This is yet another filter which cleans up the picture, averaging out visual noise between and within frames.
The Peach is designed to cope with the oddities of broadcast TV. It automatically fine tunes itself for good or poor quality signal and reception, and can even deal reasonably with some kinds of interference. It’s spiffy!
The filter examines the image, infers the amount of noise in the picture, and uses that estimate to determine how to cancel the noise. To differentiate between noise and motion, the filter makes use of correlations within and between fields as well as local color differences. The degree of certainty about motion, noise, and detail is used to decide what mix of the previous, current, and nearby colors to show.
To use the Peach, your computer will have to be able to run SSE instructions. In English, that means you will need to have a Pentium 3, Athlon, later Celeron, or any more recent processor.
This filter works in the YUY2 colorspace. So if your clip is encoded as RGB colors, you’ll first need to process it with ConvertToYUY2().
The Peach Smoother AVISynth filter was written by Lindsey Dubb, If you have any questions, comments, problems, or suggestions, you are welcome to send me an email or post to Doom9’s AVISynth Forum.
NoiseReduction
An integer from 1 to 200; 35 by default
Sometimes the filter doesn’t know whether change in the picture comes from noise or motion. When that happens, there’s a tradeoff — Guess motion, and the picture will be noisier if you’re wrong. Guess noise, and a mistake will blur the movement.
The NoiseReduction setting lets you tell the filter how to make this tradeoff. Set it low, and the computer will play it safe, avoiding any parts of the image which might be in motion. Set it high, and noise will be greatly reduced — at the cost of blurring and loss of detail in slowly moving areas. This tradeoff is affected by the amount of noise — The more noise there is in the image, the more blurring will be necessary in order to reduce it.
There’s no perfect way to choose this — Just try different values until it looks as nice as possible. Comedies and dramas are probably the best material for picking settings, since facial close-ups make any blurring really easy to see. (That’s because we’re especially good at seeing distortion in faces.)
Remember that this filter automatically adjusts for the amount of noise in the picture. So once you have a setting you like, you can use it for both clean and noisy video.
However, there are two kinds of video which deserve some special consideration: cartoons and sports. With cartoons, you can use a much higher NoiseReduction, since animation tends to have less fine detail. And with sports (or any other show with constant motion), there is very little to gain from a noise filter, so you can just skip it.
Before you fine tune this setting, make sure that the filter has been able to estimate the noise in your clip. To do so, make use of the Dot, described below.
Stability
An integer from -100 to 100; 20 by default
This setting tells the filter how much to mix colors when it is uncertain about whether there is motion. To preserve a small amount of noise throughout the picture, decrease this setting. To get a really solid looking picture, use a higher value. Don’t set it too high, though, or you’ll get posterization.
Regardless of your Stability setting, the filter will always preserve a certain amount of variation in order to maintain the color depth of the original image.
DoSpatial
A boolean; TRUE by default
This setting decides whether you want the filter to use spatial smoothing. It’s a very subtle smoother, and is used mostly in low contrast, moving parts of the picture. The filter is somewhat faster when spatial filtering is turned off.
Spatial
An integer from 0 to 400; 100 by default
This determines how much spatial smoothing to use. It’s measured in percent, relative to the amount of temporal smoothing. But the Peach prefers to use temporal smoothing whenever possible, so the spatial smoothing only kicks in when temporal smoothing fails.
Dot
A boolean; FALSE by default
It’s... a small green dot!
With this option, a tiny green dot will appear when the filter’s estimate of the noise is confirmed by the current picture. It will show up near the upper left of the screen — specifically 16 down and 16 across from the upper left corner. The dot is an indication that the filter has settled on a noise value. In general, it will turn off when all of the picture is in rapid motion — When there’s too much motion, the filter tends not to believe the current estimate. In that case, the filter makes do by weighting previous good values.
If the picture is constantly in motion, it’ll take a while before the filter can figure out the noise level. Interference also tends to make noise estimation take longer — The estimation time depends on the kind and amount of the interference.
Readout
A boolean; FALSE by default
To help decide how much to smooth, this filter measures the noise in the video. When you enable Readout, these measurements will be shown in the output.
You will see two numbers. The first one is a measurement of the mean noise level. The second is a measure of the lowest amount of noise seen anywhere in the picture. That provides some idea about how much noise varies in the video. These estimates will adjust as the video progresses.
The numbers are measured as the expected absolute change in a single pixel (counting both chroma and luma) between two adjacent frames.
This provides an objective measurement of noise. It is also the way to figure out values for the NoiseLevel and Baseline options, described below.
NoiseLevel and Baseline
Two settings — each is a floating point number from 0.0 to 100.0; Unused by default
Rather than estimate the noise level, you can tell the Peach exactly how much noise there is in the picture.
To do so, choose values for both NoiseLevel and Baseline. To get these values, use the Readout option, described above. Note that Baseline should never be more than NoiseLevel.
Why would you want to specify these numbers? Because the Peach can take a while to figure out the amount of noise. And on rare occasions it can be badly mislead. (This is usually caused by scenes with smoke and dust clouds, which the Peach can mistakenly identify as noise.) By specifying the noise values, the estimate will never waver. Skipping noise estimation also makes things run a little faster.
On the other hand, some video really does have changing amounts of noise. In that case, you’ll be much better off with the Peach’s usual noise detection.
ShowMotion
A boolean; FALSE by default
This is another obscure diagnostic. When this is turned on, the filter will interleave its estimate of whether motion is occuring. White areas are definitely motion, black areas are definitely stationary. Gray areas are gray areas.
Debug
A boolean; FALSE by default
This option is only meant for the stout of heart. For an explanation of its output, see the comments toward the bottom of FLT_AdaptiveNoise.c .
Let your eyes be the judge. I like to keep the settings pretty low — Blurred faces look worse than a little static.
Surprisingly, small amounts of color variation can sometimes improve an image. By switching back and forth between colors, the picture is able to give the impression of a color somewhere in between them. Also, noise can break up artifactual patterns in the picture, making it easier for you to ignore the errors.
With that in mind, the Peach tries to preserve a small amount of color variation. As a result, it will never give you a completely stable picture. This is probably bad for compression, but it does improve picture quality.
If you’re seeing lots of blurring, try using the Dot option. If the green dot doesn’t show up, then the problem is that the Peach wasn’t able to figure out how much noise there was.
There are a couple ways to solve this. The best is to turn on the Readout option, and watch a later part of the video where the picture is still. Make a note of the NoiseLevel and Baseline from that stationary part. Then specify them in your command, and the whole sequence should look fine.
If your video doesn’t have any stationary parts, then you should just skip this filter. A temporal smoother isn’t going to do much good for pure high motion material.
Another way to solve this problem is to put some still video (from the same clip) at the beginning of the sequence. That will allow the Peach to estimate the noise before the fast stuff shows up.
Try reducing the Stability setting.
Make sure you have a good signal. Noise can come from cable jumbles, poor connections, poor power and grounding, poorly designed video input cards, electrical gadgets (anything from a dimmer switch to various computer components) or from a bad video source. These issues are all beyond the scope of this help file. I’d suggest a look at the AV Science Forum ’s Home Theater Computers FAQ and board, where these topics are discussed at length.
It depends what you’re watching. Sports and nature shows in general — really anything with lots of moving low contrast texture — are not handled well by temporal smoothers. That’s because those textures look a lot like noise.
The Peach does well with difficult material, disabling itself where it detects motion. It causes surprisingly few problems with field sports so long as the background noise isn’t too bad. But it isn’t perfect — Road races are especially prone to blurring. When that happens, it’s best to skip any temporal smoothers.
Otherwise, it’s generally worth running a noise filter.
Be very careful about using more than one. Running multiple temporal noise filters can cause the later filters to have excess confidence in their noise/motion estimate. That usually leads to posterization, speckling, and banding.
Running a spatial filter after Peach Smoother can still work reasonably well, since its spatial smoothing is pretty subtle — Just be sure not to run the spatial filter first.
In general, it is best to filter noise as early as possible. This is especially true for the Peach, since its noise estimation can be thrown off by some filters. For example, if you subtitle your video, the Peach may detect the lack of noise in the subtitles.
With interlaced material, it is a very good idea to run the filter before deinterlacing. That’s because good noise reduction can greatly improve the accuracy of the deinterlacer.
For inverse telecine the situation is harder to judge. But as a rule of thumb, it is still better to run the noise reduction first.
The exception to this rule is comb filtering. Any comb filters should be run before noise reduction, since the noise filter will otherwise interpret color crosstalk as motion.
If you’re using a Bt8x8 card, you can turn off Odd and Even Luma Peaking. Turning on the card’s Horizontal Filter will reduce noise but will also lose some detail.
$Date: 2004/08/17 20:31:18 $