DE102005011584A1 - Image/audio data set filtering method, involves filtering data set having user and noise data with selected/automatically adjusted parameters, and selecting image of user data, which pertains to image of noise data that appears as real - Google Patents

Abstract

The method involves selecting parameters, and filtering image/audio data including user data and disturbance data such as noise data, with the selected parameters. Other filtering processes are executed on the data set by automatically adjusting the parameters. An image of the user data is selected, where the user data image pertains to an image of a disturbance data that appears as real. The image of the disturbance data appearing real is obtained by filtering. An independent claim is also included for a device for implementing a method of filtering a data set that is formed of user data and disturbance data e.g. noise data.

Description

The The invention relates to a method and a device for a determination user data from a data set, which is composed of user data and disturbance data. The invention particularly relates to a method and a device for one Noise Reduction.

In almost all flatbed scanners, film scanners and digital cameras come CCD sensors either in the form of CCD lines or CCD chips for Commitment. CCD sensors (CCD stands for "charge-coupled devices") consist of a Grid of many individual photosensitive diodes, now applies Light on one of these light-sensitive diodes, so-called "pixels", this is in converted electrical voltage. The brighter it is, the more tension is built. The picture is thus composed of a grid of brightness information. By separate brightness information of the three primary colors red, green, Blue, creates a color image.

Would not be Light falling on the CCD chip, so nothing should happen, since no light hits the photodiodes. That is not so, because electrons also move by heat, i. for example at longer Operating time. Furthermore For operating the sensor, a basic voltage is necessary, which also single electrons are set in motion.

Other Image capture devices such as digital cameras, MRI scanners, etc. also inevitably produce an undesirable Noise. unwanted Rauschenscheinungen exist not only in the capture of images, but, for example, in a comparable way among others the detection of sound. In particular, the noise leads to clutter or to a fault in the sense of the invention.

This Lets noise Reduce yourself by repeatedly measuring an image template. By averaging these pictures, you can hear the noise in the picture Reduce the result. This is because the noise is a random process is. The noise produces a different pattern on each image these patterns partially cancel each other out in the averaging on. However, the object is completely preserved in the averaging. However, such a method is especially for moving pictures not practicable. Furthermore the multiple detection is disadvantageous consuming and takes relatively a lot of time. Alternatively or additionally, noise filters are used. With a noise filter, for example, a sought after image or Sound record, so useful data in the sense of the invention separated from a background noise become.

It are a lot of noise filters. Representative is the anisotropic Called diffusion filter, with the typically sought image information should be freed from background noise. Such a noise filter For example, in "Current Methods of Laser and Medical Physics, 2nd Remagener Physikertage 2004, VDE Verlag GmbH, Berlin, Offenbach ", page 96 et seq.

The Principle of a diffusion filter is based on the so-called diffusion equation. A filtering step is expressed with a time step in the equation. The Special with the nonlinear diffusion is that one the diffusion process controls such that it is suppressed at edges. That causes a smoothing within structures. Edges remain untouched. A continuation non-linear diffusion is anisotropic diffusion, The diffusion processes also performs on edges, but only parallel (along) to the edges. So the edge shape is preserved, the noise at edges but also suppressed.

The Results that provide nonlinear diffusion are very dependent from the parameterization. An unfavorable parameterization leads either to incomplete suppression noise or structural alteration of image information. The optimum parameterization can vary greatly from data record to data record.

A the two main parameters of the diffusion filter is the number the iterations. You have to try out how many steps or Iterations feasible are without suffering information loss in the picture.

Of the other essential parameters for an anisotropic diffusion filter is the so-called diffusion factor. This can be approximate be calculated and / or adjusted by trial and error. The diffusion factor is ideally chosen to be that of the gradient amount noise, but not the gradient of an edge of the picture you are looking for.

Not only in the aforementioned anisotropic diffusion filter, but also with numerous other filters for a noise suppression exists the problem of having to set parameters manually.

So There are in image editing programs such as the commercially available Program "PHOTOSHOP" for removal of Dust and scratches on a digitized image the parameter "Radius" and the parameter "Threshold". "Radius" determines how big the area is where the filter looks for dissimilar pixels examined. The higher the radius is chosen becomes, the stronger will the blur of the picture. The "threshold" determines in which Measure that Brightness and color values of the pixels must differ from each other, so that the Filter is applied to it. The lower the value, the stronger it works the filter.

To free sound recordings of noise, there are commercially available filters, such as the so-called WavePurity ^TM FFT noise filter. With this music recordings are to be separated from a noise. The basic principle is that a noise (ideally a "white noise" is assumed) imparts a constant offset - that is to say a uniformly distributed energy to all wavelengths of the spectral range 0..22 kHz.

The Entrauschen takes place in the frequency range, that is, the signal is block by block transformed from the time domain to the frequency domain (FFT) and there a reduction of the noise by editing the Amplitu denwerte executed and subsequently back to the time domain transformed. The following steps are carried out:

Step 1

Analyze the volume of the piece of music

It will be first the volume of the entire title. there is block by block (block size = FFT window size). From Each block is the average volume. Step 1 notes now the minimum minAvgVolume and the maximum maxAvgVolume these middle block volumes. Not by artificial Silence to get confused, Step 1 accepts no blocks whose medium volume is less than 1 0 digits. Furthermore, the total mean value all block volume minima meanMinVolume formed.

Step 2:

sampling

Now comes in Random number generator that takes samples from the piece of music, but only accepted if the average volume of the Random position meets certain criteria. So you meet a certain preselection on blocks, which are close to the lower limit of the mean lowest noise, but still above the mean total block volume minimum are located.

Step 3:

Optimize the random values

Now runs one Optimization. Namely, the number of found digits will increase reduced relevant points. Criterion for good matching is the least variance of the mean amplitude values of all blocks.

Step 4:

Search for a cluster point for the sough

Now The noise filter starts to search in small increments for blocks that are random fit well with the previous points. Criterion for this is whether an improvement in the variance is achieved by the new point.

Step 5:

Determine noise profile

For the found Points will now calculate the noise profile.

Step 6:

Smoothing the noise profile

Often owns the noise profile "needles", which is characterized by Imprint interspersed signals (for example the 19kHz sound on stereo radio). Farther the noise profile itself is often very noisy. Based on experience the noise reduction reduces artifacts when using a moving average filter over the Noise profile is running to it to smooth. The highest frequency 22kHz is loaded as start value and then slowly down towards 0kHz.

Step 7:

Linearize the noise profile

Around one possible straight profile of the profile is the noise profile interpolated in the lower frequency band.

Step 8:

Determining the reduction profile

experimental was found out the calculated noise profile is always calculated as a reduction profile must become, which ultimately for the amplitude subtraction is used. There is 0% effect, if the detected noise profile is simply subtracted 1: 1 from the signal blocks becomes. This is a fact that is experimentally demonstrable.

Step 9:

Custom customization the reduction factor

Of the WavePurity FFT Noisefilters has a "slider" 50% to 150% noise canceling. The user so has the opportunity at critical pieces the strenght to influence the reduction profile.

Step 10:

Dynamic pursuit of the noise level

This Step is running virtually online during the Entrauschvorgangs from. And there are many recordings of especially older tape recorders where the Noise of the signal "pumps". That's an effect the FFT noise filter when determining the noise profile extreme Headache, because the noise is simply not stable is, and thus no clear accumulation in the histogram recognizable is. The determined noise profile is always for such recordings something falsified be. Leads to compensation one while of de-noise the reduction profile slightly after. So will the pumping the noise counteracts.

at this commercially available Need to process also be set manually parameters. Such a parameter is called "Slider" in this program.

at the setting of parameters is regularly particularly problematic that these influence each other. So it's a parameter already set optimally and will be another parameter then changed So this can have the consequence that now also the already optimal adjusted parameters must be readjusted.

With The aim of the present invention is to provide parameters Improved filtering and set in particular automated to be able to.

to solution of the problem with a filter so to speak tentatively several Filtering performed. From filtering to filtering, parameters are changed. The change The parameter may be random, among others of estimates and / or depending from previously used parameters and thus obtained Results are controlled. It results in this way in pairs various mappings of user data and associated clutter data. The way through filtering obtained information from images of a disorder or of disturbances, So for example, the reproduction of a noise are checked to see if it Realistically, it is an image of a real disorder can. The next only tentatively selected Parameters become final selected, with which an image of a disorder could be determined that seems realistic. The with these Parameters performed Filtering is selected around the clutter to obtain reduced mapping of the payload.

The actual filtering is done so with the parameters, with which the image of a disorder, So in particular a noise could be achieved, which an expected image of the disorder, so one of them Noise resembles. Preferably, the filtering is selected that has displayed disturbance data that was most similar to an expected mapping of clutter.

The Invention is based on the idea that before detection and Determination of image information, sound information or the like Although the searched user data are unknown. Are known, however, different Properties of the image of the disturbance to be filtered out or the disturbances. Is a filtering of a data set in the sense mentioned above carried out, so the correctness of the filtering on the basis of the "quality" of the image of the disorder be assessed. Corresponds to that determined by application of a filter Image of a fault not a real image (as it was to be expected), that means this is because the parameters of the filter were set improperly. The more similar z. B. detected by applying a noise filter image a noise resembles the properties of an image which was expected, the more likely the parameters were expected set. This principle makes the present invention use, in order to set parameters improved.

The each to be performed Choice and change the parameter of a filter as well as the verification of the pictures of the disturbances can be automated carried out become. In this way it is possible automatically set parameters of a filter.

In In one embodiment of the method, the variance of the by filtering considered noise, so a realistic image a noise in the sense of the present invention. The more similar Variables obtained by filtering the variance in a real one Noise is similar, the more realistic is the image of a filter obtained by filtering Disorder. The variance of the noise is thus included as a criterion to to judge whether the image of a Noise could be real.

In In one embodiment of the method, the image of a disturbance is completely scanned. At each step of the sampling is determined locally limited, which Variance within the given locality occurs, for example within a given volume element. A volume element in the sense of the present invention is - expertly expressed - a local Surroundings around a so-called voxel or pixel. It results in such a Variety of variance values, hereafter called "local variances".

It is assumed in one embodiment of the method that a real noise has the consequence that the local variances do not differ from each other or at least only slightly. If, for example, a distribution of the determined local variances is represented graphically, an image obtained by filtering is particularly real if the distribution function of the local variances has a high, narrow, symmetrical peak. The wider the peak, the less the image of noise produced by filtering corresponds to a real image. The less symmetric a frequency distribution of the local variances, the less the image of a noise determined by filtering matches with a real noise. Another criterion for assessing the frequency distribution of the local variances are the two areas that are included by the distribution function left and right - seen from the maximum of the peak. The better the size of the two surfaces matches, the more so The image of the noise determined by filtering approaches the real image. Assessment criteria for the selection of a real-appearing image of a noise are therefore individually or in combination the distribution symmetry, the width of a peak in the distribution function, the height of a peak in the distribution function and the size comparison of the areas to the right and left of the maximum of the peak.

In A further advantageous embodiment of the invention is a first filtering under specification of freely chosen parameters. The value of such a freely given parameter will be described below denoted by X0. Subsequently Two more filters are made. In the first case, instead of of X0 is chosen as the value of the relevant parameter (X0 - Y0) and in the second case the Value (X0 + Y0). Y0 is a freely chosen positive number. Subsequently, will from the three filterings picked out the one that is the most real Image of a fault in the sense of the present invention.

• 1. Die Filterung mit dem Parameterwert (X0 – Y0) ergab das am realsten erscheinende Abbild einer Störung. Es wird dann eine weitere Filterung durchgeführt, wobei der zuvor variierte Parameter weiter verkleinert wird. Insbesondere beträgt der Wert des weiter verkleinerten Parameters dann (X0 – 2·Y0). Mit dem weiter verkleinerten Wert des Parameters wird eine nächste Filterung durchgeführt und überprüft, ob durch die Verkleinerung ein Abbild einer Störung erzielt werden konnte, welches weiter verbessert einem real erscheinenden Abbild der Störung entspricht.
• 2. Die Filterung mit dem Parameterwert (X0 + Y0) ergab das am realsten erscheinende Abbild einer Störung bzw. von Stördaten. Es wird dann eine weitere Filterung durchgeführt wobei der zuvor variierte Parameter weiter vergrößert wird. Insbesondere beträgt der Wert des weiter vergrößerten Parameters dann (X0 + 2·Y0). Mit dem weiter vergrößerten Wert des Parameters wird eine nächste Filterung durchgeführt und überprüft, ob durch die Vergrößerung ein Abbild einer Störung erzielt werden konnte, welches weiter verbes sert einem real erscheinenden Abbild der Störung entspricht.
• 3. Die Filterung mit dem Parameterwert X0 ergab das am realsten erscheinende Abbild einer Störung. Es wird dann eine weitere Filterung durchgeführt, wobei der zuvor variierte Parameter weiter verkleinert und/oder vergrößert wird. Der Betrag, um den verkleinert oder vergrößert wird, wird dann kleiner als Y0 gewählt. Insbesondere beträgt der Wert des weiter verkleinerten bzw. vergrößerten Parameters dann (X0 – ½ Y0) bzw. (X0 + ½ Y0). Mit dem weiter verkleinerten und/oder vergrößerten Wert des Parameters wird eine nächste Filterung durchgeführt und überprüft, ob durch die Verkleinerung bzw. Vergrößerung ein Abbild einer Störung erzielt werden konnte, welches weiter verbessert einem real erscheinenden Abbild der Störung entspricht.

There are three cases:

• 1. Filtering with the parameter value (X0 - Y0) gave the most realistic image of a disturbance. It is then carried out a further filtering, wherein the previously varied parameter is further reduced. In particular, the value of the further reduced parameter is then (X0-2 · Y0). With the further reduced value of the parameter, a next filtering is performed and checked whether the reduction could be an image of a disturbance could be achieved, which further improved corresponds to a real appearing image of the disturbance.
• 2. Filtering with the parameter value (X0 + Y0) gave the most realistic image of a disturbance or disturbance. It is then carried out a further filtering the previously varied parameter is further increased. In particular, the value of the further enlarged parameter is then (X0 + 2 * Y0). With the further enlarged value of the parameter, a next filtering is carried out and checks whether the magnification of an image of a disturbance could be achieved, which further verbes sert corresponds to a real appearing image of the disturbance.
• 3. The filtering with the parameter value X0 gave the most realistic image of a disturbance. It is then carried out a further filtering, wherein the previously varied parameter is further reduced and / or increased. The amount to be reduced or increased is then selected smaller than Y0. In particular, the value of the further reduced or enlarged parameter is then (X0 - ½ Y0) or (X0 + ½ Y0). With the further reduced and / or increased value of the parameter, a next filtering is carried out and checks whether the reduction or enlargement an image of a disturbance could be achieved, which further improved corresponds to a real appearing image of the disorder.

By the aforementioned embodiment of the method becomes the change controlled by a parameter. The controller makes sure that changed a parameter like that will that improvements in filtering with comparatively high probability are to be expected. The probability of a Improvement is following the first three filters performed in principle at 50%. Overall, this is compared to a purely random change of parameters avoided, the filtering and computing costs become large allow.

The The aforementioned principle can be applied several times in succession. It is checked with each new filtering, whether the result of a improved by filtering obtained image of a disorder and in dependence one of them will be next Filtering performed and thereby the parameter changed in the aforementioned manner controlled.

The Procedure can be terminated if a predetermined number of changes of a parameter has been achieved and / or no noticeable Improvements in the image of clutter obtained by filtering more achieve.

For example in the case of an anisotropic noise filter, it is possible to Approximately the presumably set parameter "diffusion factor" to be calculated. In such cases, a first value of a Parameters advantageously calculated once. From the calculated Based on value then the method is applied. In this way let yourself further reduce the computational effort.

If an optimum value for one of two parameters A and B of a filter, that is to say for example for a parameter A, the optimum value A0 has been determined, the parameter B is subsequently changed in an advantageous embodiment of the invention. For example, parameter B initially had the value B0, and for parameter B0 the value A0 was determined to be optimal. Subsequently, the parameter B is changed and its value is now B1. An optimal value for the parameter A is again determined. This amounts to A1. Subsequently, it is compared which image of disturbance data obtained by the filtering appears more real, ie the image obtained with the parameters A0 and B0 or the image which has been obtained by selecting the parameters A1 and B1. This refinement of the method makes it possible to take account of the circumstance that the change of one parameter can result in the fact that the optimum value of the other parameter can also change as a result. This constellation is the case, for example, with the anisotropic diffusion filter. An "optimal" value for the "diffusion factor" parameter changes depending on the value of the "number of iterations" parameter.

The each to be performed change the parameter B, so for example the number of iterations in the case of an anisotropic diffusion filter, in turn can be controlled respectively. It will then be first images of by filtering won disturbances of a total of three parameter pairs (A0, B0), (A1, B1) and (A2, B2) determined. Let A0, A1 and A2 is the optimal parameter value that was found when the parameter value B0, B1 and B2 has been specified. It applies for example in one embodiment of the invention then B0 = B1 - Y1 and B2 = B1 + Y1, where Y1 is a positive number. It results again the three above-mentioned cases. So it may turn out that the value pair (A2, B2) gave the best result. Then would be the Parameter B to be further enlarged. This is then advantageous B3 = (B0 + 2 · Y1). It becomes an associated optimum value A3 determined. It is then checked if that by filtering obtained image of disturbances could be further improved in the context of the invention. Let it controlled in this embodiment, both parameters optimally adjust.

in the The case of a diffusion filter will be advantageous in the number of iterations given and depending on it the optimal diffusion factor is determined. The parameter "iteration" is then the Parameter "B" the number of iterations, so the parameter B changed and an associated optimal parameter "diffusion factor" has been determined It turns out that the computational effort is kept small Comparison to the reverse case.

Around to get to a good filtering faster, is in one advantageous embodiment of the invention, a parameter B within a given range varies, the experience shows good or at least leads to useful results. For example, have yourself in the case of an anisotropic diffusion filter, that the number of iterations between 11 and 15 is around a noise to filter out correctly from a data record. In such a Case is advantageous in one embodiment of the method Fixed range within which the parameter B varies becomes. So lets the required computational effort can be significantly reduced.

It For example, let the parameter A be the diffusion factor and the parameter B is the number of iterations. Experience has shown that a well or optimally chosen one Parameter B is always between 11 and 15. It can be precautionary the given range increases slightly or exactly this Limits are met. The limits for parameter B then become 10 to 16 or 11 to 15 fixed. Within this predetermined Boundaries, the parameter B, inter alia, at random, systematically or controlled varies.

It it has turned out that in such a case it is beneficial is to change the parameter B systematically. This is to be understood that the parameter B varies according to a predetermined fixed rule For example, in the case of boundaries 10 through 16, the entire Area gradually z. B. from bottom to top or from top to bottom is traversed below. For example, the result may be: 10, 12, 14, 16. So it will be first B0 = 10 is selected and subsequently the associated one optimal parameter A0 determined. Subsequently, the procedure with the parameter B1 = 12, the associated optimal parameter A1 determined and so on. From pair A, B becomes the pair selected, which has given a real-appearing result of clutter. It has It turned out that such a good compromise between desirable Speed and desired Accuracy is obtained.

Around to further accelerate the process without too much loss to accept accuracy is found in a further embodiment of the method of the Parameters A, so for example the values A1, A2, A3, A4, A5 the median or central value formed, ie the boundary between two same size halves picked out. From a sorted number of values A becomes so the value chosen, the lies in the middle. For an even number, the arithmetic Means of the two middle values are formed.

For example, the following value pairs have been determined after systematic variation of parameter B:

out Wanted becomes the median at the A-values, in our example case the value 15. The associated B value is then 40. The filter is then with the parameters A = 15 and B = 40. The result of a mapping of User data then represents the final result of the method.

embodiment

The The invention was in an example with an anisotropic diffusion filter carried out. Filtered was a high-resolution, T1-weighted nuclear magnetic tomographic 3D image acquired by gradient echo sequence was won. Due to its high resolution (<1 mm) and the limited recording time less than 20 minutes proves the quality of a such a 3D image as not ideal. There is therefore a need for that Suppress noise.

Step 1: Choice of diffusion filters

Of the User can choose between different diffusion functions to to process the data. The second Perona-Malik function (PMAD2), if it should arrive at a smooth result, or the Weickert m = 2 function (Weickert2m), if better preservation of the edges in the foreground, exposed.

In At this stage, the user can already specify a number of neighbors, which is considered in the method in the context of local variance should be. For two-dimensional datasets is a 4 or 8 neighborhood advantageous, while in the three-dimensional case a ber-, 18er, or 26er neighborhood is to be preferred.

Step 2: Estimation of the noise levels

The estimation of the noise level is performed by calculating the "median absolute deviation" (MAD) First, it is necessary to calculate a gradient image from the data set (∇I ( u )) is determined by the sum of the differences between each element of the image and its neighbors (defined in the previous step). The MAD can be calculated using the following formula: MAD = 1.4826 · median I (| ∇I ( u ) - median I (| ∇I ( u ) |) |)

By this formula becomes an expected (global) variance of the image determines that for the first estimate the diffusion factor is necessary.

Step 3: Choice of the first Filter parameter:

The first diffusion factor value is given by the formula k1 = 0.5 · MAD for the PMAD2 function or k1 = 0.45 · MAD for the Weickert2m function determined while the number of iterations (lt0) is selected based on experience (typically between 10 and 20).

Two further diffusion factor values are calculated, for example according to k0 = k1 - (x% k1) and k2 = k1 + (x% k1).

Advantageous x% is between 10% and 30%, for example at 20%.

There are now three parameter combinations available, namely the parameters
[k0, lt0], [k1, lt0] and [k2, lt0],

Step 4: Processing the first 3 parameter combinations

The Three combinations are processed with the selected diffusion function. After each result the picture of the disturbance data (residual or disturbance picture), by subtracting the image of the user data (original) from the image of the data record (processed image).

A homogeneous texture in the disorder picture would, due to the assumed uniform variance distribution of the Noise, mean that the filter parameters that produces the image have, are optimal. An irregular texture would mean that either the picture was not smoothed enough, or that the picture smoothed too much and some structures in the original were also smoothed out.

Step 5: Variance Analysis

The three resulting perturbation images are now analyzed according to the invention. First, the local variance is calculated for each pixel in each image. The local variance calculation is an operation that compares the intensity of its neighbors to its own intensity value for each pixel using the following formula.

The number of neighbors (N-1) was given in step 1 and corresponds to a two-dimensional image, for example, an eight-neighborhood and in the case of the three-dimensional image, for example, a neighborhood of eighteen. μ _u is the mean in the selected neighborhood.

Step 6: Calculation of the histogram

For each Image from step 5, a histogram is calculated. The histogram is a one-dimensional representation of the frequency distribution of the local Variances in the respective picture. The histogram comes with a one-dimensional smoothing filter (Gaussian filter) processed.

Step 7: Evaluation function

To extract the noise characteristics described in step 4 from the histogram, an evaluation function is used. This function considers six different properties of the resulting histogram to define the quality indicators.

The meanings of the individual placeholders of the evaluation function result from the 1 and 2 , These figures show a determined distribution of local variances. FWHM corresponds to the width of the peak at half height. FW20% M corresponds to the width of the peak at a height of 20% of the total height (height) of the peak. These widths are subdivided into Left _HWHM , Right _HWHM , where "Left" is the left side and "Right" is the right side, as viewed from a vertical, passing through the maximum of the peak. The same applies to Left _{HW20% HM} and Right _{HW20% M.} "Lf" means "if".

Balance is the quotient between the two surfaces (areas 1 and 2) of the curve below the function of the frequency distribution, where one surface is separated from the other surface by a vertical line passing through the maximum of the frequency distribution peak. This will be in 2 shown. One surface has been hatched. The smaller value is divided by the larger value.

If the fault picture would correspond to the ideal case the local variance operator present a homogeneous picture. In the histogram would this homogeneous operator response represents a long and narrow peak. The evaluation function then returns a large value as a result.

If in the fault picture an irregular texture is contained, the local variance operator becomes inhomogeneous Present picture, with a strong dispersion of values. The histogram then sets a broad and flat curve. The result of the evaluation function is then a small value.

After this the three parameter combinations from step 3 go through the process you have three statements about the quality the results that each of these combinations produced.

Step 8: Optimization rule

With corresponding results of the three parameter combinations it now possible a statement about the direction and size of the required change of the diffusion factor.

• 1.: Die Größe des Ergebnisses der Evaluierungsfunktion bestätigt die Richtung der Änderung des Diffusionsfaktors, das heißt, dass die vorletzte Parameterkombination ein schlechteres Ergebnis als die letzte Parameterkombination erzeugt hat und dass die vor-vorletzte Parameterkombination ein schlechteres Ergebnis als die vorletzte erzeugt hat. Dies bedeutet, dass der neue Diffusionsfaktor eine Verkleinerung bzw. Vergrößerung in der ermittelten Richtung mit der gleichen Schrittgröße darstellt.
• 2.: Die Größe des Ergebnisses der Evaluierungsfunktion steht im Widerspruch zur bisherigen Richtung, das heißt, dass die vorletzte Parameterkombination bessere Ergebnisse als die letzte erzeugt hat und, dass die vor-vorletzte Parameterkombination bessere Ergebnisse als die vorletzte erzeugt hat. Dies bedeutet, dass der neue Diffusionsfaktor eine Verkleinerung bzw. Vergrößerung in der entgegengesetzten Richtung mit der gleichen Schrittgröße darstellt.
• 3.: Das Ergebnis der Evaluierungsfunktion zeigt, dass innerhalb zwei der drei Parameterkombinationen ein Maximum liegt. Entweder ist die vorletzte Parameterkombination besser als die letzte und die vor-vorletzte schlechter als die vorletzte, oder die vorletzte Parameterkombination ist schlechter als die letzte und die vor-vorletzte besser als die vorletze. Dies bedeutet, dass der neue Diffusionsfaktor genau in der Mitte des Bereiches der besten zwei Parameterkonfigurationen liegt.

There are three ways:

• 1: The size of the result of the evaluation function confirms the direction of the change in the diffusion factor, that is, the penultimate parameter combination has produced a worse result than the last parameter combination and that the pre-penultimate parameter combination has produced a worse result than the penultimate. This means that the new diffusion factor represents a reduction or enlargement in the determined direction with the same step size.
• 2 .: The size of the result of the evaluation function is in contradiction to the previous direction, that is, the penultimate parameter combination has produced better results than the last one, and that the pre-penultimate parameter combination has produced better results than the penultimate one. This means that the new diffusion factor represents a reduction in the opposite direction with the same step size.
• 3 .: The result of the evaluation function shows that there is a maximum within two of the three parameter combinations. Either the penultimate parameter combination is better than the last and the before-penultimate worse than the penultimate, or the penultimate parameter combination is worse than the last and the pre-penultimate better than the vorletze. This means that the new diffusion factor is exactly in the middle of the range of the best two parameter configurations.

With the new calculation of the diffusion factor becomes a new parameter combination compiled, namely [kneu, lt0].

Step 9: Update the parameter combinations

in the Framework of the iterative method and in particular the optimization rule becomes the new parameter combination and the corresponding result the evaluation function in the group of the last three examined Integrated parameters. The worst of the three previous parameter combinations is discarded.

Example: If [k0, lt0] [k1, lt0] and [k2, lt0] are the worst result with [k2, lt0], the new parameter combination is composed as follows:
[k2, lt0] = [k1, lt0], [k1, lt0] = [k0, lt0] and [k0, lt0]. = [kneu, lt0]
and their results of the evaluation function:
Mp ([k2, lt0]) = Mp ([k1, lt0]), Mp ([k1, lt0]) = Mp ([k0, lt0]), Mp ([kneu, lt0])
will be calculated in the next step.

Step 10: Processing of the original data record with the new parameter combination.

in the Unlike step 4, this time only the filtering for the new one Calculated parameter combination. The information from the others two combinations is already calculated.

Steps 11 to 13:

Complies Steps 5 through 7. The only difference is that this time only the new parameter combination is evaluated.

Step 14: Optimization rule (same as in step 9)

At At this point, the loop is closed and the procedure can be repeated until a certain number of loop iterations carried out is or until the difference between the results obtained the three parameter combinations are very small.

Step 15: Selection of best parameter combination based on lt0

To the Final receives so the parameter combination that produces the largest evaluation value Has. This combination is considered an optimal combination based on lt0 selected.

Step 16: New number the iterations

One new value for the number of iterations is selected to produce three new parameter combinations, namely [k0, ltl] [k1, lt1] and [k2, lt1]. Steps 4 to 15 will be done again.

Step 17

Of the Step 16 is repeated three more times to finish with five different "optimal" parameter combinations to obtain

Step 18: Selection of optimal combination

Of the five determined optimal parameter combinations [k _Optimum 0, lt0], [k _Optimum 1, It1], [k _Optimum 2, lt2], [k _Optimum 3, lt3], [k _Optimum 4, lt4] becomes the median value of the diffusion factor selected to obtain the final optimal solution for the original image.
Median (k _optimum 0, k _optimum 1, ...)

The associated Number of iterations is already given.

Step 20: Final processing

The original data record is processed with the determined optimal parameter combination or the already determined result. The result is saved as a result of the procedure.

The method was applied to such a 3D image. 3 shows the result of the evaluation function M _p over the parameters "iteration" and "diffusion factor" of an anisotropic diffusion filter.

Claims (11)

Method for filtering one of payload data and Disturbance formed record with the steps: - Selection of parameters and execution filtering with the selected ones parameters, - Implementation of one or more further filterings with changed parameters or changed ones parameters, - Selection a filtered image of disturbance that appears real, - Selection the image of user data that to the appearing as real image of disturbances belongs. The method of claim 1, wherein the variance of a disorder is included to select a real-world image of clutter. Method according to one of the preceding claims, in which frequency distributions of local variances in filtered images of Disturbance be determined and included as a criterion to make a real-life Image of the disturbances select. Method according to the preceding claim, in which the width of a peak of the frequency distribution, the height a peak of the frequency distribution, the symmetry of a peak of the frequency distribution and / or a size comparison of surfaces included left and right below the maximum of the peak as a criterion be as a real-looking image of the by filtering won disturbances select. Method according to one of the preceding claims, in the parameters of a filter are set automatically. Method according to one of the preceding claims, in the parameters "iteration" and "diffusion factor" in a diffusion filter be set automatically. Method according to one of the preceding claims, in a noise from an image data set or sound record filtered out becomes. Method according to one of the preceding claims, in the three filters performed exactly one parameter of the filter used by filtering changed to filtering becomes, from the three images of disturbances, the most realistic image of disturbances is selected, and subsequently at least one further filtering is carried out, the parameter being further increased, if filtering with the largest parameter the most realistic image of clutter has given the parameter is further reduced if the filtering with the smallest parameter the most realistic image of clutter has given the parameter enlarged or reduced becomes, if the filtering with the middle parameter the most real Appearing image of disturbances which parameter is then changed within the limits, those through the previously used largest and smallest parameters be formed. Method according to one of the preceding claims, in a parameter varies within pre-set limits in particular systematically in predefined steps. Apparatus for carrying out the method according to one of the preceding claims, with a programmed unit of calculation, the following steps carried out become: - Selection of parameters and implementation filtering with the selected ones parameters, - Implementation of one or more further filterings with changed parameters or changed ones parameters, - Selection a filtered image of disturbance that appears real, - Selection the image of user data that to the appearing as real image of disturbances belongs. Device according to the preceding claim, characterized in that the device has a Camera or a scanner is.