ADAPTIVE CLEANING OF MEASUREMENT DATA TO REMOVAL OF DIFFERENT TYPES OF DISTURBANCES

Description

[0001] The present invention relates to signal processing for physical measurement data that are subject to noise and other types of disturbances. State of the art

[0002] The automation of many previously manually managed physical processes requires the automated evaluation of physical measurement data obtained during process observation. The evaluation results are used to generate control signals, which in turn can be used to intervene in the process.

[0003] An important example of such a process is at least partially automated driving, in which measurement data from the observation of the vehicle's surroundings form the basis for the next actions performed by the vehicle.

[0004] A challenge here is that the physically collected measurement data is usually subject to noise and other types of interference. Such interference can be amplified during automated analysis, such as classification and semantic segmentation using artificial neural networks, and lead to misclassifications. Therefore, the measurement data is usually preprocessed. However, inappropriate preprocessing can lead to artifacts and further exacerbate the influence of the interference.

[0005] WO 2019 / 007 632 A1 discloses tailoring the preprocessing of images in such a way that a downstream evaluation module designed to recognize specific objects or features can evaluate the preprocessed image in the best possible way with regard to these objects or features.

[0006] NI ZEHAO ET AL: "Adaptive Image Filtering Based on Convolutional Neural Network," September 26, 2018 (2018-09-26), ANNUAL INTERNATIONAL CONFERENCE ON THE THEORY AND APPLICATIONS OF CRYPTOGRAPHIC TECHNIQUES, EUROCRYPT 2018; [LECTURE NOTES IN COMPUTER SCIENCE; LECT.NOTES COMPUTER], SPRINGER, BERLIN, HEIDELBERG, PAGES 369 - 379, XP047487199, ISBN: 978-3-642-17318-9 discloses a convolutional neural network (CNN) capable of automatically identifying salt and pepper noise, Gaussian noise, and random noise based on deep learning training. Median filters, mean filters, and Wiener filters are then used to filter the corresponding images. Disclosure of the invention

[0007] The invention is described in the appended set of claims.

[0008] Within the scope of the invention, a method for training a trainable module was developed. This module is used to process recordings of a measurement signal that contain noise.

[0009] The term "noise" includes not only statistical disturbances such as fluctuations or noise, but also systematic errors such as blurring of an image due to defocus or motion blur.

[0010] The measurement signal is obtained through a physical measurement process, and / or through a partial or complete simulation of such a measurement process, and / or through a partial or complete simulation of a technical system observable by such a measurement process. This means that technically produced recordings of the measurement signal, unlike purely synthetic data, are subject to noise.

[0011] A trainable module is, in particular, considered to be a module that embodies a function parameterized with adjustable parameters with a high degree of generalization power. During training of a trainable module, the parameters can be adjusted in particular such that, when learning input values ​​are input into the module, the corresponding learning output values ​​are reproduced as accurately as possible. The trainable module can, in particular, contain an artificial neural network (ANN), such as a convolutional neural network, and / or it can be an ANN.

[0012] In this method, a trainable module is provided which, in the trained state, is configured to select a combination of interference suppression modules from a predefined catalog of interference suppression modules for a recording of the measurement signal, the application of which to the recording is expected to improve the suppression of the interference contained in the recording. The trainable module thus receives the recording of the measurement signal as input and delivers the selection as output.

[0013] The recording of the measurement signal can in particular include, for example, an image representing a spatial distribution of at least one physical measurement variable and / or a time series representing a time course of at least one physical measurement variable.

[0014] For example, the selection can indicate for each available interference suppression module whether, and if so, to what extent, this interference suppression module should affect the recording of the measurement signal in order to suppress the interference contained therein. For example, the selection can provide a numerical rating for each interference suppression module indicating the extent to which the application of the respective interference suppression module is considered appropriate. This numerical rating can then be further refined, for example, by selecting only the interference suppression module with the best rating and ignoring all other interference suppression modules.

[0015] For at least one training recording of the measurement signal, a combination of noise suppression modules is selected using the trainable module. The selected combination is applied to the training recording to obtain a noise suppression result. Using a predefined cost function, an evaluation is determined of the extent to which the noise is suppressed in the noise suppression result.

[0016] During training, a set of parameters is optimized. This set of parameters characterizes, at a minimum, the behavior of the trainable module. The optimization is aimed at improving the evaluation of the resulting denoising result by the cost function when reprocessing the same training recording and / or additional training recordings.

[0017] It was recognized that in this way, the processing of the measurement signal can be adaptively adjusted to the respective disturbances present, even if these fluctuate over time. Different disturbances can be based on very different physical mechanisms, so different countermeasures are indicated to suppress them. For example, the influence of noise can be at least partially eliminated by blurring, while the influence of defocus blur or motion blur can be exacerbated by blurring. Furthermore, adaptive selection can, for example, prevent interference suppression modules from taking effect when they are not really necessary.For example, filtering an image in the spatial frequency domain using Fourier analysis or wavelet analysis can largely remove noise from a noisy image, while the same filtering in a noise-free image merely results in a loss of detail. If a disturbance is only present temporarily, for example, when a vehicle drives through a poorly lit area while taking pictures, the corresponding noise reduction modules only affect images taken during that time.

[0018] Finally, using the trainable module to select the noise reduction modules to be applied allows for greater freedom and transparency regarding the noise reduction methods that can be applied to the recording compared to using the trainable module directly as a noise reduction module.

[0019] For example, if a trainable module contains an ANN with a predefined structure whose behavior is characterized by a set of parameters, it is difficult to determine at which parameter values ​​a specific desired denoising method, such as a specific filtering, will be considered by the ANN. By decoupling the actual denoising process from the selection, the catalog of denoising methods to be considered can be specified in concrete terms.

[0020] Furthermore, this decoupling prevents highly nonlinear or even non-differentiable transfer functions in noise suppression modules from directly affecting the ultimate dependence of the cost function on the parameters that characterize the behavior of the trainable module. The cost function, expressed as a function of these parameters, is then more "benign," allowing, for example, parameter optimization using a gradient descent method to converge better.

[0021] Each interference suppression module can thus translate the measurement signal received as input into an interference suppression result using any transfer function and deliver this as output.

[0022] For example, there may be transfer functions that depend only on the image received as input, such as subtracting the mean of all numerical values ​​contained in the image (e.g., intensity values ​​of image pixels), filtering with a fixed filter characteristic, or even an unaltered copy of the image in the denoising result. The latter "dummy denoising module" can advantageously be included in the predefined catalog, for example, so that the trainable module can signal that the process has converged toward a final result by selecting precisely this denoising module in a recursively guided denoising process.

[0023] However, there may also be transfer functions that have free parameters. Instead of multiple filters, each with predefined filter characteristics, only one filter can be stored in the predefined catalog, with the filter characteristics being set using the still free parameters. The noise reduction module can also be implemented, for example, as an autoencoder network that first translates the input recording into an intermediate state with low dimensionality and then reconstructs it from this intermediate state. Such an autoencoder network has a large number of free parameters that need to be trained.

[0024] Therefore, in a particularly advantageous embodiment, the parameter set optimized with a view to improving the evaluation by the cost function additionally characterizes the behavior of at least one interference suppression module. For example, the selection of an interference suppression module and the interference suppression module itself can then be trained together.

[0025] Using the cost function, any type of prior knowledge ("ground truth") available regarding the training image can be fed into the training process. Based on a sufficient number of training images, the trainable module can then learn to evaluate any image to determine which types of interference these images contain and which measures can be used to suppress these interferences. The trainable module can then fulfill a function comparable to that of a human image processing expert. If you present such an expert with a photo, they will immediately recognize which problems impair the quality of the photo. They can then accurately access the tools in the image processing software that will correct these problems.

[0026] For example, the cost function can depend on the result of a comparison between the noise reduction result and a known learning noise reduction result. Learning data for the trainable module can be obtained, for example, by subjecting undisturbed recordings of a measurement signal, such as images, to noise of a defined type and intensity. The noise-impregnated recording then becomes the learning recording, for whose noise reduction the trainable module should select one or more suitable noise reduction modules. The undisturbed original state of the recording is the learning noise reduction result, which should ideally be reproduced when applying the selected noise reduction modules to the learning recording.

[0027] Alternatively, or in combination with this, the cost function can depend, for example, on an automated evaluation of the noise reduction result based on at least one predefined criterion. For example, the intensity of noise in an image can be determined based on statistical criteria, such as a histogram of intensity values ​​and / or the standard deviation. This does not require an undistorted version of the image.

[0028] Alternatively, or in combination with this, the cost function can depend, for example, on an assessment of the denoising result obtained from at least one external evaluator. In this way, the expert's knowledge regarding the quality of denoising recordings can be preserved and converted into a machine-usable form. Until now, this type of knowledge has been difficult to document.

[0029] In a further advantageous embodiment, the interference suppression result is processed into a work result by a predefined processing module. The cost function then depends on the result of a comparison between the work result and a known learning work result. The processing module can be designed, for example, as a classifier and / or regressor that provides a classification and / or regression of the interference suppression result as a work result.

[0030] In the example mentioned above, in which an initially undisturbed recording is subjected to disturbances of a defined type and intensity, the undisturbed recording can be fed to the processing module to obtain the learning work result. If the disturbed recording has been de-noised by the de-noise module(s) selected by the trainable module and the resulting de-noise result is also fed to the processing module, the processing module should deliver a result that is as close as possible to the said learning work result.

[0031] In a particularly advantageous embodiment, a trainable module is selected that contains an artificial neural network. In such a network, the task of selecting the appropriate interference suppression module(s) for a given recording of the measurement signal can be particularly easily broken down into two subtasks. In a first step, the interference contained in the recording can be characterized in more detail, and in a second step, the selection of the interference suppression module(s) can be made based on this knowledge.

[0032] The characterization of the disturbance can be performed, for example, with an initial sequence of convolutional layers that gradually reduce the dimensionality of the image. For example, this sequence can contain between 2 and 7 convolutional layers, with one or more pooling layers optionally arranged between the convolutional layers.

[0033] The interference suppression module(s) can be selected, for example, using a second sequence of fully cross-linked layers. This sequence can, for example, comprise between two and four fully cross-linked layers.

[0034] In another particularly advantageous embodiment, an intermediate result provided as output by at least one of the convolutional layers is evaluated with a pre-training cost function. Parameters that characterize the behavior of the first sequence of convolutional layers are optimized in advance with the aim of improving the value of this pre-training cost function.

[0035] The intermediate result can, for example, be taken from the penultimate convolutional layer, i.e., the convolutional layer immediately adjacent to the first fully connected layer. For the exemplary application of denoising images, the intermediate result can, for example, comprise approximately 500 numerical values. This type of pre-training essentially focuses only on the perturbation itself and largely ignores the semantic content of the image. This suppresses the tendency of the subsequent training of the complete parameter set to always link the detected perturbations to a semantic content of the recorded measurement signal.

[0036] The pretraining cost function can depend, in particular, on the result of a comparison between the intermediate result and known additional information about the disturbance. In the example mentioned, in which an initially undisturbed recording is subjected to a defined disturbance, it can be evaluated, for example, whether this disturbance is correctly detected.

[0037] The invention also relates to a further method for processing recordings of a measurement signal that contain interference.

[0038] Within the scope of this method, at least one trainable module trained using the method described above is first provided. This provision may, for example, involve training the trainable module as described above. However, the trainable module can also be obtained from any source in a fully trained state.

[0039] At least one recording of the measurement signal is fed to the trainable module, allowing the trainable module to select a combination of interference suppression modules from the predefined catalog of interference suppression modules. The selected combination is applied to the recording, yielding an interference suppression result. The method thus relates to the effective operation (inference) of the trainable module in the trained state for the purposes of a practical application.

[0040] As explained above, the training described above has the effect that exactly those noise suppression modules are applied to each recording that are really necessary to suppress the specific noise present in that recording.

[0041] Advantageously, the noise reduction result is further processed by a predefined processing module into a working result. As previously described, in this context, the improved noise reduction of the recording has the effect of increasing the probability of obtaining the correct working result in the context of the respective application. This is especially true if the processing module is designed as a classifier and / or regressor that provides a classification and / or regression of the noise reduction result as the working result.

[0042] In another particularly advantageous embodiment, a control signal is generated from the interference suppression result and / or the work result. This control signal is used to control a vehicle, and / or a classification system, and / or a system for quality control of mass-produced products, and / or a medical imaging system. In this context, the improved quality of the interference suppression result or the work result results in a higher probability that the respective technical system will execute the optimal action from the perspective of the respective application.If, for example, traffic signs or other road users are semantically classified with greater accuracy as part of object recognition for at least partially automated driving, the vehicle controlled on the basis of this object recognition will be more likely to drive as expected in light of the traffic situation.

[0043] Both during training and in actual operation, the trainable module can, for example, provide a weighting with which the interference suppression results delivered by various interference suppression modules are to be combined, as a selection of a combination of interference suppression modules. This reflects the realization that many interference is complex and cannot be remedied with a simple individual measure. This is especially true when the specified catalog of interference suppression modules is very finely granular, i.e., each individual interference suppression module has a comparatively simple structure. A catalog with this level of granularity offers, on the one hand, greater flexibility for dealing with a larger class of possible interference. On the other hand, implementation, for example in the form of software, is also facilitated.Testing several small modules independently can be much easier than testing a monolithic module that combines the functionalities of these small modules.

[0044] However, multiple noise reduction modules can be applied to the measurement signal recording even if the trainable module always selects only one for application. For this purpose, the noise reduction result can be fed back to the trainable module as a noise-affected recording of the measurement signal. This means that noise reduction can be performed recursively, with one aspect being processed at a time. This recursion can be repeated, for example, until the noise reduction result converges according to a specified criterion.

[0045] As previously explained, the trainable module can indicate convergence, for example, by proposing a "dummy denoising module," which copies its input identically to its output. This could be due, for example, to the denoising result reaching a Pareto optimum. It is then no longer possible to improve one aspect of the denoising result without worsening another aspect.

[0046] In another particularly advantageous embodiment of both training and operational mode, at least one interference suppression module contains several submodules and a further trainable module. The further trainable module controls the extent to which the effect of this interference suppression module is combined from the effects of one or more submodules. This combination can be achieved, for example, through the weighting explained above or through the recursion explained above.

[0047] Arrangements consisting of a trainable module and several interference suppression modules, which this trainable module has "at hand" for selection, can thus be nested. One possible application of such nesting is to expand an existing arrangement in which the trainable module is already fully trained with additional interference suppression modules. The previously completed training then does not need to be discarded, but can continue to be used in its entirety. Integrating the additional interference suppression modules with the additional trainable module requires significantly less complex training of this additional trainable module.

[0048] For example, if the existing arrangement contains 50 interference suppression modules and is to be expanded by three additional interference suppression modules, it is not necessary to start training from scratch with a total of 53 interference suppression modules. The newly added trainable module now has the choice of selecting the existing arrangement of the 50 interference suppression modules as a whole or selecting one of the three newly added interference suppression modules. This choice between only four modules is significantly easier to train than a choice between 53 modules. If the existing arrangement is selected, the original, already trained trainable module decides within this arrangement to what extent which of the original 50 interference suppression modules are used.

[0049] The invention also relates to a parameter set with parameters that characterize the behavior of a trainable module and were obtained using the training method described above. These parameters can, for example, be weights with which inputs from neurons or other computing units in an ANN are calculated to activate these neurons or computing units. This parameter set embodies the effort invested in the training and is therefore a standalone product. This means that anyone who has this parameter set available in a comparable situation does not have to incur the effort for training again.

[0050] According to the above explanation, the invention also relates to a device for carrying out one of the methods described above. This device comprises a plurality of interference suppression modules, each of which receives a recording of a measurement signal as input and is configured to output versions of the recording at least partially corrected for the effect of specific interference. The device further comprises at least one trainable module, which also receives the recording of the measurement signal as input and controls the extent to which the overall interference suppression result provided by the device is composed of the versions provided by the individual interference suppression modules.

[0051] The device may, in particular, further comprise a processing module configured to further process the interference suppression result into a working result. This processing module may, in particular, be configured as a classifier and / or regressor.

[0052] The methods can be fully or partially computer-implemented, for example, in the form of software. Therefore, the invention also relates to a computer program with machine-readable instructions that, when executed on one or more computers, cause the computer(s) to execute one of the described methods. In this sense, control units for vehicles and embedded systems for technical devices that are also capable of executing machine-readable instructions are also considered computers.

[0053] The invention also relates to a machine-readable data carrier and / or to a downloadable product containing the computer program. A downloadable product is a digital product that can be transmitted over a data network, i.e., downloaded by a user of the data network, and which can be offered for immediate download, for example, in an online shop.

[0054] Furthermore, a computer can be equipped with the computer program, the machine-readable data carrier or the download product.

[0055] Further measures improving the invention are presented in more detail below together with the description of the preferred embodiments of the invention with reference to figures. Examples of implementation

[0056] It shows: Figure 1 Embodiment of the method 100 for training the trainable module 3; Figure 2Embodiment of the method 200 for processing recordings 11 of a measurement signal 10; Figure 3 Embodiment of a device 1; Figure 4 Detailed view of a trainable module 3 with neural network 30; Figure 5 Example removal of Gaussian noise ( Figure 5a ) and shot noise ( Figure 5b ).

[0057] Figure 1is a flowchart of an embodiment of the method 100. In step 110, a trainable module 3 is provided. In step 120, this trainable module 3 is supplied with learning recordings 11a of a measurement signal 10. These learning recordings 11a are subject to interference 11b. For each learning recording 11a, the trainable module 3 selects from a predetermined catalog 2 of interference suppression modules 2a-2d those interference suppression modules 2a-2d that it considers most suitable for removing the interference 11b from the learning recording 11a. According to the invention, the selected interference suppression modules 2a-2d are assigned weightings 5a-5d with which they are to operate.

[0058] In step 130, the selected noise suppression module(s) 2a-2b are applied to the learning recordings 11a. This results in a noise suppression result 12.

[0059] Optionally, in step 180, it is checked whether the interference suppression result 12 has converged according to a predefined criterion. If this is not the case (truth value 0), the interference suppression result 12 is fed back to the trainable module 3 in step 190 as a recording 11a of the measurement signal 10 containing a disturbance 11b. If, however, the interference suppression result 12 has converged (truth value 1), it is final and is further evaluated.

[0060] In step 140, an evaluation 14 of the interference suppression result 12 is determined using a cost function 13. In step 150, the interference suppression result 12 is processed by a processing module 4 into a work result 15.

[0061] Example components of the cost function 13 are broken down in more detail within box 13. According to block 13a, the denoising result 12 can be compared with a known learning denoising result 12a. According to block 13b, the denoising result 12 can be automatically evaluated based on a predetermined criterion 18. According to block 13c, an evaluation of the denoising result 12 obtained from at least one external evaluator 19, such as an expert or via crowdsourcing, can be used. According to block 13d, the work result 15 formed in step 150, such as a classification, can be compared with a known learning work result 15a.

[0062] The evaluation 14 determined with the cost function 13 is used in step 160 to optimize the parameter set 33, which at least characterizes the behavior of the trainable module 3. According to block 161, for example, an intermediate result 16 taken from the trainable module 3, which characterizes the disturbance 11b, can be evaluated with a pre-training cost function 17, and a portion of the parameters 33 can be optimized in advance according to block 162 with the aim of improving the value of this pre-training cost function 17. The pre-training cost function 17 can, for example, according to block 17a, contain a comparison between the intermediate result 16 and known additional information 16a about the disturbance 11b.

[0063] Figure 2is a flowchart of an embodiment of the method 200 for processing recordings 11. In step 210, a trainable module 3 trained with the previously described method 100 is provided, wherein the trainable module 3 can first be trained for this purpose, but can also be obtained from any source in the fully trained state.

[0064] In step 220, at least one recording 11 of the measurement signal 10 is fed to the trainable module 3, so that the latter selects one or more interference suppression modules 2a-2d for application to the recording 11, with weightings 5a-5d. The interference suppression modules 12a-12d thus selected are applied to the recording 11 in step 230.

[0065] Optionally, in step 280, it is checked whether the interference suppression result 12 has converged according to a predefined criterion. If this is not the case (truth value 0), the interference suppression result 12 is fed back to the trainable module 3 in step 290 as a recording 11 of the measurement signal 10 containing a disturbance 11b. If, however, the interference suppression result 12 has converged (truth value 1), it is final and is further evaluated.

[0066] In step 240, the interference suppression result 12 is further processed by a predetermined processing module 4, such as a classifier, to produce a working result 15.

[0067] In step 250, a control signal 20 is formed from the interference suppression result 12 and / or from the work result 15. In step 260, this is used to control a vehicle 50, and / or a classification system 60, and / or a system 70 for quality control of mass-produced products, and / or a system 80 for medical imaging.

[0068] Figure 3shows an exemplary embodiment of a device 1 for carrying out the previously described methods 100, 200. The catalog 2 of interference suppression modules here comprises four interference suppression modules 2a-2d, which each generate versions 2a*-2d* from the recording 11, 11a affected by the interference 11b. The trainable module 3 also receives the recording 11, 11a as input and determines weightings 5a-5d, with which the versions 2a*-2d* are calculated by a weighting unit 5 to produce an interference suppression result 12. Furthermore, a processing module 4 is provided to further process the interference suppression result 12 into a work result 15.

[0069] The first noise reduction module 2a contains submodules 2a1-2a3 as well as a further trainable module 3, which decides how the submodules 2a1-2a3 each affect the version delivered by the noise reduction module 2a as a whole. This can be done, analogous to the interaction of the noise reduction modules 2a-2d, for example, with an additional weighting unit 5.

[0070] Figure 4shows a detailed view of the trainable module 3, which here contains a neural network 30. The neural network 30 contains a first sequence 31 of three convolutional layers 31a-31c, which receives the recording 11, 11a as input and primarily characterizes the noise 11b contained in the recording 11, 11a. Adjacent to this is a second sequence 32 of two fully connected layers 32a, 32b, which determines the selection of noise reduction modules 2a-2d with weights 5a-5d. The intermediate result 16 is extracted from the penultimate convolutional layer 31b for use in the pre-training cost function 17.

[0071] Figure 5 shows schematically how different disturbances 11b in recordings 11 can be eliminated by applying different interference suppression modules 2a-2d.

[0072] Figure 5ashows a first example in which the image 11 contains Gaussian noise as interference 11b. Here, a noise reduction module 2a is activated, which blurs the image.

[0073] Figure 5b shows a second example in which the image 11 contains shot noise as interference 11b. Here, a first noise reduction module 2c is first activated, which "morphologically opens" the image 11. The result 12 is then recursively passed to a second noise reduction module 2d, where it is "morphologically closed." Both operations together yield a final result 12' that is free of the shot noise and essentially still retains all the details.

Claims

1. Computer-implemented method (100) for training a trainable module (3) for preparing recordings (11) of a measurement signal (10) that contain interference (11b), comprising the following steps: • a trainable module (3) is provided (110) which is designed so as, in the trained state, to select, from a predetermined catalogue (2) of interference suppression modules (2a-2d), a combination of interference suppression modules (2a-2d) for a recording (11) of the measurement signal (10) whose application to the recording (11) is expected to improve suppression of the interference (11b) contained in the recording (11); • the trainable module (3) is used to select a combination of interference suppression modules (2a-2d) for at least one learning recording (11a) of the measurement signal (10) (120); • the selected combination is applied to the learning recording (11a) (130) to obtain an interference suppression result (12); • a predetermined cost function (13) is used to determine an evaluation (14) (140) of the extent to which the interference (11b) is suppressed in the interference suppression result (12); • a parameter set (33) that at least characterizes the behaviour of the trainable module (3) is optimized (160) with the aim of, when the same learning recording (11a), and / or further learning recordings (11a), is / are processed again, improving the evaluation (14) of the interference suppression result (12) that is then obtained by way of the cost function (13); wherein the trainable module (3) supplies a weighting (5a-5d), which can be used to combine the interference suppression results (12) supplied by various interference suppression modules (2a-2d), as a selection of a combination of interference suppression modules (2a-2d).

2. Method (100) according to Claim 1, wherein the parameter set (33) additionally characterizes the behaviour of at least one interference suppression module (2a-2d).

3. Method (100) according to either of Claims 1 and 2, wherein the cost function (13) depends on the result of a comparison (13a) between the interference suppression result (12) and a known learning interference suppression result (12a), and / or on an automated evaluation (13b) of the interference suppression result (12) on the basis of at least one predetermined criterion (18), and / or on an evaluation (13c) of the interference suppression result (12) that has been received from at least one external evaluator (19).

4. Method (100) according to one of Claims 1 to 3, wherein • the interference suppression result (12) is processed by a predetermined processing module (4) to produce a work result (15) (150) and • the cost function (13) depends on the result of a comparison (13d) between the work result (15) and a known learning work result (15a).

5. Method (100) according to Claim 4, wherein the processing module (4) is in the form of a classifier and / or regressor that supplies a classification and / or regression of the interference suppression result (12) as the work result (15).

6. Method (100) according to one of Claims 1 to 5, wherein a trainable module (3) is selected which contains an artificial neural network (30), this artificial neural network (30) comprising a first sequence (31) of between 2 and 7 convolution layers (31a-31c) and a second sequence (32) of between 2 and 4 fully cross-linked layers (32a-32b).

7. Method (100) according to Claim 6, wherein an intermediate result (16) supplied by at least one of the convolution layers (31a-31c) as an output is evaluated using a pre-training cost function (17) (161) and wherein parameters (33) that characterize the behaviour of the first sequence (31) of the convolution layers (31a-31c) are optimized in advance (162) with the aim of improving the value of this pre-training cost function (17).

8. Method (100) according to Claim 7, wherein the pre-training cost function (17) depends on the result of a comparison (17a) between the intermediate result (16) and known additional information (16a) about the interference (11b).

9. Computer-implemented method (200) for preparing recordings (11) of a measurement signal (10) that contain interference (11b), comprising the following steps: • at least one trainable module (3) trained using the method (100) according to one of claims 1 to 8 is provided (210); • at least one recording (11) of the measurement signal (10) is routed to the trainable module (3) (220) so that the trainable module (3) selects a combination of interference suppression modules (2a-2d) from the predetermined catalogue (2) of interference suppression modules (2a-2d) by supplying a weighting (5a-5d) that can be used to combine the interference suppression results (2a*-2d*) supplied by the various interference suppression modules; • the selected combination is applied to the recording (11) (230) so that an interference suppression result (12) is obtained.

10. Method (200) according to Claim 9, wherein the interference suppression result (12) is further-processed by a predetermined processing module (4) to produce a work result (15) (240).

11. Method (200) according to Claim 10, wherein the processing module (4) is in the form of a classifier and / or regressor that supplies a classification and / or regression of the interference suppression result (12) as the work result (15).

12. Method (200) according to one of Claims 9 to 11, wherein the interference suppression result (12), and / or the work result (15), are / is used to generate a control signal (20) (250) and wherein a vehicle (50), and / or a classification system (60), and / or a system (70) for quality control of mass-produced products, and / or a system (80) for medical imaging, are / is controlled using this control signal (20) (260).

13. Method (100, 200) according to one of Claims 1 to 12, wherein the interference suppression result (12) is routed to the trainable module (3) once more as a recording (11, 11a) of the measurement signal (10) that contains the interference (11b) (190, 290).

14. Method (100, 200) according to Claim 13, wherein the returning (190, 290) is repeated until the interference suppression result (12) converges according to a predetermined criterion (180, 280).

15. Method (100, 200) according to one of Claims 1 to 14, wherein at least one interference suppression module (2a-2d) contains multiple submodules (2a1-2a3) and a further trainable module (3a) that controls the extent to which the effect of this interference suppression module (2a-2d) is compiled from the effects of one or more submodules (2a1-2a3).

16. Method (100, 200) according to one of Claims 1 to 15, wherein the recording (11) of the measurement signal (10) includes an image representing a spatial distribution of at least one physical measured variable, and / or a time series representing a time characteristic of at least one physical measured variable.

17. Computer program containing machine-readable instructions that, when executed on one or more computers, cause the one or more computers to carry out a method (100, 200) according to one of Claims 1 to 16.

18. Machine-readable data carrier and / or download product having the computer program according to Claim 17.

19. Computer equipped with the computer program according to Claim 17 and / or with the machine-readable data carrier and / or download product according to Claim 18.