A method and system for improving ground penetrating radar signal resolution based on deep learning

Abstract

The application discloses a kind of based on deep learning's ground penetrating radar signal resolution improvement method and system, high frequency radar and low frequency radar are respectively used to collect radar data as training data set;Cycle-GAN network model is constructed, and Cycle-GAN network includes two mirror symmetry generation adversarial network structures;Cycle-GAN network model is trained using training data set;The low-resolution ground penetrating radar data actually collected is tested using the trained Cycle-GAN network model, and the based on deep learning's ground penetrating radar signal resolution improvement method and system of the application can be based on the data set of unpaired high-frequency ground penetrating radar data label and low-frequency ground penetrating radar data sample, build the Cycle-GAN network model including two mirror symmetry generation adversarial network structures, using data set to train the Cycle-GAN network model constructed, adaptively learn the mapping relationship between high-resolution and low-resolution data domain, and high-quality ground penetrating radar data resolution is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of ground penetrating radar data processing, and particularly relates to a ground penetrating radar signal resolution improvement method and system based on deep learning. BACKGROUND

[0002] Ground penetrating radar (GPR) is a geophysical exploration tool commonly used for shallow underground structure exploration. By using the strong penetration ability of high-frequency electromagnetic waves to the ground surface, ground penetrating radar transmits electromagnetic waves to the ground surface and receives the transmitted or back-reflected signals to achieve the detection of underground targets.

[0003] In actual exploration, the selection of the frequency of the ground penetrating radar transmission signal directly affects the detection depth and spatial resolution. Since high-frequency radar waves have a shorter wavelength, they can provide more information in shallow underground medium exploration, but their energy decays quickly and the detection depth is shallow. Low-frequency radar waves have a longer wavelength and can detect information about deeper underground media, but the data obtained often has low resolution. Radar data super-resolution allows the detection information resolution to be improved while retaining the detection depth of low-frequency radar data, greatly improving the detection efficiency and reducing the detection cost. Therefore, accurately and effectively improving the resolution of low-frequency ground penetrating radar data is of great significance in actual ground penetrating radar detection.

[0004] Traditional methods for improving the resolution of radar data include deconvolution, inverse Q filtering, and spectral whitening. The deconvolution method is based on the assumption of convolution model, which often differs from the actual geological model. The inverse Q filtering method compensates for the loss according to the existing attenuation model, but its effect is directly affected by the accuracy of the quality factor Q estimation. The spectral whitening method is simple and intuitive to implement, but its fidelity is poor, which directly affects the signal-to-noise ratio of the radar data. The above methods can effectively improve the resolution of radar data, but the results are highly dependent on the selection of model parameters. In recent years, with the rapid development and wide application of deep learning and other methods, many researchers have proposed using deep learning methods to improve the resolution of ground penetrating radar signals to reduce the impact of parameter selection.

[0005] Existing methods use continuous wavelet transform to process ground penetrating radar data; use the Morlet complex wavelet function as the basis function of continuous wavelet transform, perform continuous wavelet transform on single-channel data, select a reference frequency based on octave for spectrum expansion, and perform inverse wavelet transform on the expanded information to improve the data resolution. However, the selection of the reference frequency directly affects the effect of signal bandwidth expansion, and the effect on signals with rapid high-frequency information decay is limited, with a small expansion width. SUMMARY

[0006] The technical problem to be solved by the present application is to provide a ground penetrating radar signal resolution improvement method and system based on deep learning to solve the technical problem of complex manual selection of large-scale data reference frequency and expansion parameters in view of the deficiencies in the prior art.

[0007] The present application adopts the following technical solutions:

[0008] A ground penetrating radar signal resolution improvement method based on deep learning comprises the following steps:

[0009] S1, constructing a training data set;

[0010] S2, constructing a Cycle-GAN network model, the Cycle-GAN network model comprising two mirror-symmetrical generative adversarial network structures;

[0011] S3, using the training set data set obtained in step S1 to train the Cycle-GAN network model constructed in step S2;

[0012] S4, using the trained Cycle-GAN network model in step S3 to perform resolution improvement processing on the actually collected low-resolution ground penetrating radar data.

[0013] Specifically, in step S1, the training data set comprises radar data actually collected by a high-frequency radar and radar data actually collected by a low-frequency radar, the low-resolution data collected by the low-frequency ground penetrating radar is preprocessed by interpolation, the data after interpolation preprocessing is used as sample data, the high-resolution data collected by the high-frequency radar is used as label data, and the training data set is constructed using the sample data and the label data.

[0014] Further, the sampling time interval of the low-resolution data after interpolation processing is the same as that of the high-resolution data.

[0015] Specifically, in step S2, the generative adversarial network structure shares a generator GeneratorA2B and a generator GeneratorB2A, and carries an discriminator respectively; the generator GeneratorA2B can convert the sample in the data domain D A to high-resolution data in the target domain D B , and the generator GeneratorB2A can convert the high-resolution data in the data domain D B to the original data in the domain D A .

[0016] Further, the network structure of the generator GeneratorA2B and the generator GeneratorB2A is the same, comprising an encoder, a converter and a decoder.

[0017] Further, the one-way workflow of the Cycle-GAN network model is as follows:

[0018] An input sample Input_A is obtained from the data domain D A , the input sample Input_A is transmitted to the generator GeneratorA2B to generate a new sample Generated_B, the new sample Generated_B is transmitted to the generator GeneratorB2A, and the original domain D A is converted into the original data Pseduo_A; the discriminator is used for judgment, and the network parameters are iteratively updated.

[0019] Further, the new sample Generated_B is specifically:

[0020] Data features are extracted from the input sample Input_A by using a convolutional neural network, the data features are compressed into 256 128*128 feature vectors, and the features that are not close to each other are combined, so that the sample can be converted from the data domain D A to the data domain D B ; finally, low-level features are restored by using a deconvolutional layer to generate a new sample Generated_B.

[0021] Specifically, in step S3, the initial learning rate is set to 0.0002 and is attenuated with iterations; the adversarial loss and the cycle consistency loss function are defined as the optimization objective, and the Adam optimization algorithm is used to bias correct the network parameters to find the optimal solution through iterative learning.

[0022] Further, the adversarial loss and the cycle consistency loss function L(G,F,D X ,D Y ) are as follows:

[0023] L(G,F,D X ,D Y )=L GAN (G,D Y ,X,Y)+L GAN (F,D X ,Y,X)+λL cyc (G,F)

[0024] Wherein, L GAN (G,D Y ,X,Y) is the adversarial loss from the X domain to the Y domain, L GAN (F,D X ,Y,X) is the adversarial loss from the Y domain to the X domain, L cyc (G,F) is the cycle consistency loss for judging the consistency of the generators G and F, and λ is a non-negative parameter for controlling the relative importance of the two generators G and F.

[0025] In a second aspect, the embodiment of the present application provides a ground penetrating radar signal resolution improvement system based on deep learning, comprising:

[0026] a data module configured to construct a training data set;

[0027] a construction module configured to construct a Cycle-GAN network model, wherein the Cycle-GAN network model comprises two mirror-symmetrical generative adversarial network structures;

[0028] a training module configured to train the Cycle-GAN network model constructed by the construction module using the training data set obtained by the data module;

[0029] a processing module configured to perform resolution improvement processing on the actually collected low-resolution ground penetrating radar data using the Cycle-GAN network model trained by the training module.

[0030] Compared with the prior art, the present application has at least the following beneficial effects:

[0031] The ground penetrating radar signal resolution improvement method based on deep learning uses the Cycle-GAN network as the ground penetrating radar data resolution improvement scheme, can adaptively learn the mapping relationship between unpaired samples, and realizes high-quality data resolution improvement; has high fidelity to radar data effective signals, and can effectively protect the effective signals; after constructing the data set, the data set is sent to the Cycle-GAN network comprising two mirror-symmetrical generative adversarial network structures for sufficient training, and the network after sufficient training comprises the parameter weight for mapping from low-resolution data to high-resolution data, so that the target data sent into the network can obtain the resolution improved result in real time.

[0032] Further, in step S1, the constructed data set comprises the radar data actually collected by the high-frequency radar and the radar data actually collected by the low-frequency radar, and based on the different sampling time intervals of the high-frequency radar and the low-frequency radar, the low-resolution data collected by the low-frequency ground penetrating radar is subjected to interpolation preprocessing, so that the high-resolution data and the low-resolution data are matched in the time direction, the sampling time interval of the low-resolution data after interpolation processing is the same as that of the high-resolution data, the complexity of network parameter learning is simplified, the data after interpolation preprocessing is used as sample data, the high-resolution data collected by the high-frequency radar is used as label data, and the sample data and the label data are used to construct the training data set.

[0033] Further, in step S2, the Cycle-GAN network model is constructed, and the generated adversarial network structure shares the generator GeneratorA2B and the generator GeneratorB2A, and each carries a discriminator; the generator GeneratorA2B can convert the sample in the data domain D A into high-resolution data in the target domain D B , and the generator GeneratorB2A can convert the high-resolution data in the data domain D B into original data in the domain D A ; the discriminator is used to judge whether the converted data meets the requirements in the training process and is used for the next training; in one training, the input sample from the original data domain is sent into the sample generated by the first generator, and then the sample is sent into another generator, and finally the original input sample is judged to update the network parameters.

[0034] Further, since the functions and purposes of the settings of the generator GeneratorA2B and the generator GeneratorB2A are the same, the network structures of the generator GeneratorA2B and the generator GeneratorB2A are the same, including an encoder, a converter, and a decoder, the encoder uses convolution to extract the attributes of the input sample, the converter is used to convert the attributes between domains, and the decoder uses deconvolution to generate new samples.

[0035] Further, in one direction, the data is sent into one generator in the constructed Cycle-GAN network model to obtain a generated sample, and the generated sample data is sent into another generator, and the output result is compared with the initial input data to update the network parameters; in the overall training process of the network, the process is bidirectional, which further strengthens the learning of the network on the mapping between the unpaired data domains.

[0036] Further, the initial learning rate is set to 0.0002 and is attenuated with iterations, thereby reducing the possibility of network gradient diffusion, defining the adversarial loss and the cycle consistency loss function as the optimization objective to enhance the constraints of the network on the data domain mapping learning, using the high-performance Adam optimization algorithm to bias correct the network parameters, and iteratively learning to find the optimal solution.

[0037] Further, the network loss function can be further divided into an adversarial loss and a cycle consistency loss function, the main purpose of the adversarial loss is to make the distribution of the network generated data matched with the distribution of the high-resolution data, and the cycle consistency loss is used to constrain the generated samples of the generator, and to strengthen the corresponding relationship between the forward and reverse mappings.

[0038] It can be understood that the beneficial effects of the above-mentioned second aspect can be referred to the related description in the above-mentioned first aspect, which will not be repeated here.

[0039] To sum up, the application constructs two mirror-symmetrical generative adversarial network structures, uses the adversarial loss and the cycle consistency loss function as the optimization target, learns the feature attributes and mapping functions of the low-resolution data and high-resolution data adaptively without manual selection of parameters, and realizes a high level of radar data resolution improvement target.

[0040] The technical solutions of the application will be further described in detail below with reference to the drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 It is a Cycle-GAN network structure diagram;

[0042] Figure 2 It is a low-frequency radar data diagram;

[0043] Figure 3 It is a high-frequency radar data diagram;

[0044] Figure 4 It is a low-resolution radar data interpolation sample diagram;

[0045] Figure 5 It is a high-resolution radar data label diagram;

[0046] Figure 6 It is a low-resolution data diagram;

[0047] Figure 7 It is high-resolution data obtained by the method of the application;

[0048] Figure 8 It is a high-resolution data diagram for comparison;

[0049] Figure 9 It is Figure 6 low-resolution data, Figure 7 high-resolution results obtained by the application, and Figure 8 high-resolution comparison diagram amplitude spectrum comparison;

[0050] Figure 10 It is a flowchart of the application. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0052] In the description of the application, it should be understood that the terms "include" and "contain" indicate the presence of described features, whole, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, whole, steps, operations, elements, components and / or sets thereof.

[0053] It should also be understood that the terms used in the specification of the application are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in the specification and the appended claims of the application, the singular forms "a", "an" and "the" are intended to include the plural forms, unless the context clearly indicates otherwise.

[0054] It should be further understood that the term "and / or" used in the specification and the appended claims of the application means one or more of the associated listed items in any combination and all possible combinations, and includes these combinations, for example, A and / or B can represent three cases: A exists alone, A and B exist together, and B exists alone. In addition, the character " / " in this paper generally represents that the front and rear associated objects are in an "or" relationship.

[0055] It should be understood that although the terms first, second, third, etc. may be used in the embodiments of the application to describe the preset ranges, etc., these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from each other. For example, the first preset range can also be referred to as the second preset range, and similarly, the second preset range can also be referred to as the first preset range without departing from the scope of the embodiments of the application.

[0056] Depending on the context, the word "if" as used herein can be interpreted to mean "when" or "while" or "in response to determining" or "in response to detecting". Similarly, depending on the context, the phrase "if it is determined" or "if (a stated condition or event) is detected" can be interpreted to mean "when it is determined" or "in response to determining" or "when (a stated condition or event) is detected" or "in response to detecting (a stated condition or event)".

[0057] Various structural diagrams according to the disclosed embodiments of the application are shown in the drawings. These figures are not drawn to scale, in which certain details are exaggerated for the purpose of clarity and certain details can be omitted. The shapes of various regions, layers and their relative size and positional relationship shown in the figures are only exemplary, and in actuality, there can be deviations due to manufacturing tolerances or technical limitations, and a person skilled in the art can additionally design regions / layers with different shapes, sizes and relative positions according to actual needs.

[0058] The application provides a deep learning-based ground penetrating radar signal resolution improvement method, which performs interpolation preprocessing on low-resolution data collected by a low-frequency ground penetrating radar, uses the data after interpolation preprocessing as samples, uses high-resolution data collected by a high-frequency radar as labels, uses the sample data and the label data to construct a training data set, constructs a Cycle-GAN network, and inputs the training data set into the Cycle-GAN network for training, and after the training is completed, uses the Cycle-GAN network to achieve the purpose of resolution improvement on test low-resolution data; the data set can be constructed based on unpaired high-frequency ground penetrating radar data labels and low-frequency ground penetrating radar data samples, a Cycle-GAN network model containing two mirror-symmetrical generative adversarial network structures is built, the Cycle-GAN network model is trained by using the data set, the mapping relationship between the high-resolution and low-resolution data domains is adaptively learned, and the resolution improvement of the ground penetrating radar data is realized in a high quality.

[0059] Referring to Figure 10 , the application is a deep learning-based ground penetrating radar signal resolution improvement method, which comprises the following steps:

[0060] S1, constructing a training data set;

[0061] The radar data collected by the high-frequency radar and the low-frequency radar in the training data set respectively; since the radar wave with high frequency has a shorter wavelength, rich information can be obtained in the exploration of shallow underground media, but the energy attenuation is faster, and the detection depth is shallower; the radar wave with low frequency has a longer wavelength, and can detect information of deeper underground media, but the obtained data often has low resolution.

[0062] According to the sampling time interval, the low-resolution radar data is interpolated, the sampling time interval of the interpolated low-resolution data is consistent with that of the high-resolution data, and the size of the constructed sample and label is 128*128.

[0063] S2, constructing a Cycle-GAN network model, the Cycle-GAN network model contains two mirror-symmetrical generative adversarial network structures;

[0064] Referring to Figure 1 , the Cycle-GAN essentially contains two mirror-symmetrical generative adversarial networks, the two GANs share two generators, and each carries a discriminator. The generator GeneratorA2B can convert the sample in the data domain D A into high-resolution data in the target domain D B , and the generator GeneratorB2A converts the high-resolution data in D B into the domain D Athe original data in the data domain D

[0065] S201, first, input samples Input_A are obtained from the data domain D A , which are passed to the first generator GeneratorA2B to generate new samples Generated_B;

[0066] S202, the generated samples Generated_B are passed to another generator GeneratorB2A, which converts the original data Pseduo_A in the data domain D A to the original data in the data domain D

[0067] S203, the discriminator is used for judgment, so as to iteratively update the network parameters.

[0068] The network structures of GeneratorA2B and GeneratorB2A in the network model are the same, and the network structures of Discriminator A and Discriminator B are also the same. Among them, the generation network is composed of an encoder, a converter and a decoder.

[0069] First, the convolutional neural network is used to extract data features from the input samples, and the samples are compressed into 256 128*128 feature vectors. By combining non-close features, the samples can be converted from the data domain D A to the data domain D B ; finally, the low-level features are restored by using the deconvolution layer to generate new samples.

[0070] In this network, a very important residual module is also used, which is composed of two convolutional layers and can make the sample retain the original features while converting.

[0071] The discriminator network is actually a convolutional neural network, which can be used to judge whether the input sample is real data or the output data of the generator. It extracts the features of the input data, and determines whether the sample belongs to a specific category by adding a one-dimensional output convolutional layer.

[0072] S3, the Cycle-GAN network model constructed in step S2 is trained according to the training set samples in step S1;

[0073] The initial learning rate is set to 0.0002 and is attenuated with iteration; the adversarial loss and the cycle consistency loss function are defined as the optimization objective, and the Adam optimization algorithm is used to bias correct the network parameters to find the optimal solution through iterative learning.

[0074] The loss function of the network can be generally divided into adversarial loss and cycle consistency loss function, including:

[0075] L(G,F,D X Y )=L GAN (G,D Y ,X,Y)+L GAN (F,D X ,Y,X)+λL cyc (G,F)

[0076] Wherein, G and F are generators, D X and D Y are discriminators, X and Y are two different mapping domains, and λ is a non-negative parameter for controlling the relative importance of the two generators G and F;The adversarial loss makes the generated sample consistent with the target sample data distribution, and the specific form is:

[0077]

[0078] Wherein, the mapping function is G:X→Y, G(x) is a generated sample, and D Y is used for discriminating the generated sample G(x) and the target sample y;For the mapping function F:Y→X, the loss function is L GAN (F,D X ,Y,X);

[0079]

[0080] The above formula is the cycle consistency loss, which is used to judge the consistency of G and F;It mainly includes forward cycle consistency loss and reverse cycle consistency loss, namely:

[0081]

[0082] According to the designed loss function and network structure, the network is trained alternately to converge quickly.

[0083] S4, the Cycle-GAN network model trained in step S3 is used to test the actual collected low-resolution ground penetrating radar data, so as to improve the resolution.

[0084] In another embodiment of the application, a deep learning-based ground penetrating radar signal resolution improvement system is provided, which can be used to implement the above deep learning-based ground penetrating radar signal resolution improvement method, specifically, the deep learning-based ground penetrating radar signal resolution improvement system comprises a data module, a construction module, a training module and a processing module.

[0085] Wherein, the data module constructs a training data set.

[0086] ​The constructing module constructs a Cycle-GAN network model, and the Cycle-GAN network model comprises two mirror-symmetrical generative adversarial network structures.

[0087] The training module trains the Cycle-GAN network model constructed by the constructing module by using the training set data set obtained by the data module.

[0088] The processing module performs resolution improvement processing on the actually collected low-resolution ground penetrating radar data by using the trained Cycle-GAN network model.

[0089] In another embodiment of the present application, a terminal device is provided, which comprises a processor and a memory, the memory is used to store a computer program, the computer program comprises program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, and is specifically suitable for loading and executing one or more instructions to realize the corresponding method flow or corresponding function. The processor in the embodiment of the present application can be used for the operation of the ground penetrating radar signal resolution improvement method based on deep learning, which comprises:

[0090] The training data set is constructed, the Cycle-GAN network model is constructed, the Cycle-GAN network model comprises two mirror-symmetrical generative adversarial network structures, the Cycle-GAN network model is trained by using the training data set, and the resolution improvement processing is performed on the actually collected low-resolution ground penetrating radar data by using the trained Cycle-GAN network model.

[0091] In still another embodiment of the present application, the present application also provides a storage medium, specifically a computer readable storage medium (Memory), which is a memory device in a terminal device, used for storing programs and data. It can be understood that the computer readable storage medium herein can include a built-in storage medium in the terminal device, and of course can also include an extended storage medium supported by the terminal device. The computer readable storage medium provides a storage space, which stores an operating system of the terminal. Moreover, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and the instructions can be one or more computer programs (including program codes). It should be noted that the computer readable storage medium herein can be a high-speed RAM memory or a non-volatile memory such as at least one disk memory.

[0092] The one or more instructions stored in the computer readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the above-mentioned embodiment of the method for improving the resolution of ground penetrating radar signals based on deep learning.

[0093] The training data set is constructed, a Cycle-GAN network model is constructed, the Cycle-GAN network model includes two mirror-symmetrical generative adversarial network structures, the Cycle-GAN network model is trained by using the training data set, and the resolution of the actually collected low-resolution ground penetrating radar data is improved by using the trained Cycle-GAN network model.

[0094] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0095] The following will be based on a method and system for improving the resolution of ground penetrating radar signals based on deep learning of the present application applied to the actually collected radar data, and it is found that the method of the present application can improve the resolution of radar data, and the data spectrum is effectively widened.

[0096] Please refer to Figure 2 , for the actual collection of 900MHz low-resolution radar data, the sampling points are 512, the sampling interval is 24ns, and there are 5916 channels.

[0097] Please refer to Figure 3 , for the actual collection of 2GHz high-resolution radar data, the sampling points are 512, the sampling interval is 12ns, and there are 11640 channels per meter of exploration.

[0098] Please refer to Figure 4 , for the constructed low-resolution sample example, the sample size is 128*128, which is obtained by interpolating and pre-processing the low-resolution radar data shown in Figure 2 .

[0099] Please refer to Figure 5 , for the constructed high-resolution label example, the label size is 128*128, which is obtained by pre-processing the high-resolution radar data shown in Figure 3 , the high-resolution label and the low-resolution sample do not correspond in space and time, and the high-resolution radar data contains more information than the low-resolution radar data, which can be used for network training.

[0100] Please refer to Figure 6 , for the test original low-resolution radar data, the data size is 512*512, which is obtained by interpolating and cropping the actual collected low-resolution data.

[0101] Please refer to Figure 7 , for the high-resolution radar data obtained by the method of the present application, compared with Figure 6 , it can be seen that the results obtained by the method of the present application can effectively reflect the information of the deeper position in the low-resolution radar data, and can more detailedly show the shallower stratum.

[0102] Please refer to Figure 8 , for the original high-resolution radar data, compared with Figure 7 , it can be seen that the high-resolution results obtained by the method of the present application can more accurately correspond to the shallow information in the original high-resolution radar detection data.

[0103] Please refer to Figure 9 , for the amplitude spectrum comparison of Figure 6 , Figure 7 and Figure 8 ; from the amplitude spectrum comparison analysis, it can be seen that the data spectrum obtained by the method of the present application is widened, containing more high-frequency fine stratum information, which shows that the method has certain effectiveness and feasibility.

[0104] The above model and actual data examples show that the ground penetrating radar signal resolution improvement method and system based on deep learning can realize high-quality resolution improvement of ground penetrating radar data, effectively compressing the shallow high-energy signal while widening the frequency spectrum.

[0105] In summary, the ground penetrating radar signal resolution improvement method and system based on deep learning can realize radar data resolution improvement by constructing a Cycle-GAN network, using unpaired high-frequency radar acquisition data and low-frequency data to construct a data set for training, and learning the mapping from low-resolution data to high-resolution data. The method of the present application can realize ground penetrating radar resolution improvement and has strong adaptability. At the same time, the method of the present application is weakly supervised learning and does not depend on complete pairing of data sets, so it has stronger performance and applicability.

[0106] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the above division of each functional unit and module is exemplified, and in actual application, the above functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The above integrated unit can be realized in the form of hardware or software. In addition, the specific names of each functional unit and module are only for easy distinction, and do not limit the protection scope of the present application. The specific working process of the units and modules in the above system can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0107] In the above embodiments, the description of each embodiment has its own emphasis, and the parts not described or recorded in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0108] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0109] In the embodiments of the present application, it should be understood that the disclosed apparatus / terminal and method can be implemented in other manners. For example, the embodiments of the apparatus / terminal described above are merely schematic, and the division of the modules or units is merely logical function division, and there can be another division manner in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between the units can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical, mechanical or in other forms.

[0110] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0111] In addition, each functional unit in the various embodiments of the present application can be integrated into a processing unit, or each unit can be a physically independent unit, or two or more units can be integrated into a unit. The integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0112] The integrated module / unit, if implemented in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, all or part of the flow of the above-mentioned embodiment methods can be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. When the processor executes the computer program, the steps of each method embodiment described above can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, executable file or some intermediate form. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the computer readable medium can include or exclude content according to the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, the computer readable medium does not include electrical carrier signals and telecommunication signals.

[0113] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps in one or more flow or blocks

[0114] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps in one or more flow or blocks

[0115] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps in one or more flow or blocks

[0116] The above merely provides the technical idea of the present application, and cannot be used to limit the protection scope of the present application. Any modification made according to the technical idea of the present application, on the basis of the technical solutions, falls within the protection scope of the claims of the present application.

Claims

1. A method for improving the resolution of ground-penetrating radar signals based on deep learning, characterized in that, Includes the following steps: S1. Construct the training dataset; S2. Construct a Cycle-GAN network model. The Cycle-GAN network model contains two mirror-symmetric generative adversarial network (GAN) structures. These GAN structures share a generator (GeneratorA2B) and a generator (GeneratorB2A), and each carries a discriminator. Generator A2B can process the data domain... Transform samples from the target domain to the target domain For high-resolution data, the GeneratorB2A can transform the data domain High-resolution data in the domain is converted into a domain. The original data in the generator; the network structures of GeneratorA2B and GeneratorB2A are the same, including encoder, converter and decoder; The unidirectional workflow of the Cycle-GAN network model is as follows: From the data domain Obtain input samples Input_A Input sample Input_A Pass the sample to the GeneratorA2B to generate a new sample. Generated_B ; New sample Generated_B Passed to GeneratorB2A, in the original domain Convert to raw data Pseduo_A Use a discriminator to make judgments and iteratively update network parameters; S3. Train the Cycle-GAN network model constructed in step S2 using the training dataset obtained in step S1. Define adversarial loss and cycle consistency loss functions as optimization objectives. as follows: in, From Domain generation to Domain confrontation losses, for from Adversarial loss from domain generation to domain, To determine the generator and The circular consistency loss of consistency For controlling two generators and The nonnegative parameter of relative importance; S4. Use the Cycle-GAN network model trained in step S3 to improve the resolution of the actual low-resolution ground-penetrating radar data.

2. The method for improving the resolution of ground-penetrating radar signals based on deep learning according to claim 1, characterized in that, In step S1, the training dataset includes radar data actually collected by high-frequency radar and radar data actually collected by low-frequency radar. The low-resolution data collected by low-frequency ground-penetrating radar is preprocessed by interpolation, and the preprocessed data is used as sample data. The high-resolution data collected by high-frequency radar is used as label data. The training dataset is constructed using the sample data and label data.

3. The method for improving the resolution of ground-penetrating radar signals based on deep learning according to claim 2, characterized in that, After interpolation, the sampling time interval of the low-resolution data is the same as that of the high-resolution data.

4. The method for improving the resolution of ground-penetrating radar signals based on deep learning according to claim 1, characterized in that, Generate new samples Generated_B Specifically: Using convolutional neural networks to extract from input samples Input_A Data features are extracted and compressed into 256 128*128 feature vectors. Dissimilar features are combined to enable samples to be extracted from the data domain. Transform to data domain Finally, deconvolutional layers are used to recover low-level features and generate new samples. Generated_B .

5. The method for improving the resolution of ground-penetrating radar signals based on deep learning according to claim 1, characterized in that, In step S3, the initial learning rate is set to 0.0002 and decays with each iteration; the Adam optimization algorithm is used to correct the bias of the network parameters, and the optimal solution is found through iterative learning.

6. A deep learning-based system for improving the resolution of ground-penetrating radar signals, characterized in that, include: The data module is used to construct the training dataset. The building block constructs a Cycle-GAN network model. This model contains two mirror-symmetric generative adversarial network (GAN) structures, Generator A2B and Generator B2A, which share a common generator and each carries a discriminator. Generator A2B can process the data domain... Transform samples from the target domain to the target domain For high-resolution data, the GeneratorB2A can transform the data domain High-resolution data in the domain is converted into a domain. The original data in the generator; the network structures of GeneratorA2B and GeneratorB2A are the same, including encoder, converter and decoder; The unidirectional workflow of the Cycle-GAN network model is as follows: From the data domain Obtain input samples Input_A Input sample Input_A Pass the sample to the GeneratorA2B to generate a new sample. Generated_B ; New sample Generated_B Passed to GeneratorB2A, in the original domain Convert to raw data Pseduo_A Use a discriminator to make judgments and iteratively update network parameters; The training module uses the training dataset obtained from the data module to train the Cycle-GAN network model constructed by the construction module. Adversarial loss and cycle consistency loss functions are defined as optimization objectives. as follows: in, From Domain generation to Domain confrontation losses, for from Adversarial loss from domain generation to domain, To determine the generator and The circular consistency loss of consistency For controlling two generators and The nonnegative parameter of relative importance; The processing module uses the Cycle-GAN network model trained by the training module to improve the resolution of the actual low-resolution ground-penetrating radar data.

Images