Generation of training data for two-dimensional scans of a ground-based radar system

DE502022002087D1Active Publication Date: 2024-11-28FRAUNHOFER GESELLSCHAFT ZUR FORDERUNG DER ANGEWANDTEN FORSCHUNG EV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
DE502022002087
Authority / Receiving Office
DE · DE
Patent Type
Patents
Current Assignee / Owner
Priority Date
2021-01-29
Filing Date
2022-01-26
Publication Date
2024-11-28
Estimated Expiration
2042-01-26

AI Technical Summary

Technical Problem

Current methods for generating training data for ground penetrating radar (GPR) systems are limited by the lack of available data sets for convolutional neural networks (CNNs), leading to poorer classification results due to mismatched dimensions, gray value spectra, and object signatures, and require extensive measurement campaigns or transfer learning strategies.

Method used

A device and method that extract interfering objects from source GPR images using quotient images and a reference image, allowing for the generation of result GPR images through manipulation and modification, thereby creating a large amount of training data with minimal effort.

Benefits of technology

This approach enables the creation of robust training data with scalable and efficient generation of GPR images, reducing the effort required for data acquisition and improving classification accuracy in machine learning algorithms.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader
Need to check novelty before this filing date? Find Prior Art

Description

Technical area

[0001] Embodiments according to the invention relate to devices and methods for generating training data for two-dimensional scans of a ground-penetrating radar system. Furthermore, embodiments address image processing techniques for creating training data for two-dimensional scans of a ground-penetrating radar (GPR) system, which can be used to train machine learning (ML) algorithms. Further embodiments relate to automatic image processing techniques for creating and expanding training data for machine learning methods for ground-penetrating radar images. Background of the invention

[0002] ML methods have enjoyed steadily increasing attention in various application areas in recent years. Convolutional neural networks (CNNs), a special form of neural networks (NNs), are particularly popular in the field of digital image processing. In order for such networks to calculate robust analysis results, very large datasets are generally required for training. Available networks are used to analyze image data captured with electro-optical sensors. For such input data, a multitude of different training datasets exist, which can be used, for example, for classification tasks. For GPR data, on the other hand, no such training data exists to date, according to current knowledge. As a solution, some people therefore resort to transfer learning (TL) strategies, which operate with smaller databases.In the case of a new construction of a network without recourse to TL methods, one is therefore dependent on acquiring suitable training data through complex measurement campaigns. State of the art

[0003] For the processing analysis of GPR image data, there are very few, if any, publicly available datasets that can be used to train CNNs. As a result, methods such as TL are often used to adapt pre-trained networks to the respective GPR dataset. This requires a relatively small amount of GPR data. Key problems with this approach are that the data underlying the pre-trained network do not match the GPR data in terms of dimension, grayscale, and object signature, and must be adapted accordingly. This, in turn, often leads to poorer classification results.

[0004] The alternative approach, the creation and testing of GPR-specific network structures from scratch, is again only feasible without large amounts of training data and at the same time results in a loss of classification quality.

[0005] The following approaches are also known from the state of the art.

[0006] In "Some good practices for applying convolutional neural networks to buried threat detection in Ground Penetrating Radar," (2017 9th International Workshop on Advanced Ground Penetrating Radar (IWAGPR), 2017, pp. 1-5, doi: 10.1109 / IWAGPR.2017.7996100), D. Reichman, L.M. Collins, and J.M. Malof address the application of convolutional neural networks to GPR data. In particular, the authors discuss a pretraining initialization step and a dataset augmentation protocol.

[0007] Besaw, Lance E., in "Detecting buried explosive hazards with handheld GPR and deep learning." (Detection and Sensing of Mines, Explosive Objects, and Obscured Targets XXI. Vol. 9823. SPIE, 2016), discusses data augmentation techniques for GPR training data.

[0008] H. Jo, Y.-H. Na, and J.-B. Song, in "Data augmentation using synthesized images for object detection," (2017 17th International Conference on Control, Automation and Systems (ICCAS), 2017, pp. 1035-1038, doi: 10.23919 / ICCAS.2017.8204369), discuss methods for generating a dataset by synthesizing images of the background and object. Different images can be generated through post-processing, such as adding noise and changing the brightness of images of objects obtained from different angles.

[0009] For a better understanding of the background of the invention, reference is further made to the documents cited in the above-mentioned documents.

[0010] In view of this, there is a need for a concept that enables the creation of a large amount of training data based on a small number of scans (e.g., GPR images) from a ground-penetrating radar system or based on a small number of GPR images, so that the training data can be used, for example, to train machine learning algorithms. Furthermore, there is a need for an alternative approach to reduce the effort associated with acquiring suitable training data, e.g., for reconstructing a neural network.

[0011] The object of the present invention is therefore to provide a device and a method as well as a computer program product which make it possible to provide a large amount of training data based on a small number of scans of a ground-penetrating radar system and / or to provide training data with little effort.

[0012] This problem is solved by the independent patent claims. Further developments according to the invention are defined in the subclaims. Summary of the invention

[0013] Embodiments according to the present invention provide a device for generating training data for two-dimensional scans of a ground-penetrating radar system, wherein the device comprises a clutter extractor configured to determine a quotient image from a source GPR image using a reference image and to extract at least one clutter based on the quotient image. Furthermore, the device comprises a result image generator configured to generate at least one result GPR image for the source GPR image based on the at least one extracted clutter object in order to generate the training data.

[0014] Further embodiments according to the present invention provide a method for generating training data for two-dimensional scans of a ground-penetrating radar system, the method comprising determining a quotient image from a source GPR image with a reference image, extracting at least one clutter object from the source GPR image based on the quotient image, and generating at least one result GPR image for the source GPR image, depending on the at least one extracted clutter object, to generate the training data.

[0015] Further embodiments according to the present invention provide a computer program with a program code for carrying out the inventive method according to embodiments of the present disclosure.

[0016] Embodiments according to the present invention are based on the core idea that, for example, a large amount of training data can be generated based on, for example, a small or limited number of source GPR images by extracting noise objects from the source GPR images using quotient images and subsequently manipulating or modifying GPR images in order to generate at least one result GPR image using the extracted noise objects. From this, a low-effort basis for potentially multiple use of the noise object in result GPR images can be created by simply calculating a quotient image, so that training data can be created with little effort and on a scalable scale.

[0017] For the extraction of interfering objects, a device according to the invention has an interfering object extractor which can determine a quotient image from a source GPR image together with a reference image. The reference image can, for example, preferably be a GPR image which does not have an interfering object. The reference image can, for example, be an average image of a large number of interference-free GPR images. It should be noted that the freedom from interference or the absence of an interfering object refers to interference or artifacts or elements which are to be detected or classified, for example, during training using training data generated according to the invention. Interference-free in no way necessarily means the absence of influences such as noise.A clutter object can, for example, be an object to be detected beneath a ground surface that was scanned during the generation of the source GPR images. Using ratio image determination, the clutter object extractor can, in simple terms, detect and extract a clutter object by comparing the source GPR image with the reference image.

[0018] In order to generate, for example, a large amount of training data from, for example, a small amount of source GPR images and associated noise objects, the device according to the invention has a result image generator by means of which, for example, the extracted noise objects can be manipulated and / or modified in order to generate, for example, a large amount of result GPR images. In this case, a noise object can, for example, be inserted into a GPR image which previously had no noise object, or which contained one or more other noise objects. Furthermore, the extracted noise objects can also be modified and likewise inserted into other images or into their original source GPR image or into the reference image. According to the invention, a large number of variations of the manipulations and / or changes and / or modifications can be used to generate the result GPR images.

[0019] The resulting GPR images can therefore be used to generate training data, for example, by forming the training data themselves. By modifying the source GPR images with regard to the interfering objects, a data basis can be created on which algorithms, such as machine learning algorithms, can be trained. The multitude of possible variations with regard to the interfering object can thus enable a robust learning process and / or a learning process that leads to a robust (e.g., classification) algorithm.

[0020] It should be noted, however, that the device according to the invention can be used without any problems in the method if an undisturbed GPR image is available, since, for example, no interference object extraction is performed from such an image in the absence of an interference object. Such an undisturbed image can then, however, be used as the basis for a final image, for example, by integrating an interference object extracted from another image into it.

[0021] In further embodiments according to the present invention, the result image generator is designed to insert the at least one interference object into a second GPR image to generate the at least one result GPR image.

[0022] The second GPR image can, for example, be another source GPR image that, for example, did not exhibit any interference in the previous image. Furthermore, the second GPR image can also be a GPR image that also exhibits interference, from which the interference object was extracted. Simply put, this allows interference objects to be exchanged between GPR images, which can lead to a rapid increase in the amount of training data.

[0023] In further embodiments according to the present invention, the second GPR image is the reference image. The reference image is, for example, preferably an undisturbed GPR image, which can be formed, for example, from an averaging or superimposition of a plurality of undisturbed source GPR images. Thus, it can form a type of generic undisturbed image. By inserting interference objects into such an image, a good data basis for detecting the interference objects can be created, for example with regard to a learning algorithm, e.g. based on the averaged and thus, for example, representative or characteristic background of the image, wherein the background can designate the undisturbed image area, i.e. the area without an interference object.

[0024] In further embodiments according to the present invention, the result image generator is configured to change the at least one interfering object and / or a position of the at least one interfering object in the second GPR image compared to the source GPR image in order to generate the at least one result GPR image.

[0025] By changing the at least one interfering object and / or the position of the at least one interfering object, for example, an almost arbitrarily large number of resulting GPR images can be generated from a source GPR image simply by extracting a single interfering object. This is due to the almost inexhaustible possibilities for changing the position of an interfering object in an image, or, with regard to a large number of extracted interfering objects and source GPR images, an additionally executable permutation of interfering objects and source GPR image combinations, for example. Accordingly, a very large amount of training data can be generated from, for example, a small basic set of source GPR images containing interfering objects.

[0026] In further embodiments according to the present invention, the result image generator is configured to change a scaling of the at least one interference object in the second GPR image compared to the source GPR image in order to generate the at least one result GPR image.

[0027] In addition to increasing the training dataset, scaling the clutter object can also improve the quality of the training data, as a large number of modified clutter objects can be used to generate the resulting GPR images. Training on data with scaled clutter objects allows for a particularly robust algorithm to be trained.

[0028] In further embodiments according to the present invention, the result image generator is configured to change a gray value intensity of the at least one interference object in the second GPR image compared to the source GPR image in order to generate the at least one result GPR image.

[0029] This allows the training data to be further improved, for example, with regard to training robust classifiers, for example, from the field of machine learning. In particular, by changing the gray values, changes in the recording conditions and / or recording hardware, for example, for scanning a ground surface and generating the source GPR images, can be taken into account in the training data.

[0030] In further embodiments according to the present invention, the source GPR image is a two-dimensional GPR image. By using two-dimensional GPR images, two-dimensional interfering objects can be extracted and, accordingly, two-dimensional result GPR images can also be generated.

[0031] In further embodiments according to the present invention, the source GPR image is a GPR image of a plurality of first GPR images. This allows interfering objects to be combined and / or allows filtering of individual image properties to be utilized by combining the first GPR images.

[0032] In further embodiments according to the present invention, the plurality of first GPR images are a plurality of temporally consecutive GPR images generated from a sequence of a plurality of recorded GPR images.

[0033] In further embodiments according to the present invention, the apparatus comprises a GPR data acquisition system for recording the sequence of the plurality of acquired GPR images.

[0034] By integrating a GPR system for data acquisition, a complete system can be created that can be deployed in the field, for example, in a specific area of ​​the ground surface that is to be scanned. This allows, for example, a data basis in the form of training data to be created directly on site, with which an algorithm can be trained.

[0035] In further embodiments according to the present invention, the GPR system for data acquisition is a multi-channel ground-penetrating radar, and the device is configured to use the multi-channel ground-penetrating radar to acquire the sequence of the plurality of acquired GPR images. By using a multi-channel ground-penetrating radar, a ground surface to be scanned can, for example, be scanned more quickly, thanks to the plurality of channels that can generate information in parallel.

[0036] In further embodiments according to the present invention, the GPR system is a single-channel system having parallel measuring paths. The device is configured to use the single-channel system to record the sequence of the majority of recorded GPR images. For example, to provide a more cost-effective version of the GPR system, the GPR system can be configured as a single-channel system. The parallel arrangement of the measuring paths can enable area-wide scanning.

[0037] In further embodiments according to the present invention, the device is designed to compensate for a distance between the sensor and the ground surface which varies in the sequence of the plurality of recorded GPR images in order to generate the plurality of first GPR images. Unevenness in the ground surface, which can lead to a varying distance between the sensor used to record the GPR images and the ground surface, can make subsequent processing of the recorded GPR images more difficult. It can therefore be advantageous to compensate for such a varying distance, wherein the compensation can be understood as a type of normalization with regard to a distance to the ground surface. Corresponding distance information can, for example, be obtained from an evaluation of a recorded GPR image. For example, the ground surface, orthe position of the soil signature can be detected as a characteristic feature in a GPR image so that the image can be adjusted accordingly.

[0038] In further embodiments according to the present invention, the device is designed to normalize a ground signature that varies in the sequence of the plurality of recorded GPR images in order to obtain a normalized ground signature, and to generate the plurality of first GPR images using the normalized ground signature. The ground signature can, for example, be radiation properties inherent in the ground, for example of ground layers. For example, depending on the existing ground conditions, different returning radar signals can occur. This influence can make detection of interfering objects more difficult, for example for subsequent extraction. And to reduce this influence, a normalized ground signature can be determined with which the majority of first GPR images are generated.This can reduce the susceptibility to errors in subsequent noise object extraction or, in other words, make it more robust.

[0039] In further embodiments according to the present invention, the device comprises a zero-offset corrector configured to perform a zero-offset correction with respect to the sequence of the plurality of acquired GPR images in order to generate the plurality of first GPR images.

[0040] By integrating a zero-offset corrector into the device, a system can be created that is particularly advantageous for mobile applications, for example, for generating training data. For example, the recorded GPR images can be corrected directly in the field, thus saving storage space by saving only processed GPR images and deleting raw data immediately after processing. The zero-offset correction of the zero-offset corrector can, for example, include the previously explained compensation for a varying distance between the sensor and the ground surface and / or normalization using a varying ground signature.

[0041] In further embodiments according to the present invention, the reference image is an average image from a plurality of interference-free GPR images of the sequence of the majority of recorded GPR images. By generating the reference image from an averaging of a plurality of interference-free GPR images of the sequence, a reference adapted to the respective sequence can be generated. By using the GPR images of the sequence, features of the soil composition or properties of the soil composition that influence the GPR images can be directly taken into account. Such generation of a reference image can, for example, be interpreted as adaptive reference generation, for example adaptive with regard to changing environmental information in the GPR images, which is associated, for example, with soil properties.

[0042] In further embodiments according to the present invention, the noise object extractor has a quotient image determiner for determining the quotient image, and the noise object extractor is designed to perform change detection based on filtering and / or a threshold comparison to extract the noise object. Since the quotient image contains differences between the corresponding source GPR image and the reference image, it is advantageous to analyze corresponding differences with regard to the presence of a noise object and / or to mitigate unwanted influences that can lead to differences in the quotient image but are independent of the noise object. In order to avoid erroneously extracting differences, which are caused, for example, by image noise, as noise objects, the noise object extractor can, according to the invention, perform filtering and / or a threshold comparison.This allows interfering objects to be detected and extracted with reduced susceptibility to errors.

[0043] In further embodiments according to the present invention, the noise object extractor comprises a filter, and the filter is configured to filter the quotient image to suppress noise pixels. This can reduce the error susceptibility of the noise object extraction, as described above.

[0044] In further embodiments according to the present invention, the filter is a structure-preserving, morphological filter, and the filter is configured to delete or retain connected components of the quotient image depending on their size. By selecting connected components of the quotient image depending on their size, the quotient image can be adapted, for example, based on the expected size of interfering objects. Accordingly, only those elements in the quotient image that meet a certain minimum size can be considered for extraction. This, in turn, reduces the susceptibility to errors in interfering object extraction.

[0045] In general, embodiments according to the present invention can comprise structure-preserving filters. A structure-preserving property can be understood as such a morphological, alternating sequential filter (ASF) resulting in no or only negligible distortion of the image content compared, for example, to various box filters (mean, etc.). "Distortion" can, for example, include the addition of filter-specific artifacts (e.g., "smearing" in mean filters) or be understood as the addition of such filter-specific artifacts.

[0046] An ASF, in its conventional form, can be described, for example, in [Soille, 1998], as particularly suitable for processing closely spaced structures of different brightness and size, or can be designed for this purpose. The inventors have recognized that such a filter can also be well suited for noise reduction in radar images.

[0047] In further embodiments according to the present invention, the clutter object extractor comprises a threshold comparator configured to subject the quotient image to a threshold comparison to determine a binary change map.

[0048] In the case of quotient images which were created from two images, e.g. source GRP image and reference image, with large deviations from each other and where at least part of the deviations are not caused by an interfering object present in one of the images, the threshold comparator can help to prevent this part of the deviations which are not caused by an interfering object (but for example by a changed background, e.g. a changed soil layer) from being falsely identified as an interfering object.

[0049] In the case of quotient images, which were generated from two images, e.g., a source GRP image and a reference image, with small or very slight deviations from each other, and where at least some of the deviations are not caused by an interfering object present in one of the images, the thresholding method itself may not be able to prevent, for example, the detection of interfering objects. This can then essentially be a "merit" of the two images or their contents and properties, or, in other words, caused by the contents and properties of the two images. In other words, the problem of false detection cannot occur with small or very slight deviations according to the invention and therefore does not necessarily have to be addressed with the threshold comparator.

[0050] For example, a minimum threshold, for example for the gray values, can be determined from empirical values, which can be met in a quotient image for an expected interference object. However, according to exemplary embodiments, the automated method according to the invention can preferably be used instead of determining the minimum threshold from empirical values, since the gray values ​​or brightnesses of the interference objects can vary greatly from one another. Therefore, it may even be advantageous, for example, to avoid manual determination. This can further improve the robustness of the interference object extraction.Furthermore, binarization for generating the binary change map can significantly reduce the required storage space and, accordingly, the computing power required for further processing the binary change map. This is because a large portion of the source information, for example, for extracting the interference object, can be discarded. The position of the interference object can be displayed with binary-valued data points in the change map. Binarization, for example of the gray values, can be performed based on threshold comparison, for example, depending on whether the threshold is at least met or undershot.

[0051] In further embodiments according to the present invention, the clutter object extractor comprises a pixel reducer configured to perform a reduction of change pixels of the binary change map, wherein binary connected components of the binary change map smaller than a number of pixels associated with a minimum size of the clutter object are deleted. Thus, the extraction of clutter objects that are, for example, uninteresting for a corresponding classification task (e.g., for the classification of large underground pipelines, although nails or screws can also be identified as clutter objects) and / or the extraction of incorrectly detected clutter objects, for example, due to recording disturbances in the GPR images, can be prevented or at least made less likely.

[0052] In further embodiments according to the present invention, the clutter extractor is configured to extract the at least one clutter object from the source GPR image based on the binary change map. By using the binary change map, the required storage space and computing power for handling the image information can be reduced due to the reduced number of pixels and their binary-valued information. Furthermore, the robustness of the clutter extraction can be improved due to the previously explained properties of the binary change map.

[0053] In further embodiments according to the present invention, the result image generator is designed to insert the at least one extracted interference object into the second GPR image by averaging the source GPR image and the second GPR image for each pixel of the binary change map to generate the at least one result GPR image. Averaging provides a simple and low-computational-intensive way to create the result GPR image from the second GPR image and the extracted interference object. By averaging the result GPR image for each pixel of the binary change map, even further computational effort can be saved, since an average only needs to be calculated for the relevant pixels.Furthermore, averaging avoids an unnatural transition between the inserted interfering object and the second GPR image, which can be advantageous, for example, for training an algorithm, for example, with regard to replicating real GPR images with interfering objects. For example, averaging can be used to simulate or reproduce the effect of altered reflection properties of the ground surrounding an interfering object at the interface between the ground and the interfering object, e.g., because the ground around the interfering object is denser or looser. Short character description

[0054] Embodiments according to the present invention are explained in more detail below with reference to the accompanying figures. With regard to the schematic figures shown, it should be noted that the illustrated functional blocks are to be understood both as elements or features of the device according to the invention and as corresponding method steps of the method according to the invention, and corresponding method steps of the method according to the invention can also be derived therefrom. They show: Fig. 1 shows a schematic block diagram of a device according to an embodiment of the present invention; Fig. 2 shows a schematic block diagram of a device with additional optional elements according to an embodiment of the present invention; Fig. 3 shows a schematic flow diagram of a method according to an embodiment of the present invention; Fig. 4 shows examples of (intermediate) result images that are generated in steps of the Fig. 3 explained method, to explain embodiments; Fig. 5Exemplary (intermediate) result images of the data augmentation for one of the Fig. 4illustrated images according to an embodiment; Fig. 6 is a flowchart of a method according to an embodiment of the present invention; and Fig. 7 is an example of an 8-neighborhood of a pixel for use in embodiments of the present invention; and Fig. 8 is an example of a 4-neighborhood of a pixel for use in embodiments of the present invention. Detailed description of the embodiments according to the figures

[0055] Before exemplary embodiments of the present invention are explained in more detail below with reference to the drawings, it is pointed out that identical, functionally equivalent or equivalent elements, objects and / or structures in the different figures are provided with the same or similar reference numerals, so that the description of these elements shown in different exemplary embodiments is interchangeable or can be applied to one another.

[0056] Fig. 1 shows a schematic block diagram of a device 100 according to an embodiment of the present invention. The device 100 has a noise object extractor 110 and a result image generator 120. The noise object extractor 110 receives a source GPR image 110a and a reference image 110b. Based on the source GPR image 110a and the reference image 110b, the noise object extractor 110 determines a quotient image and extracts at least one noise object 110c based on the quotient image. The result image generator 120 receives the at least one noise object 110c and generates at least one result GPR image 120a based on the at least one extracted noise object 110c.

[0057] The source GPR image 110a can, for example, be a GPR image from a plurality of temporally successive GPR images that was generated from a sequence of a plurality of recorded GPR images. The source GPR image 110a can, for example, in particular be a two-dimensional GPR image. The reference image 110b can be an average image from a plurality of interference-free GPR images, for example from the sequence of the majority of recorded GPR images. Furthermore, the reference image 110b can also be a single, for example preferably undisturbed, GPR image. To extract the interference object 110c, the interference object extractor 110 is designed to generate a quotient image that can have differences between the reference image 110b and the source GPR image 110a. For example,If the reference image 110b does not contain the same interference object as the source GPR image 110a, the interference object present in the source GPR image 110a can be displayed in the quotient image as a difference between the source GPR image 110a and the reference image 110b and thus also extracted. The extracted interference object 110c can then be used by the result image generator 120, e.g., by inserting it into the reference image 110b, or another GPR image (e.g., a second GPR image), e.g., additionally modifying the interference object, e.g., with respect to a position in the image, the associated gray values, and / or a scaling, to generate the result GPR image.

[0058] Fig. 2shows a schematic block diagram of a device 200 according to an embodiment. Device 200 has additions compared to device 100. It should be noted that these additions are optional, both individually and in combination, even within previously described elements, such as the interference object extractor. It should also be noted that the respective optional elements are to be regarded as optional features, both individually and in any combination.

[0059] The device 200 has a GPR system 210 for data acquisition. The GPR system 210 provides a sequence 210a of a plurality of recorded GPR images. The sequence 210a can be a plurality of temporally consecutive GPR images, such as can be acquired during scanning by a movable GPR system 210. The sequence 210a includes a plurality of source GPR images, which can be, for example, two-dimensional GPR images.

[0060] The GPR system 210 can further be a multi-channel ground-penetrating radar or a single-channel ground-penetrating radar having parallel measurement paths. Furthermore, the device 200 comprises a zero-offset corrector 220, which receives the sequence 210a from the GPR system 210. The zero-offset corrector 220 is configured to perform a zero-offset correction of the sequence 210a. The correction can, for example, comprise compensating for a varying distance between a sensor used to record the sequence 210a and a ground surface scanned by the sensor. Furthermore, the zero-offset corrector 220 can, for example during the zero-offset correction, normalize the sequence 210a with respect to a varying ground signature. This can, for example, reduce an undesirable influence of varying ground conditions on the extraction of interfering objects.The zero-offset corrector 220 thus provides a corrected source GPR image 110a or a sequence of a plurality of corrected acquired GPR images for the clutter object extractor 110 of the device 200.

[0061] The noise object extractor 110 has a quotient image determiner 230, which receives the source GPR image 110a. The quotient image determiner 230 determines a quotient image 230a based on the source GPR image 110a and a reference image 110b. The reference image 110b can, for example, be an average image from a plurality of noise-free GPR images of the sequence of the majority of the recorded GPR images. The GPR images used for averaging can, in particular, be zero-offset corrected. However, the reference image can also be a single undisturbed GPR image. It should be noted again that undisturbed with regard to the reference image describes an absence of noise objects, so that the image can still be affected by other undesirable, process-related effects, such as image noise.

[0062] The noise object extractor 110 further comprises a filter 240, which receives the quotient image 230a. The filter 240 can reduce and / or suppress previously discussed undesirable effects, such as image noise in the form of noise pixels. Furthermore, the filter can be a structure-preserving, morphological filter, which deletes or preserves connected components of the quotient image depending on their size. This allows a preselection to be made, whereby the quotient image, which, simply put, indicates a difference between a corresponding source GPR image 110a and the reference image 110b, can be freed of connected components that are smaller than the connected components of an expected noise object. This can improve the robustness and quality of the noise object extraction of the noise object extractor 110.

[0063] For better understanding, a short, non-limiting example will be given. The filter, e.g. an ASF, can proceed at this point in such a way that the quotient image has a certain number of grayscale levels, e.g., 256, although other numbers are readily possible. The quotient image can be broken down into corresponding binary images, e.g., into a corresponding number of binary images. For each of these binary images, connected components can be determined, e.g., by examining the neighborhood of a pixel. The neighborhood can, for example, be a neighborhood of 2, 4, or 8, or any other neighborhood, or can be dimensioned differently. The filter can then check for each binary image which of the connected components are smaller or larger than the specified number of pixels.

[0064] To explain the neighborhood, Figures 7 and 8 referred to. Figures 7 and 8each show an arrangement of 9 pixels. Figures 7 and 8 each show a central pixel 710 and its neighborhood, which includes surrounding pixels. The case of a neighborhood of 8 is shown in Fig. 7 shown. All 8 surrounding pixels 720 are considered as neighbors of pixel 710, for example, to determine the connected components. Fig. 8 shows the case of a neighborhood of four. Only the four pixels 720, here as an example in horizontal and vertical neighborhood, are considered as neighbors of pixel 710, while the other pixels 810, here as an example diagonal to pixel 710, are not considered.

[0065] Furthermore, the noise object extractor 110 has a threshold comparator 250, which receives the filtered quotient image 240a. The threshold comparator 250 is configured to subject the quotient image 240a to a threshold comparison and can thus determine a binary change map 250a. To create a binary image from a grayscale image, a specific grayscale value as a threshold may be advantageous or even necessary. The filter for smoothing the quotient image can serve to reduce or even eliminate potential false detections. However, it may therefore, for example, by definition, not be suitable for providing a threshold for binarizing the quotient image, which is why, for example, the threshold comparator is used in exemplary embodiments.

[0066] However, misdetections cannot be avoided, for example, simply by applying a threshold value to binarize the quotient image. Such avoidance can be provided according to embodiments by subjecting the filtered quotient image to a threshold value comparison. Thus, for example, the filter can be "responsible" for this, i.e., for example, for avoiding misdetections, not necessarily the threshold value comparison itself. Based on the threshold value, e.g., a minimum gray value change, the threshold comparator 250 can, for example, make a binary decision, so that, for example, each gray value of a pixel of the filtered quotient image 240a can be reduced to a value from a two-element value range, wherein the respective value indicates whether the threshold value is met or not met.By means of the information extraction, which leads to the generation of the binary change map 250a, a strong data compression and / or data reduction with regard to the extraction of noise objects can be achieved, for example, in comparison to an original source GPR image, for example by omitting concrete gray values ​​and describing changes by a binary value.

[0067] As a further element, the clutter object extractor 110 can include a pixel reducer 260, which can receive the binary change map 250a. The pixel reducer 260 deletes binary connected components of the binary change map 250a that are associated with a minimum size of the clutter object. In other words, connected components that are, for example, smaller than connected components of an expected clutter object can be deleted. This allows the pixel reducer 260 to reduce change pixels of the binary change map 250a. Thus, the pixel reducer 260 can provide a reduced binary change map 260a. The change map 260a then contains, for example, only information related to the position of a clutter object to be extracted from the source GPR image 110a.In other words, the change map 260a can display the region of the source GPR image 110a to be extracted, which region includes the interference object, and / or, for example, can contain this region alone. It should be noted that, for ease of understanding, the previous explanations have primarily referred to one interference object; however, analogous processing steps and results can be performed or achieved for a multitude of interference objects in a source GPR image 110a. The creation of the reduced binary change map 260a can be understood as interference object extraction, since the change map 260a extracts the information about the interference object and thus the interference object from the source GPR image 110a. Furthermore, the interference object extractor 110 can also provide the source GPR image 110a in addition to the change map 260a, so that the extracted interference object is fully described.

[0068] The device 200 may comprise the result image generator 120. The result image generator 120 receives the source GPR image 110a and the reduced binary change map 260a from the clutter object extractor 110. These two input variables of the result image generator 120 may, for example, be assigned to the extracted clutter object 110c from Fig. 1correspond. Furthermore, the result image generator 120 receives a second GPR image 120b. Thus, to generate at least one result GPR image 120a, the result image generator 120 can insert the extracted noise object, for example in the form of the source GPR image 110a and the reduced binary change map 260a, into the second GPR image 120b to generate training data. The second GPR image 120b can, for example, be the reference image 110b or another image base into which the noise object is to be inserted to generate later training data. This can ensure, for example, that the noise object is inserted into a noise-free GPR image in order to obtain, for example, a specific form of training data that should contain only one noise object.To generate the result GPR image 120a, the result image generator 120 can, for example, change the extracted interference object, a position of the extracted interference object, a scaling of the extracted interference object, and / or a gray value intensity of the extracted interference object in comparison to the source GPR image 110a from which the interference object was extracted.

[0069] This allows a large amount of training data to be generated, for example, with specific properties for training an algorithm, wherein the properties can be adjusted by manipulating and / or modifying and / or altering (e.g., changing the position, scaling, changing the gray value) the extracted noise object before inserting it into the second GPR image 120b. For example, different classes of training data can be generated that can be used for training according to the respective algorithm, for example, from the field of machine learning. The resulting GPR image 120a can then be generated, for example, by averaging between the source GPR image 110a and the second GPR image 120b for each pixel of the reduced binary change map 260a.

[0070] The device made of Fig. 2or an associated method, can form an example of an alternative procedure according to the invention in order to significantly reduce the effort associated with the acquisition of suitable training data, e.g. for the new construction of a neural network without recourse to TL methods, e.g. with regard to previous methods of the state of the art.

[0071] Based on the Figures 3 , 4 and 5 Further embodiments according to the invention are described. Fig. 3 shows a schematic flow diagram of a method 300 according to an embodiment, which can be carried out in devices described herein and can also be understood as a workflow of such a device, which is why the Fig. 3also refers to physical components of a device or system. The flowchart shows the workflow of an exemplary embodiment. Parallelograms represent data, and rectangles represent work steps.

[0072] A GPR system 210 can be used for data acquisition and can be configured with single or multi-channel technology. The GPR sensor data acquired by the GPR system 210, for example, can be used to detect objects in the ground. The "pure" ground signal without object involvement can also be referred to as the image background in this context. If an object in the ground is detected and imaged, its signature overlays the image background as the image foreground or clutter.

[0073] To create or augment a training dataset from GPR image data, an image processing technique according to the invention can be used, which aims to separate the image foreground from the image background. One of the core aspects, or even the main focus of this separation, can be so-called clutter objects, which represent the image foreground in this description. Following the separation, new GPR images can be artificially generated by varying the clutter object, for example, with regard to location, scaling, and grayscale intensity (e.g., using data augmentation methods and techniques).

[0074] Data augmentation can use methods to increase the amount of training data. Artifacts or disturbances in the image data can be detected, extracted, and moved to different locations within the image. Embodiments of the invention can be considered a novel approach to data augmentation.

[0075] The workflow according to an embodiment is shown in Fig. 3 This is explained step by step below.

[0076] The data basis, for example, is at least one 310 (GPR sequence, i=[1;x] with running index i with a total number x of GPR images) of chronologically successive GPR image data. The two-dimensionality of the data can be achieved, for example, by using a multi-channel ground-penetrating radar that records an image with each measurement, with n image columns orthogonal to the direction of movement of the sensor, i.e., for example, the sensor of the ground-penetrating radar that records the image. Alternatively, two-dimensional GPR image data can also be created, for example, by parallel measuring paths of a single-channel system. The number m of image lines can represent the number of depth points determined by sensor-related parameters. By moving the sensor along a measuring path, a three-dimensional data cube of size m × n × p can be created from p GPR images.

[0077] Based on the image data sequence, a so-called zero-offset correction (correction of the zero crossing of the signal) is first performed. This serves the purpose of compensating for a distance between the sensor and the ground surface that varies throughout the sequence and normalizing a varying ground signature. This preprocessing step is particularly advantageous for subsequent processing steps that aim to robustly estimate the background of a GPR image. The zero-offset correction takes place, for example, before the start of further image processing. The abbreviation ZC is used below to indicate that corresponding GPR images are zero-offset corrected. The corrected image data sequence forms the basis for the extraction of noise objects.

[0078] To extract clutter objects as the image foreground, a change detection method is used that is widely used in the field of SAR image analysis (SAR = Synthetic Aperture Radar). This method is based on calculating a quotient image from an input image pair. For this purpose, the currently acquired image 330 (B-scan i, ZC [int16] zero-offset corrected B-scans with gray values ​​in a 16-digit integer range with index i, in other words the i-th image of the sequence 310) is used. Image 330 can, for example, be the source GPR image 110a from Fig. 1 and 2correspond) of the sequence and a mean background image 340 (MW B-Scan ZC [int16], mean value image in the form of a zero-offset corrected B-scan with gray values ​​in a 16-digit integer range, which results from an initial settling phase; i<=N (e.g.: N = 30), where N is the number of GPR images used to determine the background image or mean image) resulting from a so-called "transient phase" of the system.

[0079] When using a multi-channel sensor, a B-scan can be composed of the individual measurements from the corresponding channels recorded simultaneously at the receiver. These channel measurements can then form the measurements "pointing" vertically downward into the subsurface. For example, if a system with 15 channels is used, a B-scan consists of 15 columns. The rows are determined by the sensor's penetration depth into the subsurface.

[0080] The middle background image can be, for example, the reference image 110b from Fig. 2 be. In other words, the mean image can be the so-called reference image. It can be determined from the most interference-free images possible of the initial transient response phase, e.g. simply by averaging, i.e. summed image stack from a number of N B-scans divided by the number N. The transient response phase is to be selected depending on the underlying recording parameters, with the recording frequency and speed of the sensor being important parameters. To ensure a representative mean image of the undisturbed background, it can be advantageous, for example, to record at least 30 target-free B-scans, i.e. free of interfering objects or targets, at the beginning of the sequence. Furthermore, it is advantageous if the ground conditions of the target-free subsurface match those of the subsurface to be analyzed well or, for example, as closely as possible.

[0081] In exemplary embodiments, the change detection is performed unsupervised by means of a calculation 350 (MaxRatio, e.g., determination of a quotient of gray values ​​of individual pixels of the input images and associated inverse values ​​and subsequent maximum formation) of the quotient image 360 ​​from the input image pair. The calculation 350 of the quotient image 360 ​​can, for example, be performed by the quotient image determiner 230 from Fig. 2 The quotient image 360 ​​from Fig. 3 corresponds, for example, to the quotient image 230a from Fig. 2 .

[0082] Optional is in Fig. 3A format change 345 (uint8 conversion, e.g., conversion to a format with unsigned 8-digit integers for describing the grayscale values) is shown, which converts the formats of the captured image 330, the quotient image 360, and the central background image 340 to corresponding images 330', 360', and 340'. This allows the use of morphological image processing methods (e.g., ASF) that require or can use grayscale images with a uint8 dynamic range [0;255].

[0083] Following the calculation of the quotient image from the input image pair, a threshold calculation 370 (threshold) is also performed, for example, unsupervised, using an entropy-based method (e.g., the Rényi method, see [Sahoo et al., 2004]). The threshold calculation 370 is performed on the basis of the quotient image 360' and / or a filtered quotient image 380. The threshold calculation 370 can, for example, be performed by the threshold comparator 250. Fig. 2 Furthermore, the filtered quotient image 83 can be Fig. 3 the filtered quotient image 240a from Fig. 2 are equivalent to.

[0084] The threshold calculation 370, for example, comprising the Renyi method, can, according to embodiments, comprise the calculation and evaluation of a two-dimensional histogram created from two input images. One of these can represent the original (here: quotient image), the second can serve to simulate the local background of the original (here: smoothed or filtered quotient image). Based on these two images, the 2D histogram can be created for each possible gray value pair of this 2D image stack. This histogram can further be used to calculate the foreground and background entropies, which ultimately results in the threshold. It should be noted that in embodiments, the terms "smoothed" and "filtered" can be synonymous, e.g., due to the function or operation of the filter.

[0085] To avoid possible false detections due to noise pixels in the quotient image 360', a filtering 390 (ASF smoothing MaxRatio i) is first performed using an alternating sequential filter (ASF). An ASF is a structure-preserving morphological filter that deletes or preserves the connected components contained in an image for a specific gray value according to their size [Boldt et al., 2019]. The filtering 390 can, for example, be performed by the filter 240 of Fig. 2 be performed.

[0086] After applying the threshold calculation, a binary map 400 results (which corresponds, for example, to the binary change map 250a of Fig. 2corresponds) with potential changes in the image foreground. Due to the successive zero-offset correction (see above), the position of the ground signature is known, which is why it can be deleted from the binary map as a stationary signature after change detection. The extraction 410 of the position of the ground signature is performed based on the central background image 340 and provides the coordinates of the ground signature 420.

[0087] Before the separated image foreground is used as input for the data augmentation 430, a reduction 440 (removal of small connected components) of possible change pixels caused by noise is performed. This step can be performed, for example, in the pixel reducer 260 from Fig. 2can be realized. Binary connected components smaller than a number of pixels are deleted, which corresponds to the minimum size of a sought-after interference object. This results in a reduced binary change map 450 (e.g., corresponding to 260a from Fig. 2 ) which includes, for example, the binary final changes, for example the changes in the quotient image.

[0088] In the following, Fig. 4 Examples of intermediate results of the image processing according to the invention are illustrated.

[0089] Fig. 4 Using two example B-scans, it contains a representation of the images that may be important or even central to the workflow. Fig. 4 shows (intermediate) result images of the workflow Fig. 3 , for two exemplary B-scans (upper row: 4940; lower row: 4420) of a sequence, e.g. sequence 310 from Fig. 3From left to right: Target-free, average background image 340 from the transient phase (MW B-scan); i-th GPR image 330, here No. 4940 or 4420; quotient image (MaxRatio) 360; filtered quotient image (MW MaxRatio) 380; binary change map (CM) 400. In Fig. 4The normalized ground signature 610 can be seen in the middle background images 340 and the GPR images 330. The ground signature is at the same relative position in all images compared to the rest of the image, which is due to the zero-offset correction. Furthermore, an interfering object 620 can be seen in the GPR images 330. In the quotient images 360, the differences between the middle background image 340 or reference image and the respective GPR image 330 or source GPR image are clearly visible. However, not only the interfering object 620 is present in the quotient images 360, but also other artifacts 630. Through the subsequent filtering and threshold comparison, the interfering objects 620 are clearly visible in the binary change map 400. Through filtering, they can be extracted easily and with minimal error, or their position can be determined easily and with minimal error.In the lower change map 400, in addition to the interference object, another artifact 640 can be seen. This can be caused, for example, by a pixel reducer (e.g., 260 from . Fig. 2 ) or the reduction 440 of Fig. 3 by deleting binary connected components of the binary change map 400 smaller than a number of pixels associated with a minimum size of the interference object.

[0090] The change pixels contained in the change map 450 are used as input for the final processing step, data augmentation 430. For this purpose, the original grayscale intensity at the locations of these pixels is extracted 460 (extraction of pixel values) from the original GPR image 330' (here: i-th GPR image) and entered, for example, at the changed position into the middle, target-free image 340' of the settling phase. This results in the result image 470 (result image [uint8]) of the process flow, which can, for example, form the end 480 of the processing sequence. For the entry, the mean value of the i-th GPR image and the image of the settling phase is calculated for each change pixel position. These values ​​are then entered in the result image at the corresponding image coordinates with their correct sign.

[0091] In other words, for the extraction 460 of the pixel values, the change map 450, which indicates the position and / or orientation of the interfering object, and the coordinates 420 of the ground signature can be used. These can be used to determine the pixel values ​​of the ground signature 490 (Pixel Values ​​Ground Sign [uint8]) and the pixel values ​​of the changes 500 (Pixel Values ​​Changes [uint8]) from the GPR image 330' and the central background image 340', based on which the data augmentation 430, together with the central background image 340', can generate the result image 470. The result image 470 is, for example, the result GPR image 120a from Fig. 2 , wherein the data augmentation 430 is generated by the result image generator 120 Figures 1 and 2 can be carried out and the middle background image 340' used for this purpose corresponds to the second GPR image 120b from Fig. 2 can correspond.

[0092] In Fig. 5 are for one of the Fig. 4The (intermediate) result images of the data augmentation are visualized using the given exemplary or exemplary GPR images. The respective interfering object 620 was shifted from left to right at three different depths within the GPR image. Fig. 5 illustrates the enormous potential of embodiments according to the invention for generating large amounts of training data. Although Fig. 5 Since only one displacement of the interfering object within the same image was used, six result GPR images were generated from each source GPR image. The combinatorial possibilities of swapping the interfering objects 620 in the images shown, as well as the additional options such as scaling and / or changing the gray value, allow for the generation of a very large amount of training data.

[0093] In the following, Fig. 6an overview of a method according to the invention according to an embodiment will be shown. Fig. 6 shows a flowchart of a method 800 for generating training data for two-dimensional scans of a ground-penetrating radar system according to an embodiment of the present invention. The method comprises determining 810 a quotient image from a source GPR image with a reference image; and extracting 820 at least one noise object from the source GPR image based on the quotient image; and generating 830 at least one result GPR image for the source GPR image, depending on the at least one extracted noise object, to generate the training data. Conclusions and further remarks

[0094] In general, embodiments of the invention have the advantage, which may be the most important advantage and which is associated with the method in connection with Figures 3 , 4 and 5explained, that a training dataset can be expanded very quickly using a small number of GPR image data. This enables a massive reduction in the effort required for data acquisition. A prerequisite for creating, for example, a representative image background, e.g. as a reference for change detection, is the existence of an undisturbed sequence of spatially consecutive ground-penetrating radar images (transient phase), e.g., one undisturbed by interfering objects. By applying fully automated processing steps, such as calculating the threshold using an entropy-based method (e.g., the Rényi method, see above), a high degree of user-friendliness can be achieved or even ensured.

[0095] In general, embodiments of the invention are not determined by the specific change detection method used. For example, the evaluation of quotient images can be replaced by other methods that, while using a different strategy, produce comparable results. Alternatively or in addition to the concepts mentioned, the quotient image can be smoothed using other methods, e.g., using a median filter or the like, and / or other methods or structures for calculating the threshold can be used. Depending on the method chosen, slightly different results may result.

[0096] Embodiments according to the present invention include, as important or even central steps of their process chain, quotient image-based change detection, which further includes ASF-based filtering and threshold calculation using entropy. Furthermore, the inventive process chain sequence can be used to create training and / or test data, for example, for classifying GPR image data.

[0097] Embodiments according to the present invention are designed to create a large number of suitable training data for machine learning applications based on ground-penetrating radar image data with relatively little effort.

[0098] All lists of materials, environmental influences, electrical properties and optical properties listed herein are to be considered exemplary and not exhaustive.

[0099] Although some aspects have been described in the context of a device, it should be understood that these aspects also represent a description of the corresponding method, so that a block or component of a device can also be understood as a corresponding method step or as a feature of a method step. Analogously, aspects described in the context of or as a method step also represent a description of a corresponding block, detail, or feature of a corresponding device. Some or all of the method steps may be performed by (or using) a hardware apparatus, such as a microprocessor, a programmable computer, or an electronic circuit. In some embodiments, some or more of the key method steps may be performed by such an apparatus.

[0100] Depending on specific implementation requirements, embodiments of the invention may be implemented in hardware or software. The implementation may be performed using a digital storage medium, such as a floppy disk, a DVD, a Blu-ray Disc, a CD, a ROM, a PROM, an EPROM, an EEPROM, or a FLASH memory, a hard disk, or other magnetic or optical storage device storing electronically readable control signals that can interact or cooperate with a programmable computer system to perform the respective method. Therefore, the digital storage medium may be computer-readable.

[0101] Some embodiments according to the invention thus comprise a data carrier having electronically readable control signals capable of interacting with a programmable computer system such that one of the methods described herein is carried out.

[0102] In general, embodiments of the present invention may be implemented as a computer program product having a program code, wherein the program code is effective to perform one of the methods when the computer program product is run on a computer.

[0103] The program code can, for example, also be stored on a machine-readable medium.

[0104] Other embodiments include the computer program for performing one of the methods described herein, wherein the computer program is stored on a machine-readable carrier.

[0105] In other words, an embodiment of the method according to the invention is thus a computer program which has a program code for carrying out one of the methods described herein when the computer program runs on a computer.

[0106] A further embodiment of the method according to the invention is thus a data carrier (or a digital storage medium or a computer-readable medium) on which the computer program for performing one of the methods described herein is recorded. The data carrier, the digital storage medium, or the computer-readable medium is typically physical and / or non-perishable or non-transient.

[0107] A further embodiment of the method according to the invention is thus a data stream or a sequence of signals that represents the computer program for carrying out one of the methods described herein. The data stream or the sequence of signals can be configured, for example, to be transferred via a data communication connection, for example, via the Internet.

[0108] A further embodiment comprises a processing device, for example a computer or a programmable logic device, which is configured or adapted to carry out one of the methods described herein.

[0109] A further embodiment comprises a computer on which the computer program for performing one of the methods described herein is installed.

[0110] A further embodiment according to the invention comprises a device or system designed to transmit a computer program for performing at least one of the methods described herein to a recipient. The transmission can be electronic or optical, for example. The recipient can be, for example, a computer, a mobile device, a storage device, or a similar device. The device or system can, for example, comprise a file server for transmitting the computer program to the recipient.

[0111] In some embodiments, a programmable logic device (e.g., a field-programmable gate array, an FPGA) may be used to perform some or all of the functionalities of the methods described herein. In some embodiments, a field-programmable gate array may interact with a microprocessor to perform any of the methods described herein. In general, in some embodiments, the methods are performed by any hardware device. This may be general-purpose hardware such as a computer processor (CPU) or method-specific hardware such as an ASIC.

[0112] The devices described herein may be implemented, for example, using a hardware apparatus, or using a computer, or using a combination of a hardware apparatus and a computer.

[0113] The devices described herein, or any components of the devices described herein, may be implemented at least partially in hardware and / or in software (computer program).

[0114] The methods described herein may be implemented, for example, using a hardware apparatus, or using a computer, or using a combination of a hardware apparatus and a computer.

[0115] The methods described herein, or any components of the methods described herein, may be implemented at least partially by hardware and / or by software.

[0116] The above-described embodiments are merely illustrative of the principles of the present invention. It is understood that modifications and variations of the arrangements and details described herein will be apparent to others skilled in the art. Therefore, it is intended that the invention be limited only by the scope of the following claims and not by the specific details presented in the description and explanation of the embodiments herein. References

[0117] Sahoo, PK, Arora, G. (2004): A thresholding method based on two-dimensional Renyi's entropy, Pattern Recognition, Vol. 37, No. 6, pp. 1149–1161. Boldt, M., Thiele, A., Schulz, K., Meyer, FJ, Hinz, S. (2019): Practical approach for synthetic aperture radar change analysis in urban environments, Journal of Applied Remote Sensing, Vol. 13, No. 3, pp. 034528-1–034528-29. Soille, P., "Morphological Image Processing: Fundamentals, Methods, Applications," Springer-Verlag Berlin Heidelberg, 1998, 316 pages, ISBN: 3540643230

Claims

1. Apparatus (100, 200) for generating training data for two-dimensional scans of a ground radar system, wherein the apparatus (100, 200) comprises an interference object extractor (110) that is configured to determine a quotient image (230a, 360, 360') with a reference image (110b, 340, 340') from a source GPR image (110a, 330, 330') and to extract at least one interference object (110c, 620) on the basis of the quotient image (230a); and wherein the apparatus (100, 200) comprises a result image generator (120) that is configured to generate at least one result GPR image (120a, 470) for the source GPR image (110a, 330, 330') on the basis of the at least one extracted interference object (110c, 620) in order to generate the training data.

2. Apparatus (100, 200) according to Claim 1, wherein the result image generator (120) is configured to insert the at least one interference object (110c, 620) into a second GPR image (120b) in order to generate the at least one result GPR image (120a, 470).

3. Apparatus (100, 200) according to Claim 2, wherein the second GPR image (120b) is the reference image (110b, 340, 340').

4. Apparatus (100, 200) according to any one of Claims 2 to 3, wherein the result image generator (120) is configured to change the at least one interference object (110c, 620) and / or a position of the at least one interference object (110c, 620) in the second GPR image (210b) in comparison with the source GPR image (110a, 330, 330') in order to generate the at least one result GPR image (120a, 470).

5. Apparatus (100, 200) according to any one of Claims 2 to 4, wherein the result image generator (120) is configured to change a scaling of the at least one interference object (110c, 620) in the second GPR image (120b) in comparison with the source GPR image (110a, 330, 330') in order to generate the at least one result GPR image (120a, 470).

6. Apparatus (100, 200) according to any one of Claims 2 to 5, wherein the result image generator (120) is configured to change a grayscale intensity of the at least one interference object (110c, 620) in the second GPR image (120b) in comparison with the source GPR image (110a, 330, 330') in order to generate the at least one result GPR image (120a, 470).

7. Apparatus (100, 200) according to any one of the preceding claims, wherein the source GPR image (110a, 330, 330') is a two-dimensional GPR image.

8. Apparatus (100, 200) according to any one of the preceding claims, wherein the source GPR image (110a, 330, 330') is a GPR image of a plurality of first GPR images.

9. Apparatus (100, 200) according to Claim 8, wherein the plurality of first GPR images is a plurality of temporally successive GPR images that were generated from a sequence (210a, 310) of a plurality of captured GPR images.

10. Apparatus (100, 200) according to Claim 9, wherein the apparatus (100, 200) is configured to compensate for a distance between sensor and ground surface that varies in the sequence (210a, 310) of the plurality of captured GPR images in order to generate the plurality of first GPR images.

11. Apparatus (100, 200) according to any one of Claims 9 to 10, wherein the apparatus (100, 200) is configured to normalize a ground signature that varies in the sequence (210a, 310) of the plurality of captured GPR images in order to obtain a normalized ground signature (610), and to generate the plurality of first GPR images using the normalized ground signature (610).

12. Apparatus (100, 200) according to any one of Claims 9 to 11, wherein the reference image (110b, 340, 340') is a mean value image (340, 340') from a plurality of interference-free GPR images of the sequence (210a, 310) of the plurality of captured GPR images.

13. Apparatus (100, 200) according to any one of the preceding claims, wherein the interference object extractor (110) comprises a quotient image determiner (230) for determining the quotient image (230a, 360, 360'), and wherein the interference object extractor (110) is configured to perform a change detection on the basis of a filtering (390) and / or a threshold value comparison (370) in order to extract the interference object (110c, 620).

14. Method (300, 800) for generating training data for two-dimensional scans of a ground radar system, the method comprising: determining (810) a quotient image (230a, 360, 360') from a source GPR image (110a, 330, 330') with a reference image (110b, 340, 340'); and extracting (820) at least one interference object (110c, 620) from the source GPR image (110a, 330, 330') on the basis of the quotient image (230a, 360, 360'); and generating (830) at least one result GPR image (120a, 470) for the source GPR image (110a, 330, 330') depending on the at least one extracted interference object (110c, 620) in order to generate the training data.

15. Computer program having a program code for preforming the method (300, 800) according to Claim 14.