A Radar Target Detection Method for External Radiation Sources Based on Feature Parallax Augmentation Network

Abstract

This invention discloses a target detection method for external radiation source radar based on a feature disparity augmentation network, comprising: acquiring a target RD spectrum based on the external radiation source radar; and performing sliding window detection on the target RD spectrum based on the feature disparity augmentation network to determine whether the target is within the target RD spectrum in the current sliding window. The feature disparity augmentation network performs target detection based on the stereo disparity features of the target RD spectrum, which are obtained by splicing the average energy features and peak energy features of the target RD spectrum. This invention enables target detection using external radiation source radar based on neural networks and improves detection performance.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing technology, specifically relating to a radar target detection method for external radiation sources based on a feature parallax amplification network. Background Technology

[0002] In the field of radar target detection, traditional methods are mainly based on constant false alarm rate (CFAR) detectors. However, in real-world environments with complex and changing background electromagnetic environment and poor target echo signal quality, the target echo features are easily masked by background noise, making it difficult for CFAR detectors to effectively extract target echo features, resulting in a significant decrease in detection performance.

[0003] In recent years, many advanced neural network models have demonstrated excellent performance in tasks such as image recognition and sequence information processing. This has inspired researchers to integrate these methods into radar target detection tasks to achieve better detection results. For neural network detectors, after training with sufficient relevant sample pairs, they can better extract target features under low signal-to-noise ratio conditions and achieve higher probability detection of weak-energy targets that are difficult to observe, thus compensating for the shortcomings of traditional CFAR detectors and demonstrating the advantages of neural network methods over traditional methods. However, traditional neural network-based target recognition methods mainly use millimeter-wave radar and synthetic aperture radar signals as input, and there are few methods that use neural networks to process external radiation source radar signals for target detection. This is because, compared to ordinary radar, the target echoes and clutter acquired by external radiation source radar are more unstable, posing a greater challenge to the neural network in performing its tasks.

[0004] Therefore, there is an urgent need for a target detection method based on neural networks for radar with external radiation sources. Summary of the Invention

[0005] This invention provides a target detection method for external radiation source radar based on a feature parallax amplification network, which can solve the problem of poor detection effect caused by unstable echo when the traditional technology is used to detect targets based on external radiation source radar signals.

[0006] In a first aspect, embodiments of the present invention provide a method for detecting external radiation source radar targets based on a feature parallax amplification network, the method comprising: Acquiring the RD spectrum of the target based on external radiation source radar; Based on the feature disparity augmentation network, sliding window detection is performed on the RD spectrum to be tested to determine whether the target to be tested is in the RD spectrum to be tested within the current sliding window. The feature disparity augmentation network performs target detection based on the stereo disparity features of the RD spectrum to be tested. The stereo disparity features are obtained by splicing the average energy features and peak energy features of the RD spectrum to be tested.

[0007] Secondly, embodiments of the present invention provide an external radiation source radar target detection device based on a feature parallax amplification network, including an acquisition unit and a sliding window processing unit; The acquisition unit is used to acquire the RD spectrum to be measured based on the external radiation source radar; The sliding window processing unit is used to perform sliding window detection on the RD spectrum to be tested based on the feature disparity augmentation network, so as to determine whether the target to be tested is in the RD spectrum to be tested within the current sliding window; The feature disparity augmentation network performs target detection based on the stereo disparity features of the RD spectrum to be tested. The stereo disparity features are obtained by splicing the average energy features and peak energy features of the RD spectrum to be tested.

[0008] Thirdly, embodiments of the present invention provide an electronic device, including a processor and a memory, wherein the memory is used to store a computer program; the processor can be used to execute a calculator program (instructions) stored in the memory to implement the method of the first aspect described above.

[0009] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed, can implement the method described in the first aspect above.

[0010] It is understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here.

[0011] The beneficial effects of this invention compared to the prior art are as follows: This invention extracts the average energy feature and peak energy feature of the RD spectrum as primary features. The average energy feature can reflect the overall level of local energy, forming a relatively stable background characterization, while the peak energy feature can retain the local extreme value information of the energy distribution, making those target peaks that are weak in energy but have significant shapes stand out. Then, these two features are spliced ​​together for sliding window target detection, so that the stereo parallax feature can simultaneously contain stable background information and prominent target information. The sliding window detection mechanism allows the detection and analysis process of the feature parallax amplification network to be repeated in different local regions of the RD spectrum, adapting to the echo changes caused by the uncertainty of the target position; thereby improving the detection effect. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the structure of a feature disparity augmentation network provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a ResNet module provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the implementation process of a training method for a feature disparity augmentation network provided by the present invention. Figure 4 A schematic diagram of a training scenario for a feature disparity augmentation network provided in an embodiment of the present invention; Figure 5 A schematic diagram illustrating the implementation process of an external radiation source radar target detection method based on a feature parallax amplification network provided in this embodiment of the invention; Figure 6 A schematic diagram of a target detection scenario provided by an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of an external radiation source radar target detection device based on a feature parallax amplification network, provided in an embodiment of the present invention. Figure 8 A schematic diagram of an RD spectrum after inversion processing provided in an embodiment of the present invention; Figure 9 A schematic diagram of the RD spectrum after masking processing provided in an embodiment of the present invention; Figure 10 This is a schematic diagram showing the difference in the distribution of features extracted from the backbone network before and after training, after clustering, as provided in an embodiment of the present invention. Figure 11a -d is a schematic diagram of the detection results of different frame RD spectra provided in an embodiment of the present invention; Figure 12 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0013] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0014] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0015] It should also be understood that the term “and / or” as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0016] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0017] Furthermore, in the description of this invention and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0018] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of the invention include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0019] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0020] Example 1 Figure 1 The diagram shown illustrates the structure of a feature disparity augmentation network according to an embodiment of the present invention. As an example and not a limitation, the feature disparity augmentation network may include a backbone network 110, a neck network 120, and a classifier 130.

[0021] In some embodiments, the backbone network 110 can extract features from the input RD spectrum to obtain preliminary stereo features (including average energy features and peak energy features); then, the neck network 120 further extracts stereo disparity features to obtain deep features; finally, the classifier 130 separates the background information and target information of the RD spectrum based on the deep features to complete target detection.

[0022] For example, stereo parallax features are obtained by stitching together average energy features and peak energy features.

[0023] In one possible implementation, the backbone network 110 may include a random dropout layer and two parallel feature channels. After the RD spectrum is processed by the random dropout layer with 10% of the data randomly zeroed out, it is then processed by the two internal feature channels to extract the average energy features (see...). Figure 1 (flatten1) and peak energy characteristics (see [reference]). Figure 1 (flatten2 in the middle).

[0024] For example, the input size of the RD spectrum can be 200×51.

[0025] In one example, the peak feature channel in the backbone network 110 can be based on a max pooling layer, which focuses on the peak energy information in the RD spectrum by selecting the maximum value of the elements in the unit, and extracting the peak energy features.

[0026] For example, see Figure 1 Peak feature channels can include cascaded convolutional layers, batch normalization layers, activation layers, max pooling layers, and residual blocks. For example, the pooling process of the max pooling layer can be represented as:

[0027] in, Indicates the ( , ) output pooling windows, Indicates the ( , ) input pooling windows, , These represent the height and width of the pooling window, respectively. This is the stride size of the pooling window. , These are the row and column traversal indices for the pooled window, respectively.

[0028] Optionally, the feature map size of the max / average pooling layer output can be 100×26.

[0029] In one example, the average feature channel in the backbone network 110 can be based on the average pooling layer (AvgPool) to focus on the average energy information of the RD spectrum by selecting the average value of all elements in the unit, and extract the average energy feature.

[0030] For example, the average feature channel can be composed of convolutional layers, batch normalization layers, activation layers, average pooling layers, and residual blocks.

[0031] Optionally, the aforementioned convolutional layers, batch normalization layers, activation layers, and residual blocks can respectively employ conv2d convolutional layers, batchnorm2d layers, ReLU activation function layers, and 4-layer ResNet modules (see [link to documentation]). Figure 2 ).

[0032] Specifically, in the ResNet module, Input represents the input, Conv represents the convolutional layer, BN represents the batch normalization layer, ReLU represents the ReLU activation function layer, and AdaptiveAvgPooling represents the adaptive average pooling layer.

[0033] For example, the pooling process of the average pooling layer can be represented as: .

[0034] Optionally, the backbone network 110 can choose different architectures depending on the task being performed. For example, when the feature parallax augmentation network is used in other complex scenarios, the backbone network 110 can adopt architectures such as ResNet34 or conventional CNN.

[0035] It should be understood that after replacing the backbone network 110, the input size of the neck network should also be modified to ensure the normal transmission of network parameters, so that the feature parallax augmentation network can adapt to the feature extraction needs in different scenarios.

[0036] In one possible implementation, see Figure 1 The neck network 120 can adopt the BNNeck architecture.

[0037] Specifically, the neck network 120 can be composed of a fully connected layer, a batch normalization layer, a ReLU activation layer, and a random deactivation layer cascaded together.

[0038] For example, the stereo disparity feature of size 1×1024 obtained by horizontally concatenating two 1×512 features output by the backbone network 110 can be input into the neck network 120, and the deep feature output by the neck network 120 can be 1×510.

[0039] Classifier 130 performs target detection based on deep features processed by a random deactivation layer, which can enhance the training effect.

[0040] In one possible implementation, the classifier 130 can consist of cascaded fully connected layers (FC) and activation layers (Softmax).

[0041] Specifically, classifier 130 can output the probability that a target exists in the RD spectrum.

[0042] The feature disparity network provided by this invention extracts the average energy feature and peak energy feature of the RD spectrum through average pooling and max pooling, respectively. Since the features extracted by average pooling and max pooling are inherently different, compared with the structure where the two branches use the same pooler, the two channels are more likely to form extracted features with a larger distribution difference, thus achieving a better stereo perception effect. This enables the backbone network to provide higher quality and stronger representational features for subsequent classification, improving the robustness and completeness of the decision basis of the classifier when detecting objects.

[0043] Example 2 The method provided in this embodiment of the invention can be applied to electronic devices such as mobile terminals, personal laptops, and supercomputers. This embodiment of the invention does not impose any restrictions on the specific type of electronic device.

[0044] Figure 3 The diagram shown illustrates the implementation flow of a training method for a feature disparity augmentation network provided by the present invention. As an example and not a limitation, the method may include steps S301-S310, which are described below.

[0045] S301, input the RD spectrum of the current round of samples into the backbone network after the previous round of training to obtain the average energy feature and peak energy feature of the current round of samples.

[0046] In one example, the backbone network 110 can be trained on two feature channels using only unlabeled RD spectra through unsupervised learning, enabling the two branches to extract features from different perspectives. This process of expanding the difference in features extracted between the two channels through multiple rounds of training to obtain features with greater perspective differences is also known as the feature parallax amplification mechanism.

[0047] Optionally, before inputting the samples into the backbone network, the RD spectrum of the N×M samples can be augmented by flipping or masking to improve the robustness of the network.

[0048] S302, determine the maximum mean difference loss for this round based on the average energy characteristics and peak energy characteristics of the samples in this round.

[0049] In one possible implementation, the maximum mean discrepancy loss can be constructed using the reciprocal of the maximum mean discrepancy (MMD). Thus, during iteration, as the error decreases, the MMD value increases, gradually widening the distribution difference between the peak features extracted from the two channels and the background environment features. This allows for the extraction of perturbation features of the target from two different perspectives, forming a three-dimensional perception mechanism for the network.

[0050] In one example, the loss function used to calculate the maximum mean loss can be expressed as:

[0051] in, Represents the set of features of average energy of samples and sample peak energy feature set The maximum mean difference loss between them The multiplication coefficient is... for , The maximum mean difference between them.

[0052] For example, the basic idea of ​​MMD is to compare two distributions using mean embeddings in the Reproducing Kernel Hilbert Space (RKHS). According to the definition of MMD, It can be calculated using the following formula:

[0053] in, For mapping functions, The MMD distance is represented by The data is obtained by mapping it to a regenerated Hilbert space for measurement.

[0054] Furthermore, expanding the above formula yields:

[0055] If introduced The kernel function is used to express the inner product of mapping functions, i.e. And the Gaussian kernel function will be used as the kernel function. We can obtain: .

[0056] Accordingly, the training objective of the backbone network can be expressed as:

[0057] in, The training objective for the backbone network is... Indicates the first Average energy characteristics of each sample Indicates the first Peak energy characteristics of each sample express , The maximum mean difference loss between them.

[0058] S303, determine whether the first convergence condition is met.

[0059] In one example, if the first convergence condition is not met, step S304 can be performed to update the backbone network under the constraint of the maximum mean difference loss, resulting in the backbone network after this round of training. Afterwards, step S301 can be returned to continue training the backbone network 110.

[0060] For example, the first convergence condition can be that the loss is less than a threshold or the number of training epochs is greater than a threshold.

[0061] In another example, if the convergence condition is met, the training of the backbone network 110 can be completed, and the trained backbone network 110 can be obtained. Then, the following step S305 is performed.

[0062] S304 updates the backbone network under the constraint of maximum mean difference loss, resulting in the backbone network after this round of training.

[0063] S305 inputs the stereo disparity features of the current round of samples into the neck network trained in the previous round to obtain the deep features of the current round.

[0064] In one possible implementation, the neck network 120 and classifier 130 can be trained in a supervised manner using labeled sample RD spectra. By analyzing the stereo disparity features of the samples, the characteristics of the target and background clutter signals are learned, ultimately achieving target detection.

[0065] For example, the sample stereo disparity features can be obtained by splicing the sample average energy features and sample peak energy features extracted by the trained backbone network 110.

[0066] Specifically, see Figure 4 The network parameters of the trained backbone network can be fixed. Then, the labeled RD spectrum is input into the trained backbone network to extract the sample average energy feature and the sample peak energy feature. Then, the sample average energy feature and the sample peak energy feature are concatenated to obtain the sample stereo disparity feature.

[0067] S306: Input the deep features of this round into the classifier trained in the previous round to obtain the target prediction position of this round.

[0068] S307, Calculate the cross-entropy loss for this round based on the predicted and actual target positions.

[0069] In one example, the loss function used to calculate the cross-entropy loss can be expressed as:

[0070] in, Represents cross-entropy loss, For the target's true location, This represents the probability that a target exists in the sample RD spectrum output by the network.

[0071] For example, the training objective of the second stage (i.e., the training objective of the neck network and classifier) ​​can be expressed as:

[0072] in, The total number of samples, This represents the predicted probability of the target's existence, determined based on the stereo disparity features of the sample. True label of sample stereo parallax features Determined cross-entropy loss.

[0073] S308, determine whether the second convergence condition is met.

[0074] In one example, if the second convergence condition is not met, step S309 can be performed, and then the process can return to step S305 for the next round of training.

[0075] Similarly, the second convergence condition can be that the loss is less than a threshold or the number of training epochs is greater than a threshold.

[0076] In another example, if the convergence condition is met, step S210 can be performed to stop training and output the neck network 120 and classifier 130 trained in the previous round as the trained neck network 120 and classifier 130, thereby obtaining the trained feature disparity augmentation network.

[0077] S309, under the constraint of cross-entropy loss, update the network parameters of the neck network and classifier after the previous round of training to obtain the neck network and classifier after the current round of training.

[0078] S310, stop training, and use the neck network and classifier from the previous training round as the output of the trained neck network and classifier to obtain the trained feature disparity augmentation network.

[0079] This invention uses maximum mean difference loss to iteratively train the backbone network, which makes the two types of features output by the backbone network more closely resemble the ideal data pattern in terms of statistical properties, thereby enhancing the stability and discriminative power of feature representation. By decoupling feature learning from classification decision through a two-stage training strategy, it can avoid gradient imbalance or convergence instability problems that may be caused by loss function competition in end-to-end training, making the optimization process of each module of the entire network more coordinated.

[0080] Example 3 Figure 5The diagram illustrates the implementation flow of an external radiation source radar target detection method based on a feature parallax amplification network according to an embodiment of the present invention. This is an example, not a limitation, and can be applied to the aforementioned electronic device. The method may include steps S501-S502, which are described below.

[0081] S501 acquires the RD spectrum of the target object based on external radiation source radar.

[0082] For example, see Figure 6 The echo signal acquired by the external radiation source radar and the reference signal can be mutually fuzzy and accumulated to obtain the RD spectrum to be measured.

[0083] S502, based on the feature disparity augmentation network, performs sliding window detection on the RD spectrum to be tested in order to determine whether the target to be tested is in the RD spectrum to be tested within the current sliding window.

[0084] In one example, see Figure 6 The test RD spectrum can be gridded, and a sliding window can be used to scan the gridded test RD spectrum. Part of the test RD spectrum scanned by the sliding window is input into a trained feature disparity augmentation network. The feature disparity augmentation network outputs the possible location and probability of the target as the detection result. Then, the detection results of each sliding window are integrated and stitched together to obtain the position of the target in the RD spectrum and visualize it.

[0085] For example, the size of the sliding window can be set to N×M, the horizontal step size to X, and the vertical step size to Y.

[0086] This invention extracts the average energy feature and peak energy feature of the RD spectrum as primary features. The average energy feature can reflect the overall level of local energy, forming a relatively stable background representation, while the peak energy feature can retain the local extreme value information of the energy distribution, making the target peaks that are weak in energy but have significant shapes stand out. Then, these two features are spliced ​​together for sliding window target detection, so that the stereo parallax feature can simultaneously contain stable background information and prominent target information. The sliding window detection mechanism allows the detection and analysis process of the feature parallax amplification network to be repeated in different local regions of the RD spectrum, adapting to the echo changes caused by the uncertainty of the target position, thereby improving the detection effect.

[0087] Example 4 Figure 7 The diagram shown illustrates the structure of an external radiation source radar target detection device based on a feature parallax amplification network, according to an embodiment of the present invention. As an example and not a limitation, the device may include an acquisition unit 710 and a sliding window processing unit 720.

[0088] For example, the acquisition unit 710 is used to acquire the RD spectrum to be measured based on the external radiation source radar, and the sliding window processing unit is used to perform sliding window detection on the RD spectrum to be measured based on the feature disparity amplification network to determine whether the target to be measured is in the RD spectrum to be measured within the current sliding window; wherein, the feature disparity amplification network is used to detect the target based on the stereo disparity features of the RD spectrum to be measured, and the stereo disparity features are obtained by splicing the average energy features and peak energy features of the RD spectrum to be measured.

[0089] This invention extracts the average energy feature and peak energy feature of the RD spectrum as primary features. The average energy feature can reflect the overall level of local energy, forming a relatively stable background representation, while the peak energy feature can retain the local extreme value information of the energy distribution, making the target peaks that are weak in energy but have significant shapes stand out. Then, these two features are spliced ​​together for sliding window target detection, so that the stereo parallax feature can simultaneously contain stable background information and prominent target information. The sliding window detection mechanism allows the detection and analysis process of the feature parallax amplification network to be repeated in different local regions of the RD spectrum, adapting to the echo changes caused by the uncertainty of the target position, thereby improving the detection effect.

[0090] To better illustrate the beneficial effects of the present invention, the following simulation experiments were conducted: Simulation Experiment 1 For example, in simulation experiment 1, to simulate the sliding window effect, the complete RD spectrum was cropped to 200×51, and samples containing targets were labeled "1", while those without targets were labeled "0". See also... Figure 8 The cropped RD spectrum can be directly augmented by flipping the x-axis, y-axis, and x+y-axis. Further augmentation is achieved through block partitioning and reassembly, simulating target behavior under more conditions to obtain enhanced sample RD spectra. During training, the network, faced with the enhanced data, can better extract target echo features from various sources and beams. See also... Figure 9 After the flipping operation, a masking mechanism was also introduced in the experiment to randomly set the unit elements in the RD spectrum matrix to zero at a certain ratio, thereby achieving the final data augmentation.

[0091] Specifically, see Figure 9During the masking process, the proportion of random zeroing can be set between 0.1% and 5%, ensuring that the target echo portion is not completely zeroed out. This adds extra interference to the samples, strengthening the network's memory of target features. Furthermore, due to the introduction of the random zeroing process, the overall energy of the RD spectrum will also fluctuate after zeroing. This also simulates, to some extent, the changes in target feature energy under scenarios with large energy variation, preventing the network from directly judging the existence of target echo features based solely on the overall energy of the RD spectrum, further enhancing the network's robustness and generalization ability.

[0092] For example, in simulation experiment 1, All values ​​are set to 1. Training uses the Adam optimizer with a learning rate of 0.001 and a batch size of 64. After 10 epochs of the first stage of training, the parameters of the first two channels of the network are saved and locked before proceeding to the second stage of training. In the second stage of training, the Adam optimizer is also used with a learning rate of 0.001 and a batch size of 64.

[0093] See Figure 10 The diagram shows the distribution difference of the features extracted by the backbone network before and after training. It can be seen that the constraint of the maximum mean difference loss makes the distribution difference between the two channels more obvious, which improves the network's three-dimensional perception of the energy features in the RD spectrum and achieves better detection results.

[0094] Referring to Table 1 below, which compares the performance of the method provided by this invention with those of the Cell Averaging Constant False Alarm Rate (CA-CFAR), Ordered Statistical Cell Averaging Constant False Alarm Rate (OS-CA-CFAR), and Ordered Statistical Constant False Alarm Rate (OS-CFAR) detectors in different signal-to-noise ratio (SNR) ranges, it can be seen that for measured data with an SNR higher than 15 dB, the detection rate of this invention is higher than that of the three traditional CFAR detectors. When the SNR is higher than 16 dB, the detection rate of the method provided by this invention approaches 100%, outperforming the other comparative methods. This indicates that the feature parallax augmentation network provided by this invention has better adaptability to complex environments in external radiation source radar scenarios.

[0095] Table 1. Measured Data and Detection Results

[0096] Simulation Experiment 2 For example, in simulation experiment 2, a larger RD spectrum sample before clipping is selected within a continuous time period, with each RD spectrum corresponding to a duration of 0.4s; the sliding window unit size is set to 200×51, the horizontal step size is set to 51, and the vertical step size is set to 200, and the target detection is performed using the method provided by this invention and the CA-CFAR detector.

[0097] See Figure 11a -d indicates the detection results of the RD spectrum in different frames, with the red box showing the location of the target peak. It can be seen that in the 2nd and 3rd frames ( Figure 11b In scenario c), the target echo energy fluctuates due to factors such as motion and background noise, resulting in insufficient accumulated gain within a specific timeframe. This prevents the CFAR detector from calculating a suitable threshold value and thus fails to detect the target. The method provided by this invention effectively extracts the weak target energy features, enabling target detection and filling the gap in the CFAR detector's detection capabilities. In practical applications, when facing similar weak target scenarios, if the CFAR detector starts detecting the target from the second frame, it cannot detect the target within the first 0.8 seconds due to the weak echo. However, by combining the method provided by this invention with the CAFR detector, the gap in the CFAR detector can be filled to some extent, allowing for earlier target detection and significantly improving early warning time.

[0098] Example 5 Figure 12 The diagram shown is a structural schematic of an electronic device provided in an embodiment of the present invention. Figure 12 The illustrated electronic device 1200 may include: at least one processor 1210 ( Figure 12 The diagram shows only one processor, a memory 1220, and a computer program 1230 stored in the memory 1220 and executable on the at least one processor 1210, which, when executing the computer program 1230, implements the steps in any of the above-described method embodiments.

[0099] The electronic device 1200 may be a robot or other processing device capable of implementing the above methods. This embodiment of the invention does not impose any restrictions on the specific type of electronic device.

[0100] Those skilled in the art will understand that Figure 12 This is merely an example of electronic device 1200 and does not constitute a limitation on the electronic device. It may include more or fewer components than shown, or combine certain components, or use different components. For example, the electronic device 1200 may also include input / output interfaces.

[0101] The processor 1210 may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASTCs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0102] In some embodiments, the memory 1220 may be an internal storage unit, such as a hard disk or RAM. In other embodiments, the memory 1220 may be an external storage device, such as a plug-in hard disk, a smart memory card (SMC), a secure digital card (SD), or a flash card. Furthermore, the memory 1220 may include both internal and external storage units. The memory 1220 is used to store the operating system, applications, boot loader, data, and other programs, such as the program code of the computer program. The memory 1220 can also be used to temporarily store data that has been output or will be output.

[0103] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0104] Those skilled in the art will understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the functions described above can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this invention. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0105] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0106] This invention provides a computer program product that, when run on an electronic device, enables the electronic device to perform the steps described in the various method embodiments above.

[0107] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0108] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0109] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

Claims

1. A radar target detection method for external radiation sources based on a feature parallax amplification network, characterized in that, include: Acquiring the RD spectrum of the target based on external radiation source radar; Based on the feature disparity augmentation network, sliding window detection is performed on the RD spectrum to be tested to determine whether the target to be tested is in the RD spectrum to be tested within the current sliding window. The feature disparity augmentation network performs target detection based on the stereo disparity features of the RD spectrum to be tested. The stereo disparity features are obtained by splicing the average energy features and peak energy features of the RD spectrum to be tested.

2. The method according to claim 1, characterized in that, The feature disparity augmentation network includes a backbone network, a neck network, and a classifier; The backbone network is used to extract the average energy feature and peak energy feature of the RD spectrum under test through two internal feature channels, respectively. The neck network is used to extract features based on the stereo disparity features to obtain the deep features of the RD spectrum to be tested; The classifier is used to separate the target information and background information of the RD spectrum to be tested based on the deep features, so as to perform target detection.

3. The method according to claim 2, characterized in that, The backbone network includes average feature channels and three-dimensional feature channels; The average feature channel is composed of cascaded convolutional layers, batch normalization layers, activation layers, average pooling layers, and residual blocks, while the stereo feature channel is composed of cascaded convolutional layers, batch normalization layers, activation layers, max pooling layers, and residual blocks.

4. The method according to claim 2, characterized in that, The training method for the feature disparity augmentation network includes: Under the constraint of maximum mean difference loss, the backbone network is trained iteratively for multiple rounds until a well-trained backbone network is obtained. Subsequently, under the constraint of cross-entropy loss, the neck network and classifier are trained in multiple rounds of iterations until a trained neck network and classifier are obtained; thus, a trained feature disparity augmentation network is obtained.

5. The method according to claim 4, characterized in that, The maximum mean difference loss satisfies the following formula: in: in, Represents the set of features of average energy of samples and sample peak energy feature set The maximum mean difference loss between them The multiplication coefficient is... for , The maximum mean difference between them for The total number of elements in , They represent The first in , Average energy characteristics of each sample for The total number of elements in , They represent The first in , Peak energy characteristics of each sample express, This is the kernel function.

6. The method according to claim 5, characterized in that, The training objective of the backbone network is expressed as: in, The training objective of the backbone network is... Indicates the first Average energy characteristics of each sample Indicates the first Peak energy characteristics of each sample express , The maximum mean difference loss between them.

7. A radar target detection device for external radiation sources based on a feature parallax amplification network, characterized in that, Includes an acquisition unit and a sliding window processing unit; The acquisition unit is used to acquire the RD spectrum to be measured based on the external radiation source radar; The sliding window processing unit is used to perform sliding window detection on the RD spectrum to be tested based on the feature disparity augmentation network, so as to determine whether the target to be tested is in the RD spectrum to be tested within the current sliding window; The feature disparity augmentation network performs target detection based on the stereo disparity features of the RD spectrum to be tested. The stereo disparity features are obtained by splicing the average energy features and peak energy features of the RD spectrum to be tested.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by an electronic device, it implements the method as described in any one of claims 1-6.

Images