Band extension method based on generative adversarial network

Abstract

The application provides a band expansion method and system based on a generative adversarial network, a storage medium and an electronic device, and relates to the technical field of radar signal processing.The application realizes the fusion of multi-band radar signals and generates large-bandwidth radar signals by constructing a GAN network embedded with a lightweight UNet-like network.On the one hand, the lightweight UNet-like feature extraction network automatically extracts features from the original echo signal, which can facilitate the phase matching between sub-bands and avoid the processes of establishing a signal model and phase registration, and compared with a traditional algorithm, the application is simpler and more efficient.On the other hand, the GAN network can significantly improve the correctness of the fused signal through the game between the generator and the discriminator.In addition, the GAN network can expand the bandwidth while denoising, and even in the case of low signal-to-noise ratio of the original signal, the large-bandwidth signal can still be synthesized.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing technology, and more specifically to a bandwidth extension method, system, storage medium, and electronic device based on generative adversarial networks. Background Technology

[0002] In the process of radar target detection, its range resolution and imaging resolution are directly related to its bandwidth. How to improve the bandwidth and resolution of radar systems is a key research topic in the field of radar detection. Increasing the operating bandwidth of radar from a hardware perspective is costly. However, researching new multi-band fusion algorithms that effectively utilize existing radar signals from different frequency bands and fuse them to generate a wideband radar signal can improve the resolution of radar systems without increasing hardware costs.

[0003] In related technologies, early multi-frequency fusion algorithms mainly relied on radar echo signals for frequency band extrapolation, performing coherent registration and frequency band fusion on sub-band signals. Wideband extrapolation algorithms required high accuracy in the signal model, and the width of the intermediate blank frequency band could not exceed twice the width of the narrowest bandwidth. Subsequent spectral estimation methods did not require spectral extrapolation, but necessitated multiple spectral estimation algorithms, resulting in significant computational overhead. Both of these algorithms predicted missing frequency bands based on existing data, leading to a decrease in synthesized quality under low signal-to-noise ratio conditions.

[0004] Therefore, it is necessary to provide a new method for bandwidth extension to overcome the above-mentioned shortcomings. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a bandwidth extension method, system, storage medium, and electronic device based on generative adversarial networks, solving the technical problem of reduced synthesis quality under low signal-to-noise ratio conditions.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A bandwidth extension method based on generative adversarial networks includes:

[0010] The radar measurement values ​​that are angularly coherent within the target's full frequency band are obtained as true full-band data. The true full-band data are then filtered using different bandpass filters to obtain sub-band data with intervals in the coverage of multiple frequency bands.

[0011] Obtain a pre-built GAN network; wherein the GAN network includes a generator employing a lightweight encoder-decoder network similar to UNet and a discriminator employing a multi-scale architecture; perform the following training steps:

[0012] Gaussian white noise is added to the multiple sub-band data for enhancement processing and used as input to the generator to generate pseudo full-band data;

[0013] The pseudo full-band data and the real full-band data are used as inputs to the discriminator to determine the authenticity of the full-band data.

[0014] If the GAN network meets the training completion condition, the multiple sub-band data to be expanded are used as input to the trained generator to obtain the required full-band data.

[0015] Preferably, the bandwidth extension method further includes:

[0016] If the GAN network does not meet the training completion conditions, the relevant parameters in the GAN network are adjusted to optimize the loss function, and the above training steps are continued based on the adjusted GAN network; wherein adjusting the relevant parameters of the GAN network to optimize the loss function includes:

[0017] Keeping the parameters of the generator unchanged, the discriminator loss is minimized by adjusting the parameters of the corresponding discriminator.

[0018] Keeping the parameters of the discriminator unchanged, the generator loss is minimized by adjusting the parameters of the corresponding generator.

[0019] The discriminator loss and the generator loss are alternately optimized until the network converges.

[0020] Preferably, the discriminator loss is expressed as:

[0021]

[0022] Among them, L D E represents the discriminator loss; y represents the true full-band data; E represents the discriminator loss. y This represents the average of all real full-band data in the training batch; K represents the number of discriminators, N k D represents the total number of output elements of the k-th discriminator; max represents maximizing; k,n (y) represents the nth value of the output vector of the k-th discriminator when the input is y; x represents the multiple sub-band data after enhancement processing, E x G(x) represents the average of multiple sub-band data after enhancement processing in the training batch; G(x) represents the pseudo-full-band data generated by the generator when the input is x; D k,n(G(x)) represents the nth value of the output vector of the kth discriminator when the input is x.

[0023] Preferably, the generator loss is expressed as:

[0024]

[0025] Among them, L G Indicates generator loss; Represents the hinge loss term; Let N represent the feature matching loss term, where λ represents the weight of this term; k,l F represents the total number of elements in a single feature of the l-th layer of the k-th discriminator; k,l D represents the number of features in the l-th layer of the k-th discriminator; ||·|| represents the L2 norm; D k,n (G(x)) represents the feature vector of the l-th layer of the k-th discriminator when the input is G(x); D k,n (x) represents the feature vector of the l-th layer of the k-th discriminator when the input is x.

[0026] Preferably, the generator includes a UNet-like network, which is represented by a U-shaped structure consisting of multiple encoders and decoders of equal number, wherein the encoders and decoders include convolutional layers and residual modules.

[0027] Preferably, the discriminator consists of at least three sub-discriminators. The first sub-discriminator is used to receive pseudo full-band data generated by the generator, and the remaining sub-discriminators are used to receive pseudo full-band data after downsampling according to a corresponding preset multiple.

[0028] A bandwidth extension system based on generative adversarial networks, comprising:

[0029] The sample acquisition module is used to acquire radar measurement values ​​that are angularly coherent within the target's full frequency band as real full-band data, and to filter the real full-band data using different bandpass filters to obtain sub-band data with intervals in multiple frequency band coverage areas.

[0030] A network acquisition module is used for a pre-built GAN network; wherein the GAN network includes a generator employing a lightweight encoder-decoder network similar to UNet and a discriminator employing a multi-scale architecture;

[0031] The data generation module is used to add Gaussian white noise to multiple sub-frequency band data for enhancement processing and use it as input to the generator to generate pseudo full-frequency band data;

[0032] The authenticity judgment module is used to use the pseudo full-band data and the real full-band data as inputs to the discriminator to judge the authenticity of the full-band data.

[0033] The frequency band extension module, if the GAN network meets the training completion conditions, uses the multiple sub-frequency band data to be extended as input to the trained generator to obtain the required full-frequency band data.

[0034] Preferably, the bandwidth extension system further includes:

[0035] The parameter adjustment module is used to adjust the relevant parameters in the GAN network to optimize the loss function if the GAN network does not meet the training completion conditions, and to continue the above training steps based on the adjusted GAN network; wherein adjusting the relevant parameters of the GAN network to optimize the loss function includes:

[0036] Keeping the parameters of the generator unchanged, the discriminator loss is minimized by adjusting the parameters of the corresponding discriminator.

[0037] Keeping the parameters of the discriminator unchanged, the generator loss is minimized by adjusting the parameters of the corresponding generator.

[0038] The discriminator loss and the generator loss are alternately optimized until the network converges.

[0039] A storage medium storing a computer program for bandwidth extension based on generative adversarial networks, wherein the computer program causes a computer to perform the bandwidth extension method as described above.

[0040] An electronic device, comprising:

[0041] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the band extension method as described above.

[0042] (III) Beneficial Effects

[0043] This invention provides a bandwidth extension method, system, storage medium, and electronic device based on generative adversarial networks. Compared with existing technologies, it has the following advantages:

[0044] This invention achieves the fusion of multi-band radar signals and generates high-bandwidth radar signals by constructing a GAN network embedded with a lightweight UNet-like network. On one hand, the lightweight UNet-like feature extraction network automatically extracts features from the original echo signal, promoting phase matching between sub-bands and avoiding processes such as signal model building and phase registration, making it simpler and more efficient than traditional algorithms. On the other hand, the GAN network significantly improves the accuracy of the fused signal through the game between the generator and discriminator. Furthermore, this GAN network can perform denoising while expanding bandwidth, enabling the synthesis of high-bandwidth signals even when the original signal-to-noise ratio is low. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 A flowchart illustrating a bandwidth extension method based on generative adversarial networks provided in an embodiment of the present invention;

[0047] Figure 2 This is an example diagram of a network structure for a generator provided in an embodiment of the present invention;

[0048] Figure 3 This is an example diagram of the network structure of a discriminator provided in an embodiment of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] This application provides a bandwidth extension method, system, storage medium, and electronic device based on generative adversarial networks, which solves the technical problem of reduced synthesis quality under low signal-to-noise ratio conditions.

[0051] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows:

[0052] Generative Adversarial Networks (GANs) have demonstrated powerful technical advantages in image generation and speech synthesis. Compared to traditional signal processing methods that can only extend existing frequency content, GAN networks can regenerate missing frequencies. Inspired by a series of lightweight convolutional encoder-decoder architectures, this invention creates a novel deep learning model designed to infer missing portions from radar signals containing only a subset of frequency bands, achieving ultra-bandwidth extension.

[0053] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0054] Example 1:

[0055] like Figure 1 As shown, this embodiment of the invention provides a bandwidth extension method based on generative adversarial networks, including:

[0056] S1. Obtain radar measurement values ​​that are angularly coherent within the target's full frequency band as true full-band data, and filter the true full-band data using different bandpass filters to obtain sub-band data with intervals in multiple frequency band coverage areas.

[0057] S2. Obtain the pre-built GAN network; wherein the GAN network includes a generator employing a lightweight encoder-decoder network similar to UNet and a discriminator employing a multi-scale architecture; perform the following training steps:

[0058] S3. Gaussian white noise is added to the multiple sub-frequency band data for enhancement processing and used as input to the generator to generate pseudo full-frequency band data;

[0059] S4. Use the pseudo full-band data and the real full-band data as inputs to the discriminator to determine the authenticity of the full-band data;

[0060] S5. If the GAN network meets the training completion conditions, the multiple sub-band data to be expanded are used as input to the trained generator to obtain the required full-band data.

[0061] In an optional implementation, the bandwidth extension method provided in this embodiment of the invention further includes:

[0062] S6. If the GAN network does not meet the training completion conditions, adjust the relevant parameters in the GAN network to optimize the loss function, and continue the above training steps based on the adjusted GAN network; wherein adjusting the relevant parameters in the GAN network to optimize the loss function includes:

[0063] Keeping the parameters of the generator unchanged, the discriminator loss is minimized by adjusting the parameters of the corresponding discriminator.

[0064] Keeping the parameters of the discriminator unchanged, the generator loss is minimized by adjusting the parameters of the corresponding generator.

[0065] The discriminator loss and the generator loss are alternately optimized until the network converges.

[0066] This invention employs a GAN network with embedded UNet-like features to fuse multi-band radar signals, generating ultra-wideband radar signals, thereby improving the range, azimuth, and imaging resolution of the radar system. The designed GAN network uses an end-to-end approach for multi-frequency signal fusion, eliminating the need for target echo modeling and multi-band signal coherent registration, thus reducing algorithm complexity. Furthermore, the use of a lightweight UNet-like network for feature extraction significantly reduces computational load and improves real-time performance.

[0067] The following will detail each step of the above solution:

[0068] In step S1, radar measurement values ​​that are angularly coherent across the entire frequency band of the target are acquired as true full-band data. The true full-band data are then filtered using different bandpass filters to obtain sub-band data with intervals in the coverage areas of multiple frequency bands.

[0069] It should be noted that the embodiments of the present invention do not limit the specific numerical information such as the target full frequency band range, the number of bandpass filters, or the size of the interval between frequency band coverage areas. Those skilled in the art can choose according to actual needs.

[0070] For example, the target full-band range is set to 7-17GHz, and three bandpass filters with a bandwidth of 1G are selected, with frequency band coverage ranges of 7-8GHz, 11-12GHz, and 16-17GHz, respectively. The real full-band data is filtered to obtain sub-band data with intervals between the three frequency band coverage ranges.

[0071] In step S2, the pre-built GAN network is obtained.

[0072] The pre-constructed GAN network in this embodiment of the invention mainly includes two parts: a generator and a discriminator.

[0073] The generator includes a UNet-like network for automatic signal feature extraction. The UNet-like network is represented by a U-shaped structure consisting of multiple encoders and decoders of equal number, used for cross-layer fusion of the extracted features with the input signal. The encoders and decoders include convolutional layers and residual modules.

[0074] It should be noted that the embodiments of the present invention do not limit the specific number of encoders and decoders, and those skilled in the art can set them according to actual needs.

[0075] For example, such as Figure 2 As shown, the UNet-like network is represented by a U-shaped structure consisting of three encoders and three decoders. The encoders and decoders mainly consist of convolutional layers and residual modules, and the activation function can be set to Leaky ReLU with a slope of -0.01.

[0076] The discriminator adopts a multi-scale architecture, which consists of at least three sub-discriminators. The first sub-discriminator is used to receive pseudo full-band data generated by the generator, and the remaining sub-discriminators are used to receive pseudo full-band data after downsampling according to the corresponding preset multiple.

[0077] It should be noted that, given that the discriminator consists of at least three sub-discriminators, the embodiments of the present invention do not limit the specific number of sub-discriminators, and those skilled in the art can set them according to actual needs.

[0078] For example, such as Figure 3 As shown, the discriminator consists of three sub-discriminators, D1, D2, and D3, to learn the characteristics of different sub-bands, making the synthesized broadband signal closer to the real full-band signal. D1 deals with the original signal (i.e., the pseudo-full-band data generated by the generator), while D2 and D3 deal with the data after downsampling the original signal by multiples of 2 and 4, respectively. The downsampling is performed using average pooling with a kernel size of 4. Leaky ReLU can also be used as the activation function in the discriminator, but a stronger slope of -0.2 is preferred to allow better gradient transmission to the generator. All convolutional layers in the network have normalized weights to ensure fast convergence during training.

[0079] In step S3, Gaussian white noise is added to the multiple sub-band data for enhancement processing and used as input to the generator to generate pseudo full-band data.

[0080] Before the multiple sub-band data obtained in this step are used as input to the GAN network, noise is added to the data to enable the deep learning model to perform denoising while expanding bandwidth. In particular, using the original waveform as input can minimize manual processing and allow the network to construct its representation itself, thus achieving phase registration between sub-bands.

[0081] In step S4, the pseudo full-band data and the real full-band data are used as inputs to the discriminator to determine the authenticity of the full-band data.

[0082] In step S5, if the GAN network meets the training completion condition, the multiple sub-band data to be expanded are used as input to the trained generator to obtain the required full-band data.

[0083] It should be noted that, in the embodiments of the present invention, the training completion conditions include at least one of the following: the number of training iterations reaches a predetermined iteration threshold, the loss function reaches the optimal value, or the discrimination accuracy of the discriminator is within a predetermined range.

[0084] In step S6, if the GAN network does not meet the training completion conditions, the relevant parameters in the GAN network are adjusted to optimize the loss function, and the above training steps are continued based on the adjusted GAN network.

[0085] If the training completion condition is not met in this step, the relevant parameters of the generator and / or discriminator in the GAN network are adjusted to optimize the loss function. Then, the adjusted GAN network continues to execute the above training steps until the training completion condition is met.

[0086] The loss function includes discriminator loss and generator loss. Adjusting the relevant parameters in the GAN network to optimize the loss function includes:

[0087] Keeping the parameters of the generator unchanged, the discriminator loss is minimized by adjusting the parameters of the corresponding discriminator.

[0088] Keeping the parameters of the discriminator unchanged, the generator loss is minimized by adjusting the parameters of the corresponding generator.

[0089] The discriminator loss and the generator loss are alternately optimized until the network converges (which can also be understood as the generator and discriminator reaching a Nash equilibrium).

[0090] Alternating optimization refers to: first, keeping the parameters of the generator / discriminator unchanged, and repeatedly executing steps S3 to S6 by adjusting the parameters of the corresponding discriminator / generator until the values ​​of the discriminator loss and generator loss stabilize; then, performing the reverse setting, that is, keeping the parameters of the discriminator / generator unchanged, and repeatedly executing steps S3 to S6 by adjusting the parameters of the corresponding generator / discriminator until the values ​​of the generator loss and discriminator loss stabilize. The parameters of the generator and discriminator are trained alternately until they reach Nash equilibrium.

[0091] The process of obtaining the discriminator loss and generator loss is described below:

[0092] In this embodiment of the invention, both the discriminator loss and the generator loss are based on the classic hinge loss function. In particular, the generator loss function adds a feature matching loss to the hinge loss, which can make the end-to-end synthesized signal closer to the full-band label signal.

[0093] The discriminator loss is expressed as:

[0094]

[0095] Among them, L D E represents the discriminator loss; y represents the true full-band data; E represents the discriminator loss. y This represents the average of all real full-band data in the training batch; K represents the number of discriminators, N k D represents the total number of output elements of the k-th discriminator; max represents maximizing; k,n (y) represents the nth value of the output vector of the k-th discriminator when the input is y; x represents the multiple sub-band data after enhancement processing, E x G(x) represents the average of multiple sub-band data after enhancement processing in the training batch; G(x) represents the pseudo-full-band data generated by the generator when the input is x; D k,n (G(x)) represents the nth value of the output vector of the kth discriminator when the input is x.

[0096] The generator loss is expressed as:

[0097]

[0098] Among them, L G Indicates generator loss; Represents the hinge loss term; Let N represent the feature matching loss term, where λ represents the weight of this term; k,l F represents the total number of elements in a single feature of the l-th layer of the k-th discriminator; k,l D represents the number of features in the l-th layer of the k-th discriminator; ||·|| represents the L2 norm; D k,n (G(x)) represents the feature vector of the l-th layer of the k-th discriminator when the input is G(x); D k,n (x) represents the feature vector of the l-th layer of the k-th discriminator when the input is x.

[0099] Furthermore, during the training process of the GAN network:

[0100] For example, all training data can be divided into a training set:test set ratio of 5:1, with a batch size of 32. Within each batch, the following steps are followed: Figure 2 The provided example alternately trains the discriminator set {D} kThe algorithm uses a discriminator loss and a generator G (k = 0, 1, or 2) to minimize the discriminator loss and generator loss. The optimizer can employ the Adam optimization algorithm, which adaptively adjusts the learning rate, with a learning rate of 3e-4, a first-order moment decay rate of 0.5, and a second-order moment decay rate of 0.9.

[0101] Thus, this embodiment of the invention completes the entire process of the bandwidth extension method based on generative adversarial networks.

[0102] Example 2:

[0103] This invention provides a bandwidth extension system based on generative adversarial networks, comprising:

[0104] The sample acquisition module is used to acquire radar measurement values ​​that are angularly coherent within the target's full frequency band as real full-band data, and to filter the real full-band data using different bandpass filters to obtain sub-band data with intervals in multiple frequency band coverage areas.

[0105] A network acquisition module is used for a pre-built GAN network; wherein the GAN network includes a generator employing a lightweight encoder-decoder network similar to UNet and a discriminator employing a multi-scale architecture;

[0106] The data generation module is used to add Gaussian white noise to multiple sub-frequency band data for enhancement processing and use it as input to the generator to generate pseudo full-frequency band data;

[0107] The authenticity judgment module is used to use the pseudo full-band data and the real full-band data as inputs to the discriminator to judge the authenticity of the full-band data.

[0108] The frequency band extension module, if the GAN network meets the training completion conditions, uses the multiple sub-frequency band data to be extended as input to the trained generator to obtain the required full-frequency band data.

[0109] In an optional implementation, the bandwidth extension system provided in this embodiment of the invention further includes:

[0110] The parameter adjustment module is used to adjust the relevant parameters in the GAN network to optimize the loss function if the GAN network does not meet the training completion conditions, and to continue the above training steps based on the adjusted GAN network; wherein adjusting the relevant parameters of the GAN network to optimize the loss function includes:

[0111] Keeping the parameters of the generator unchanged, the discriminator loss is minimized by adjusting the parameters of the corresponding discriminator.

[0112] Keeping the parameters of the discriminator unchanged, the generator loss is minimized by adjusting the parameters of the corresponding generator.

[0113] The discriminator loss and the generator loss are alternately optimized until the network converges.

[0114] Example 3:

[0115] This invention provides a storage medium storing a computer program for bandwidth extension based on generative adversarial networks, wherein the computer program causes a computer to execute the bandwidth extension method as described in Embodiment 1.

[0116] Example 4:

[0117] This invention provides an electronic device, comprising:

[0118] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the bandwidth extension method as described in Embodiment 1.

[0119] It is understood that the bandwidth extension system, storage medium and electronic device based on generative adversarial networks provided in the embodiments of the present invention correspond to the bandwidth extension method based on generative adversarial networks provided in the embodiments of the present invention. The explanation, examples and beneficial effects of the relevant contents can be referred to the corresponding parts in the deep bandwidth extension method, and will not be repeated here.

[0120] In summary, compared with existing technologies, it has the following beneficial effects:

[0121] This invention achieves the fusion of multi-band radar signals and generates high-bandwidth radar signals by constructing a GAN network embedded with a lightweight UNet-like network. On one hand, the lightweight UNet-like feature extraction network automatically extracts features from the original echo signal, promoting phase matching between sub-bands and avoiding processes such as signal model building and phase registration, making it simpler and more efficient than traditional algorithms. On the other hand, the GAN network significantly improves the accuracy of the fused signal through the interplay between the generator and discriminator. Furthermore, this GAN network can perform denoising while expanding bandwidth, enabling the synthesis of high-bandwidth signals even when the original signal-to-noise ratio is low.

[0122] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0123] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A bandwidth extension method based on generative adversarial networks, characterized in that, include: The radar measurement values ​​that are angularly coherent within the target's full frequency band are obtained as true full-band data. The true full-band data are then filtered using different bandpass filters to obtain sub-band data with intervals in the coverage of multiple frequency bands. Obtain a pre-built GAN network; wherein the GAN network includes a generator employing a lightweight encoder-decoder network similar to UNet and a discriminator employing a multi-scale architecture; perform the following training steps: Gaussian white noise is added to the multiple sub-band data for enhancement processing and used as input to the generator to generate pseudo full-band data; The pseudo full-band data and the real full-band data are used as inputs to the discriminator to determine the authenticity of the full-band data. If the GAN network meets the training completion condition, the multiple sub-band data to be expanded are used as input to the trained generator to obtain the required full-band data. The generator loss is expressed as: in, Indicates generator loss; Indicates the number of discriminators. Indicates the first The total number of output elements of each discriminator; Indicates maximization; Represents the hinge loss term; This represents the feature matching loss term. This indicates the weight of the item; This represents the enhanced data from multiple sub-bands. This represents the average of multiple sub-band data after all enhancement processing in the training batch; Indicates the first The first discriminant The total number of elements in a single feature of a layer; Indicates the first The first discriminant Number of layer features; Represents the L2 norm; Indicates that the input is At that time, the generator produces pseudo full-band data; Indicates that the input is At that time, the first The output vector of the discriminator is the first... One value; Indicates that the input is At that time, the first The first discriminant The first layer of feature vectors One value; Indicates that the input is At that time, the first The first discriminant The first layer of feature vectors One value; The generator includes a UNet-like network, which is represented by a U-shaped structure consisting of multiple encoders and decoders of equal number. The encoders and decoders include convolutional layers and residual modules.

2. The bandwidth extension method as described in claim 1, characterized in that, If the GAN network does not meet the training completion conditions, the relevant parameters in the GAN network are adjusted to optimize the loss function, and the above training steps are continued based on the adjusted GAN network; wherein adjusting the relevant parameters of the GAN network to optimize the loss function includes: Keeping the parameters of the generator unchanged, the discriminator loss is minimized by adjusting the parameters of the corresponding discriminator. Keeping the parameters of the discriminator unchanged, the generator loss is minimized by adjusting the parameters of the corresponding generator. The discriminator loss and the generator loss are alternately optimized until the network converges.

3. The bandwidth extension method as described in claim 2, characterized in that, The discriminant loss is expressed as: in, Indicates discriminator loss; Represents true full-band data. This represents the average of all real full-band data in the training batch; Indicates that the input is At that time, the first The output vector of the discriminator is the first... Values.

4. The bandwidth extension method as described in claim 1, characterized in that, The discriminator consists of at least three sub-discriminators. The first sub-discriminator is used to receive pseudo full-band data generated by the generator, and the remaining sub-discriminators are used to receive pseudo full-band data after downsampling according to a corresponding preset multiple.

5. A bandwidth extension system based on generative adversarial networks, characterized in that, For performing the bandwidth extension method as described in claim 1, comprising: The sample acquisition module is used to acquire radar measurement values ​​that are angularly coherent within the target's full frequency band as real full-band data, and to filter the real full-band data using different bandpass filters to obtain sub-band data with intervals in multiple frequency band coverage areas. A network acquisition module is used for a pre-built GAN network; wherein the GAN network includes a generator employing a lightweight encoder-decoder network similar to UNet and a discriminator employing a multi-scale architecture; The data generation module is used to add Gaussian white noise to multiple sub-frequency band data for enhancement processing and use it as input to the generator to generate pseudo full-frequency band data; The authenticity judgment module is used to use the pseudo full-band data and the real full-band data as inputs to the discriminator to judge the authenticity of the full-band data. The frequency band extension module, if the GAN network meets the training completion conditions, uses the multiple sub-frequency band data to be extended as input to the trained generator to obtain the required full-frequency band data.

6. The bandwidth extension system as described in claim 5, characterized in that, include: The parameter adjustment module is used to adjust the relevant parameters in the GAN network to optimize the loss function if the GAN network does not meet the training completion conditions, and to continue the above training steps based on the adjusted GAN network; wherein adjusting the relevant parameters of the GAN network to optimize the loss function includes: Keeping the parameters of the generator unchanged, the discriminator loss is minimized by adjusting the parameters of the corresponding discriminator. Keeping the parameters of the discriminator unchanged, the generator loss is minimized by adjusting the parameters of the corresponding generator. The discriminator loss and the generator loss are alternately optimized until the network converges.

7. A storage medium, characterized in that, It stores a computer program for bandwidth extension based on generative adversarial networks, wherein the computer program causes the computer to perform the bandwidth extension method as described in any one of claims 1 to 4.

8. An electronic device, characterized in that, include: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the bandwidth extension method as described in any one of claims 1 to 4.

Images