey value of the pixel with coordinates x and y on the current frame and similarily pf[x,y] and nf[x,y] denote the grey value of the same pixel on the previous and the subsequent frame. In other words a potentially dirty pixel is out of the range spanned by the previous and the subsequent frame. We use the word potentially, because, nnfortunately, not only dirt but also some kind of motion are causing a pixel to get out of range. If one does not distinguish between both, dirt removal will ruin rather than clean a clip. This unpleasant fact makes automatic true dirt removal with low artifact risk so challenging. The extreme example is a moving thin black line with a white background. Because of motion any black pixel is white on the previous and the subsequent frame, whence in this example all black pixel are always out of range. On the other hand, if the width of the moving object is large in the direction of motion (this is the reason why the line has to be thin in the above example) and if the pixels of the moving object have similar color, then the pixels of the moving object are not significantly out of range. Hence from the point of view of cleaning there is good and bad motion. Cleaning bad motion causes artifacts, good motion does not.

2. If a cleaned block has one edge in common with a block, which has not been cleaned because of some reason, then cleaning should not increase the difference of the two adjacent border lines of both blocks by too much, otherwise the cleaning has to be considered as incorrect and must be undone. In other words cleaned blocks should fit reasonably well to blocks, which are not cleaned. This is a postprocessing technique. As technical as it may look, it is an extremely powerful idea to avoid ugly artifacts. Frankly, I think that automatic true dirt removal with low artifact risk is hardly possible without this idea. For postprocessing RemoveDirtuses the variable pthreshold and cthreshold. If the total luma difference of the two adjacent border line increases by more than pthreshold, then cleaning of the block is undone by RemoveDirt. Similarily cleaning is undone, if the chroma difference exceedes cthreshold. The maximal difference of 8 pixels is 8*255. If cthreshold is larger than this value, then chroma postprocessing is disabled. For YUY2 chroma postprocessing is always disabled because it is not yet implemented. If both, pthreshold and cthreshold, are larger than the maximal value, then postprocessing is completely disabled. pthreshold = 20 is the default value and if nothing else is specified cthreshold has the same value as pthreshold. Clearly luma postprocessing is much more important than chroma postprocessing. If cthreshold and especially pthreshold, then rather unpleasant blocky artifacts become visible. These are much more likely in areas with very flat contrast. Of course, there must be motion as well to get such artifacts. If you see these typical blocky artifacts, you should lower the thresholds. Postprocessing should only be disabled if all the cleaned frames are checked for artifacts. By the very nature of the algorithm no postprocessing will occur if all blocks were cleaned. There must be at least one block, which has not been cleaned to trigger postprocessing. The postprocessing algorithm loops through all the blocks as long as it can find blocks to be restored, nevertheless it is quite efficient. It is the basic philosophy of RemoveDirt, that motion detection needs only detect at least one but not all blocks of a moving object. The rest is then taken care by postprocessing.

Motion Detection

Any procedure which allows to distinguish between dirt and motion can be considered as some kind of motion detection. From this point of view the previous section already contains two such procedures. Both are dirt specific and can be considered as nonconventional. In this section we consider general dirt independent motion detection. The basic idea to decide whether a block is a motion block or not, is to measure the difference of this block with its counterpart on the previous or the subsequent frame. However, for the purpose of cleaning this is not appropriate, because if the difference with its counterpart in a neighbour frame is low, then it cannot contain much dirt either. Instead the principal idea is to measure the difference of the blocks in the previous and the subsequent frame but not the current frame. Thus for the basic motion detection, the current frame (i.e. the frame to be cleaned) is not used at all, only the previous and the subsequent frame are used for this purpose. Within RemoveDirtthe behavior of basic motion detection is controlled primarily by the mthreshold variable. Though a block contains 64 pixels the difference function used by RemoveDirtreturns only values between 0 (identical blocks) and 8*255 (difference between an absolutely black and an absolutely white block). If this difference exceeds the value of mthreshold, then the block is marked as a motion block. mthreshold=150 is the default value. The higher the value of mthreshold the more cleaning (but also with a higher artifact risk). In addition to the absolute threshold represented by the variable mthreshold, RemoveDirtcan use also adaptive thresholds which vary from block to block (actually RemoveDirtuses a block specific threshold only if it is smaller than mthreshold). We do not describe the somewhat complicated algorithm here. The adaptive thresholds are controlled by the athreshold variable. It may very well attain negative values, but if athreshold < -mthreshold/2, then there will be no cleaning at all. athreshold=50 is the default value. The higher the athreshold variable the more blocks are cleaned and the higher the artifact risk. If mthreshold and athreshold are both > 8*255, then all blocks are cleaned. In the presence of bad motion this choice will creat massive artifacts and should only be used for selected frames using range files as decribed below. There are two further variables dist and tolerance, which control basic motion detection. dist=1 and tolerance=12 are the default values. A block B is considered a neighbor of a block A by RemoveDirt, if both horizontally and verticaly both blocks are only dist blocks apart. For dist=0 a block has only one neighbor, the block itself. If dist=1, then a block has 9 neighbors, if it is not located at the boundary. If dist=2, then each inner block has 25 neighbors and if dist=3 then it has 49 neighbors etc.. Now for a given block RemoveDirtcounts all the neighbor blocks which are marked as motion blocks. If the percentage of motion blocks among all neighbor blocks exceeds the value of tolerance, then the block is not cleaned. Thus in the default case of 9 neighbor blocks and tolerance=12 one motion block is allowed and cleaning will nevertheless be allowed. In particular, a motion block is cleaned, if it has no other motion blocks as neighbors. This is reasonable, because motion rarely occurs on one tiny block alone. On the other, if motion blocks have a certain density then also the neighbors should not be cleaned. This is the idea behind the variables dist and tolerance. A higher value of dist results in less cleaning. The higher the value of tolerance, the more cleaning. If tolerance >= 100 then all blocks are cleaned. Using the default values RemoveDirtwill be able to distinguish a moving thin black line from dirt, if it doesn't move too fast. Motion detection only needs to catch parts of a moving area. Then the rest is usually taken care by postprocessing. The disadvantage of the basic motion detection is that it doesn't distinguish between good and bad motion. In fact, good motion is easier to catch than bad motion. I hope that one fine day I can do better in this respect.

The above method of motion detection has a serious drawback, though. If a block on a frame has a substantial amount of dirt, then the same block on the previous and on the subsequent frame will falsely be detected as a motion block. Consequently, if a lot of small dirt is scattered all over a sequence of frames, cleaning performance of RemoveDirt may degrade. The idea to fix this problem is, to take the two neighbor frames, but not the current frame, from an already cleaned version of the clip. An Avisynth script with this kind of double processing will look like the following

input = MPEG2Source("input.d2v")
firstpass = RemoveDirt(input)
cleaned = RemoveDirt(input, neighbour=firstpass)

Here neighbour is a variable, which takes clips as values (the default value is the clip specified in the first argument). This technique can be iterated further, but of course the script executes slower and slower with each iteration. Normally the number of cleaned blocks increases by far less than 1% using iteration and this may not be enough reward for doubling the execution time. All kind of debugging should be disabled for the first pass if a debugging chaos is not desired. All the other variables should probably have identical values, but there is room for many experiments. Of course, one can do a lot of nonsense with the neighbour variable. RemoveDirtonly rejects a neighbour clip, if its format is so different from the format of the primary clip, such that it would compromise stable execution of the plugin.

Cleaning Modes

Currently three cleaning modes are implemented in RemoveDirt. They are accessed through the mode variable. mode = 0 is a very simple averaging mode. If a block is marked as cleanable, then the pixels of this block are replaced by the average of the corresponding pixels in the previous and the subsequent frame. If postprocessing is disabled, it very easily generates artifacts. The default mode = 2 is much smarter but also a little slower. If cf[x,y] is the grey value of a pixel in a cleanable block and if pf[x,y], nf[x,y] are the corresponding grey values in the previous and the subsequent frames, then cf[x,y]is left unchanged if min(pf[x,y],nf[x,y]) <= cf[x,y] <= max(pf[x,y],nf[x,y]). If cf[x,y] < min(pf[x,y],nf[x,y]) then cf[x,y] is replaced by min(pf[x,y],nf[x,y]) . Finally, if cf[x,y] > max(pf[x,y],nf[x,y]), then cf[x,y] is replaced by max(pf[x,y],nf[x,y]). The chroma values are handled in the same way. This is the method I am using all time and probably yields the best results. Due to SSE it is not much slower than the simple averaging method. Finally there is the experimental mode = 1, which is somewhere in the middle between mode 0 and 2 but closer to mode 2. Mode 1 is the most complicated and slowest mode. Initially I had hoped to obtain the better compression with mode = 1, but a first test was discouraging. Further tests will show, whether this mode should be dropped or not.

Higher Speed for Black&White Clips

If you have a movie which requires cleaning it is likely to be old whence often black and white. Of course the chroma of b&w clips doesn't need to be processed and cpu time can be saved. To accomplish this, RemoveDirtshould be used with the variable grey = true (the default is of course grey = false). Then instead of simply copying the old chroma to the new one, the new chroma is simply assigned the constant value 128. This is faster than copying and, as a byproduct, erases all chroma noise. In other words, if there was any color before, it is erased (Avisynth's builtin filter greyscale has exactly the same effect). grey = true also disables chroma post processing, which doesn't make any sense for b&w clips. Because there cannot be any chroma postprocessing, pthreshold should be somewhat lower, say 30 or 40.

Debugging

The boolean variable debug and the integer variable show are used for debugging. If RemoveDirtis used with show = 1, then all blocks, which are cleaned are colored red. If RemoveDirtis used with show = 2, then all blocks, which are not cleaned, are colored red. The default value show = 0, of course, implies that no coloring is done. Of course the show mode can only be used for previewing and is useful to see quickly which areas are caught by RemoveDirtand which it cannot clean. Using VirtualDubMod's F5 key, one can nicely see how changes of variables effect cleaning. The show mode only works for YUY2 clips. However, this is an inconvenience rather than a restriction. One has simply to put Avisynth's internal filter ConvertToYUY2 in front of RemoveDirt. Of course this should only be done for previewing. If the script is compressed, ConvertToYUY2 should be removed to avoid unnecessary color space conversions.

If debug = true then RemoveDirtsends output of the following kind to the debugview utility:

[1120] [36536] RemoveDirt: 5779 blocks cleaned (93%), 229 motion blocks (3%), 0 blocks restored, 1 loops
[1120] [36537] RemoveDirt: 5745 blocks cleaned (93%), 246 motion blocks (3%), 3 blocks restored, 2 loops
[1120] [36538] RemoveDirt: 5348 blocks cleaned (86%), 378 motion blocks (6%), 17 blocks restored, 5 loops
[1120] [36539] RemoveDirt: 4772 blocks cleaned (77%), 778 motion blocks (12%), 0 blocks restored, 1 loops
[1120] [36540] RemoveDirt: 4396 blocks cleaned (71%), 1028 motion blocks (16%), 1 blocks restored, 2 loops
[1120] [36541] RemoveDirt: 4695 blocks cleaned (76%), 820 motion blocks (13%), 3 blocks restored, 2 loops
[1120] [36542] RemoveDirt: 5362 blocks cleaned (86%), 384 motion blocks (6%), 48 blocks restored, 3 loops
[1120] [36543] RemoveDirt: 4038 blocks cleaned (65%), 1275 motion blocks (20%), 6 blocks restored, 2 loops
[1120] [36544] RemoveDirt: 3780 blocks cleaned (61%), 1522 motion blocks (24%), 3 blocks restored, 2 loops
[1120] [36545] RemoveDirt: 3943 blocks cleaned (63%), 1359 motion blocks (22%), 6 blocks restored, 2 loops
[1120] [36546] RemoveDirt: 4104 blocks cleaned (66%), 1225 motion blocks (19%), 3 blocks restored, 2 loops
[1120] [36547] RemoveDirt: 4287 blocks cleaned (69%), 1099 motion blocks (17%), 1 blocks restored, 2 loops
[1120] [36548] RemoveDirt: 4041 blocks cleaned (65%), 1261 motion blocks (20%), 0 blocks restored, 1 loops
[1120] [36549] RemoveDirt: 3901 blocks cleaned (63%), 1409 motion blocks (22%), 2 blocks restored, 3 loops
[1120] [36550] RemoveDirt: 3757 blocks cleaned (60%), 1507 motion blocks (24%), 4 blocks restored, 2 loops
[1120] [36551] RemoveDirt: 3799 blocks cleaned (61%), 1502 motion blocks (24%), 6 blocks restored, 2 loops

The first number on the left hand side is the id of the process, which runs the script, the second is the frame number, the blocks cleaned number explains itself, the motion block number is the number of motion blocks, which were found by the various motion detection routines, blocks restored is the number which were restored by the postprocessing routine and the loops number is the number of loops which were needed by the postprocessing routine to complete its task. If dist = 0 and tolerance < 100, then the relation

    blocks cleaned + motion blocks + restored blocks = total number of blocks = (frame width / 8) * (frame height / 8)

holds. If a block is cleaned, it doesn't imply that there was dirt. It would be too time consuming exclude clean blocks from cleaning and using all the artifact protection discussed before, cleaning a clean block should not make a visable difference and should improve compression slightly. The blocks restored number is of particular importance. The average blocks restored number should be 1-2% of all blocks. If it is higher, then the setting is probably too aggressive, if it is lower a more aggressive setting can be used. If a sharp scene switch occurs, then there are always two frames with very low blocks cleaned numbers. the first frame is the last frame of the old scene and the second frame is the first frame of the new scene. This is a rather characteristic pattern. In general the numbers are quite motion dependent. Here are some extraordinary frames, which I observed recently, while processing a rather dirty old b&w movie:

[1008] [2652] RemoveDirt: 5495 blocks cleaned (89%), 13 motion blocks (0%), 587 blocks restored, 32 loops
[1008] [22269] RemoveDirt: 4196 blocks cleaned (68%), 729 motion blocks (11%), 445 blocks restored, 20 loops
[1008] [24016] RemoveDirt: 1783 blocks cleaned (29%), 1957 motion blocks (32%), 1017 blocks restored, 16 loops

Frame 2652 is particularily remarkable: though there were only 13 motion blocks, postprocessing did catch a whopping 587 blocks with a whopping 32 loops. I had never seen such numbers before and I thought that they are only theoretically possible. Though it costs some time, I still run RemoveDirtalways with debug = true and if I see exceptional numbers as above, then I always make a visual inspection of such exceptional frames. In the above case, all frames were without visual artifacts.

Range Files 

Unfortunately the amount of dirt per frame usually varies throughout a movie. For instance, the first minutes of a film are often particular dirty, because these first minutes are located on the outside of the film roll. Also many particularly bad single frames are scattered all over a film. These situations cannot be handled optimally by a single variable setting. For this purpose range files were implemented in RemoveDirt. One may use up to 9 range files, which are specified with the range1, range2, ... , range9 variable. Each range file has its own set of variables, which have the same name as the global variables but with the range number attached at the end. For instance, if RemoveDirtis used with 

    RemoveDirt(range3= "myrange", mthreshold3=200, athreshold3=300, tolerance3=0, dist3=2)

then RemoveDirtuses mthreshold = 200, athreshold = 300 etc. for all the frames specified in the range file "myrange", which is expected to be located in the directory of the Avisynth script, but also different locations can be used by specifying the full path. For all the other frames RemoveDirt just uses the default values, because none of the global variables has been changed. A range file like "myrange" contains either single frame numbers like 13054 or ranges of frames like 13756-64, which is a shortcut for 13756-13764 (these numbers can be found by previewing the clip VirtualDubMod). The frame numbers in a range file must be increasing. This agreement makes certain shortcuts possible, which can be seen from the subsequent example. The various ranges must also be separated by a white space. Here we have a typical example:

0-43 67 287 9
1211-39 387 1432-544
11780-2833

The above range file covers the followings frames: 0-43, 67, 287, 289, 1211-1239, 1387, 1432-1544, 11780-12833. If a frame is specified in more than one range file, then the range file with the highest range number has precedence. For instance, if a frame is specified in range2 and range4, then the variables for range4 are used. In the future, I will use values a bit more agressive than the default values for the global variables. Then I will have a range file with significantly more aggressive values for the first minutes and perhaps some other scenes, if there is not too much motion and finally I will have a range file with variables set for total cleaning, which will be used for single extremely dirty or damaged frames. Though I already use RemoveDirt for real work (over time I have filled an entire 120 GB exchangeable hard disk with dirty old b&w movies of DVB origin and this hard disk is now gradually worked down), testing is still an important issue and this is the only reason, why I am not yet using range files.

Optimal Usage

1. Never crop after RemoveDirt. Modern codecs divide the frames in the same way as RemoveDirtinto a grid of 8x8 pixel blocks (codec experts, could you please confirm this?) to perform the crucial discrete cosine transform for such blocks. Now if the clip is cropped after RemoveDirt, then the grid of RemoveDirtand the codec are likely to be different resulting in subpar compression. There is one exception, though: cropping afterwards does not hurt, if all four sides are cropped by a multiple of 8. For instance, Crop(8,64,0,-72) is ok.

2. Crop only with "align=true". RemoveDirtheavily uses SSE instructions. If you crop without "align=true" before RemoveDirt, then the data on the frames may not be properly aligned and RemoveDirt will execute substantially slower. As a consequence you should always crop with Avisynth and not with DVD2AVI.

3. Telecined movies must be inverse telecined before RemoveDirt. If a film is telecined some fields are doubled in order to increase the frame rate from 24fps to 30fps. Hence on such doubled fields the basic property of dirt, described above, is no more valid and no temporal cleaner can ever spot dirt on such doubled fields. On the other hand, after an inverse telecine usually every fourth frame is composed of fields, which originate from two different frames. Visually these two fields fit together well but both are from a different compression context, which can mislead RemoveDirtto false motion detection. In extreme cases, one field may be from an I- or P-frame, while the other is from a B-frame. But even if the fields are from from frames of identical type, the different compression context has a substantial effect. Consequently RemoveDirtperformes less well on inverse telecined movies than on natively progressive movies. By the same reason also compression of inverse telcined movies is worse than of natively progressive movies. We in Europe should thank god every day that we are not getting telecined. However, here in Germany we have digital tv broadcasters, which like to comb progressive films (about 5% of all progressive movies from ARD and especially ZDF are combed). Fortunately these idiots are not able to double fields, so RemoveDirt should work, but on combed films the dirt is always split over two frames which clearly hurts RemoveDirt. On the other hand, if these combed films are uncombed, then we have the compression context problem for any frame and not only for any fourth frame. This is also the reason why Trbarry's uncomb filter doesn't work well in practice. The same question arises for interlaced movies. Should we deinterlace before or after RemoveDirt? On the other hand, dirt and scratches usually occur only on photographic, hence progressive, film. Thus this question shouldn't matter much. From the above discussion it should be clear that for measuring the quality of RemoveDirt only progressive clips should be used, which have never been telecined, and which have never been messed up by stupid digital tv providers. My plugin AlignFields can be used to decide with near absolute certainty, whether a clip fullfills these quality constraints or not.

4. Put other filters after RemoveDirt. Except those filters mentioned before, like crop and inverse telecine, all other filters should be put after RemoveDirtin the Avisynth script, because most filters have a negative rather than a positive impact on dirt detection.

All Variables at a Glance

NameTypeDefault ValueRemarksSection
neighbourClipprimary clip Motion Detection
distInteger1>=0Motion Detection
mthresholdInteger150>=0Motion Detection
athresholdInteger50 Motion Detection
pthresholdInteger20>=0Dirt Detection
cthresholdIntegerpthreshold>=0, only for YV12 clipsDirt Detection
toleranceInteger120<=tolerance<=100Motion Detection
modeInteger2only 0,1,2Cleaning Modes
greyBooleanfalse Black&White
debugBooleanfalse Debugging
showInteger0only 0,1,2, only for YUY2 clipsDebugging

All these variables can be specified for range files by attaching the number of the range file at the end of the variable. For instance, mode5 is the mode variable for range5.

RemoveDirt.ini

For a filter with many variables like RemoveDirt(together with all the variables for range files, RemoveDirt has more than 100 variables), where the settings depend strongly on the input clip, it desirable to have more than one default setting. Rather than hard wiring these default settings into the binary, these default settings are placed into the file RemoveDirt.ini. Now, if we have RemoveDirt(default="anime") in the script, then RemoveDirt looks for the file RemoveDirt.ini in the directory of the script. If it cannot find this file it terminates with an error message. Of course, without the default option, RemoveDirt.ini is not needed. Then RemoveDirtlooks for the string anime. If it cannot find this string, it terminates with an error message. Finally it reads all the variables after anime and replaces the internal default values by the values in RemoveDirt.ini. The format is

    name of variable = value of variable

If the variable is a string variable, then the value has to be enclosed in quotes. The string must not contain the quote character ". If the comment character # appears in the script (outside a string value, of course), then everything after # until the end of the line is considered as a comment. Let us look at the following concrete RemoveDirt.ini example:

clensing #total clensing, only for selected frames specified in clensing.rmd
range1 = "clensing.rmd"
pthreshold1 = 5000
mthreshold1 = 5000
grey = false
bw # black & white mode
grey = true
pthreshold = 20
mthreshold =180
anime # mode for anime videos
pthreshold = 40
cthreshold = 30
mthreshold = 200 

In this example RemoveDirt.ini contains three default modes: clensing, bw, anime. The default mode clensing defines a range1 file clensing.rmd (it must exist if this default mode is used) and the variables pthreshold1 and mthreshold1 are chosen such that for all frames specified in clensing.rmd all pixels are cleaned. It also uses grey = false, pthreshold = 20, mthreshold = 180 and cthreshold = 30 for the other frames, because these variables are specified after clensing. For all the the other variables the internal default values are taken. In the case RemoveDirt(default="bw") grey = true, pthresholdd = 20, mthreshold =180 and cthreshold = 30 is used (of course, cthreshold is ignored because grey = true). In the case RemoveDirt(default= "anime") the values pthreshold = 40, cthreshold = 30 and mthreshold = 200 are used. If there are syntax errors in RemoveDirt.ini, then RemoveDirtterminates with an error message, which displays the line number with the first syntax error.

ImproveSceneSwitch

By the very nature of the underlying algorithm RemoveDirtcannot clean at sharp scene switches. Only very few random blocks are cleaned. On the other hand the first frame after a scene switch should be encoded as an I-frame, and if this I-frame is particularily dirty, compression will suffer. Similarily the last frame before a scene switch should be encoded as a P-frame and dirt will be negative for compression as well. Most other temporal filters have problems at sharp scene switches as well. ImproveSceneSwitchis a simple filter to improve this situation. The difference d(n) of frame n and n+1 plays a fundamental role for this filter. It is based on the SSE intruction psadbw. If YV12 is the color space of the clip, then only the luma is used for calculating the difference, otherwise luma and chroma are used. The reason for this difference is computational convenience. If field = 0 (this is the default) then the entire frame is used for calculating d(n). If field = 1, then only the top field is used for calculating the difference. If field = 2, then only the bottom field is used for calculating the difference. This variable is important for deinterlacers like AlignFields. We determine a (sharp) scene switch between frame n and n+1 if d(n) > ratio*d(n-1) and d(n)>ratio*d(n+1). Here ratio is an integer variable > 1 (ratio = 7 is the default). This variable determines the scene switch sensitivity of ImproveSceneSwitch. It should not be larger than 100 to avoid arithmetic overflow. Now, if we have a sequence n-1,,n,n+1,n+2 of frames with a scene switch between n and n+1, then ImproveSceneSwitchreplaces the sequence n-1,n,n+1 by n-1,n-1,n+2,n+2. Of course this kind of frame doubling should only occur at sharp scene switches and then the viewer should hardly notice the trick. ImproveSceneSwitchis used as follows

    ImproveSceneSwitch(clip, orig, ratio = integer, extrapolate = boolean, first = boolean, last = boolean, field=0,1,2, debug = boolean

Here clip is the output and orig is the input of the temporal filter, which should be improved by ImproveSceneSwitch. orig is only used for determining scene switches. Flashes and similar effects (very common in music videos) may mislead ImproveSceneSwitch. For such clips either ImproveSceneSwitchshould not be used at all or a very high ratio above 10 should be chosen. By default both, the last frame of the old scene and the first frame of the new scene are replaced by ImproveSceneSwitch. However, for some temporal filters like AlignFields only one of these frames should be replaced. For this purpose we have the boolean variables first and last. If first = false, then the frame remains unchanged, even if it is detected as the first frame of a new scene. Similarily, if last = false, then the frame remains unchanged even if it is detected as the last frame of a scene. first = true and last = true are the deafult values. ImproveSceneSwitchshould be applied as follows:

input = MPEG2Source("input.d2v")
dein = RemoveDirt(input)
ImproveSceneSwitch(dein, input)

If debug = true, then scene switch information is send to debugview. If extrapolate = true (false is the default), then instead of simple frame doubling extrapolation is used to calculate the scene switch frames. Though SSE can be very nicely used for this purpose, extrapolation is obvious much slower than frame doubling and compression is not as good. However, motion is handled better by extrapolation. Hence ratio may be chosen somewhat lower if extrapolate = true. On the other hand, if ratio is not too small, then the viewer should hardly be able recognize frame doubling. I would very much appreciate feedback in the forum about the question, whether to extrapolate or not to extrapolate. As soon as a new version AlignField is released, ImproveSceneSwitchwill be removed from RemoveDirt and moved to the AlignField plugin. In the AlignField the scene switch code will be removed, because ImproveSceneSwitchis better and faster. Currently I combine ImproveSceneSwitchwith AlignFields through the following Avisynth function

function AF3(clip input)
{
dein = AlignFields(input,mode=3, scene=0, topfield=true,
tff=true) return ImproveSceneSwitch(dein, input, ratio=5, first=true, last=false, field=2, debug=true)
} 

If topfield = true and tff = false, then first = false, last = true, field = 2 has to be chosen instead. If topfield = false and tff = true, then first = false, last = true, field = 1 has to be chosen instead. Finally, if topfield = false and tff = false, then first = true, last = false, field = 1 has to be chosen instead. To determine the correct values for other deinterlacers, one has to look at the source code. By the nature of the algorithm, the first two and the last two frames are always left unchanged. ImproveSceneSwitchis optimised for the situation, when the frames of the clip are requested sequentially. This is typical for an encoding process. It is almost 4 times slower if the frames are requested randomly or in reverse order.

By Rainer Wittmann "gorw at gmx.de"

$Date: 2004/08/17 20:31:19 $