Sub-domain adaptive based multi-channel synthetic aperture radar moving target detection method
By optimizing a multi-channel synthetic aperture radar system using a subdomain adaptive residual network, the problem of insufficient detection performance caused by the difference between simulation data and measured data is solved, and efficient detection of slow-moving ground targets in complex backgrounds is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
- Filing Date
- 2023-03-31
- Publication Date
- 2026-06-12
AI Technical Summary
Existing synthetic aperture radar systems are inadequate in detecting slow-moving ground targets in complex backgrounds, especially when there are differences between the distribution of simulated and measured data. Furthermore, they rely on the assumption of uniform clutter and are sensitive to the radial velocity of the target, making it difficult to accurately detect targets with tangential velocity.
A multi-channel synthetic aperture radar moving target detection method based on subdomain adaptation is adopted. The subdomain adaptive residual network is used to extract data features. The network is optimized by label cross-entropy loss and local maximum mean difference loss between simulated and measured data to reduce distribution differences and achieve efficient detection without the assumptions of uniform clutter and radial velocity.
It improves the detection performance of slow-moving ground targets in complex backgrounds, avoids gradient vanishing or exploding, and achieves high-accuracy moving target detection.
Smart Images
Figure CN116413723B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of ground slow target detection technology, specifically relating to a multi-channel synthetic aperture radar moving target detection method based on subdomain adaptive method. Background Technology
[0002] Synthetic Aperture Radar (SAR) provides high-resolution imagery in all weather conditions and plays a vital role in both military and civilian applications. Ground Moving Target Indication (GMTI), as an extension of SAR, is a research hotspot in radar signal processing. However, the detection performance of existing SAR / GMTI systems for slow-moving ground targets in complex backgrounds needs further improvement.
[0003] Based on the number of receiving channels, traditional SAR / GMTI systems can be divided into single-channel GMTI systems and multi-channel GMTI systems. Single-channel GMTI can detect moving targets on the ground through spectral filtering or time-frequency analysis, based on the differences between stationary clutter and moving targets in the Doppler domain. However, the clutter spectrum broadening caused by platform motion can overwhelm slow-moving targets, making them difficult to detect. Multi-channel GMTI increases the spatial dimension, enabling joint spatial and temporal processing, thus reducing the minimum detectable speed of the system. Common multi-access GMTI techniques include Phase Center Offset Antenna (DPCA), Track-Walk Interferometry (ATI), Clutter Suppression Interferometry (CSI), and Space-Time Adaptive Processing (STAP). While these methods achieve better moving target detection results than single-channel GMTI, they still have some problems in practical processing, such as relying on the assumption of uniform clutter that cannot be satisfied in reality, and being sensitive only to the radial velocity of the target and unable to detect targets with only tangential velocity.
[0004] Deep learning has achieved great success in target detection and has been incorporated into SAR / GMTI research due to its ability to automatically extract latent features and learn rules. In a military context, obtaining large amounts of labeled data is difficult in the radar field; therefore, simulation data is often used to support network training. Although the feasibility of this method has been verified in numerous practical applications, there are distributional differences between simulation data and real-world data. When networks trained on simulation data are used with real-world data, performance loss or even poor detection results often occur.
[0005] Transfer learning is a network framework that has emerged in recent years, utilizing various auxiliary data and external models to solve target problems. Domain adaptation is a representative method of transfer learning, typically based on aligning the distributions of two domains using maximum mean difference (MMD) to avoid network performance loss. After domain adaptation, although the global distribution is roughly the same, ground moving targets are still difficult to detect accurately. With further research, subdomain adaptation based on local maximum mean difference (LMMD) can make both local and global distributions approximately consistent. Subdomain adaptation methods have been applied to various fields such as human activity recognition, wind turbine system fault diagnosis, and bridge structure damage diagnosis, but currently have not been applied to ground moving target detection using synthetic aperture radar. Therefore, exploring its applications is necessary. Summary of the Invention
[0006] Purpose of the invention: To address the problem of difficulty in acquiring large amounts of labeled data in the radar field, this invention provides a moving target detection method for multi-channel synthetic aperture radar based on subdomain adaptive methods. This method does not rely on the assumption of uniform clutter or the existence of target radial velocity. It mainly extracts data features through subdomain adaptive residual networks and reduces the difference between simulated data and measured data, thereby improving the detection performance of slow-moving ground targets in complex backgrounds.
[0007] Technical Solution: This invention provides a multi-channel synthetic aperture radar moving target detection method based on subdomain adaptation, specifically including the following steps:
[0008] (1) Obtain simulation and measured data of n-channel synthetic aperture radar, and make training and verification datasets from the simulation and measured data; the training dataset contains simulated moving targets, measured clutter and measured data of unknown categories; the verification dataset contains measured clutter and measured moving targets, and adds real labels to clutter and moving targets;
[0009] (2) Preprocess the data in the training and validation datasets, treating the real and imaginary parts as independent channels and performing standardization;
[0010] (3) Based on ResNet18 and subdomain adaptation, a subdomain adaptive residual network for ground moving target detection is constructed. The preprocessed data is input into the subdomain adaptive residual network for training and validation, and the optimal subdomain adaptive residual network parameters are saved.
[0011] (4) Import the saved network parameters into the subdomain adaptive residual network, make the synthetic aperture radar data to be tested into a test dataset, preprocess the data of the test dataset and input it into the subdomain adaptive residual network, and window the output prediction results to obtain the ground moving target detection results.
[0012] Furthermore, the implementation process of step (1) is as follows:
[0013] (11) Set the radar system and moving target parameters, acquire n-channel radar simulation echo data, and obtain clutter-free simulated moving targets after channel alignment, imaging and image registration.
[0014] (12) Acquire n-channel radar measured echo data, and after channel alignment, imaging and image registration processing, obtain measured moving target, measured clutter and measured data of unknown type;
[0015] (13) The simulated moving target without clutter is superimposed with the measured clutter at a certain signal-to-clutter ratio to obtain the simulated moving target;
[0016] (14) Add labels 0 and 1 to clutter and moving targets respectively;
[0017] (15) The simulated moving target and some measured clutter constitute the labeled data of the source domain, and the measured data of unknown category constitute the unlabeled data of the target domain. Some measured clutter and measured moving target constitute the labeled data of the target domain.
[0018] Furthermore, the implementation process of step (2) is as follows:
[0019] Treating the real and imaginary parts of the data in the training and validation datasets as independent channels, the n-channel complex data is transformed into 2n-channel real data and standardized based on standard scores.
[0020] Furthermore, the implementation process of step (3) is as follows:
[0021] (31) Retain the convolutional blocks in ResNet18 and then connect them with three fully connected layers. Introduce subdomain adaptation to construct a subdomain adaptive residual network for ground moving target detection. The network input is 2n channels of data and the network output is 2 channels of predicted labels.
[0022] (32) During network training, the input includes labeled data and its true labels from the source domain and unlabeled data from the target domain; during network validation, the input includes labeled data and its true labels from the target domain, and the output is the predicted label.
[0023] (33) Convert the real labels of the data into one-hot encoded form, use the cross-entropy between the real labels and the predicted labels to measure the classification loss, and use the sum of the differences between the multi-kernel local maximum means of the three fully connected layers to measure the domain confusion loss.
[0024] (34) During training, the network loss is the classification loss of the source domain labeled data and the domain confusion loss regularization term; during validation, the network loss is the classification loss of the target domain labeled data.
[0025] Furthermore, step (34) is implemented by the following formula:
[0026]
[0027] Among them, Loss DC Loss due to domain obfuscation CFS For classification loss, λ is a regularization parameter representing the degree of domain confusion, J(·) is the cross-entropy loss function, and y s , z represents the true and predicted labels of the source domain data. s and z t The outputs of the fully connected layer are the source domain data and the target domain data, respectively. s The number of labeled data. This represents the local maximum mean difference between the two domains.
[0028] Beneficial Effects: Compared with existing technologies, the present invention offers the following advantages: It employs a subdomain adaptive residual network, achieving better feature representation while avoiding gradient vanishing or gradient explosion issues that may exist in deep networks. It does not rely on the assumption of uniform clutter or the existence of target radial velocity, effectively improving the network's detection performance. Furthermore, the invention utilizes simulated moving target data to support network training, uses the cross-entropy between the real and predicted labels to measure classification loss, and employs the multi-kernel local maximum mean difference of fully connected layers to measure domain confusion loss. The network loss, composed of the classification loss and the domain confusion loss regularization term, is backpropagated and optimized using an adaptive moment estimation algorithm. Ultimately, an optimal network model suitable for moving target detection based on measured data is obtained, solving the problem of insufficient labeled data in the radar field. Attached Figure Description
[0029] Figure 1 This is a flowchart of the present invention;
[0030] Figure 2 Graph of network training data structure;
[0031] Figure 3 This is a schematic diagram of a subdomain adaptive residual network;
[0032] Figure 4 The classification loss curve for network training;
[0033] Figure 5 A graph showing the classification accuracy of the network during training;
[0034] Figure 6 The experiment is based on the first set of measured data; (a) is a SAR image, (b) is the detection result of CA-CFAR, (c) is the detection result of the deep learning method, and (d) is the detection result of the subdomain adaptive residual network.
[0035] Figure 7 The experiment uses the second set of measured data; (a) is a SAR image, (b) is the detection result of CA-CFAR, (c) is the detection result of the deep learning method, and (d) is the detection result of the subdomain adaptive residual network. Detailed Implementation
[0036] The present invention will now be described in further detail with reference to the accompanying drawings.
[0037] In multi-channel SAR systems, classic moving target detection is performed pixel-by-pixel, meaning that each image pixel undergoes a series of processing steps before determining whether a moving target exists. Assuming that after preprocessing steps such as channel alignment, imaging, and image registration, the data (spatial snapshot) from the same image pixel received by different channels is represented as follows:
[0038]
[0039] Wherein, column vector s represents the moving target signal, and c and n are clutter and noise components, respectively. H0 is the null hypothesis, indicating that the target does not exist, and H1 is the alternative hypothesis, indicating that the target exists.
[0040] To address this binary assumption problem, most current multi-channel SAR moving target detection methods first perform spatial filtering on the spatial snapshot z based on certain optimization criteria, which can be expressed as:
[0041] z out =w H z
[0042] Where the vector w represents the filter coefficients, the superscript H denotes the conjugate transpose, and z out This is the output of the filter.
[0043] Then, z out The moving target is compared with a predetermined threshold value η to determine whether it exists, as follows:
[0044]
[0045] Based on this principle, this invention provides a multi-channel synthetic aperture radar moving target detection method based on subdomain adaptation, such as... Figure 1 As shown, the specific steps include:
[0046] Step 1: Set the synthetic aperture radar system and moving target parameters as shown in Table 1. Acquire n-channel radar simulated echo data. After channel alignment, imaging, and image registration processing, obtain clutter-free simulated moving targets. Acquire n-channel radar measured echo data. After channel alignment, imaging, and image registration processing, obtain measured moving targets, measured clutter, and measured data of unknown categories. Superimpose the clutter-free simulated moving target with some measured clutter at a signal-to-clutter ratio of 15-25 dB to obtain the simulated moving target.
[0047] Network training data such as Figure 2 As shown, simulated moving targets and some measured clutter are source domain data, labeled 1 and 0 respectively. Measured data with unknown categories are target domain data, and both are used together for network training. In addition, a small amount of measured moving targets and measured clutter are target domain data, labeled 1 and 0 respectively, used for network validation.
[0048] Table 1 Simulation Parameters
[0049]
[0050]
[0051] Step 2: Construct a subdomain adaptive residual network for ground moving target detection based on ResNet18 and subdomain adaptation, such as... Figure 3 As shown, this network retains the convolutional blocks of ResNet18 and adds three fully connected layers at the end, making it a binary classification network. Taking a three-channel synthetic aperture radar as an example, the network input carries the label Y. s Source domain data X s And unlabeled target domain data X t The input data has 6 channels and a size of 32×32. The network outputs predicted labels for the source domain data. Predicted labels for target domain data The output data has 2 channels and a size of 1×1.
[0052] Suppose that the source domain has n s A labeled data x s The target domain has n t x unlabeled data t Then the local maximum mean difference (LMMD) between the two domains can be expressed as:
[0053]
[0054] Where H is the regenerated Hilbert space corresponding to the feature kernel k, and φ(·) maps the variable to the regenerated Hilbert space. and These represent the categories belonging to, respectively, c. and The weights. Wherein, the weights... It can be represented as:
[0055]
[0056] Where y is the label corresponding to data x. This indicates the domain to which the data belongs; for source domain data x s and target domain data x t y ic These correspond to the true label and the predicted label, respectively. Since a single kernel function cannot achieve a proper mapping from the original space to the high-dimensional space, multiple kernel functions are typically used in LMMD. This paper selects a convex combination of five Gaussian kernels with different bandwidths; the multi-kernel function is expressed as:
[0057]
[0058] Where, β u Represents the u-th Gaussian kernel k u The weights. Therefore, the multi-kernel local maximum mean difference (MK-LMMD) of the fully connected layer Fcl (l=1,2,3) can be expressed as:
[0059]
[0060] Among them, z sl and z tl The outputs of source domain data and target domain data after passing through Fcl are respectively. This is a multi-kernel function. Therefore, the difference in local maximum mean among the multi-kernel values of the three fully connected layers is... The sum, i.e., domain confusion loss. DC , can be represented as:
[0061]
[0062] In addition, classification loss CFS The true label y of the source domain data s With predictive labels Cross-entropy metric:
[0063]
[0064] Where J(·) is the cross-entropy loss function.
[0065] Ultimately, network loss NW It can be represented as:
[0066]
[0067] Where λ is a regularization parameter representing the degree of domain confusion.
[0068] Step 3: Preprocess the data in the training and validation datasets. Treat the real and imaginary parts as independent channels, transforming the n-channel complex data into 2n-channel real data, and standardizing based on standard scores. Input the preprocessed training data into the network for training, calculate the network loss and backpropagate it, using an exponentially decaying learning rate and adaptive moment estimation optimization algorithm. Input the preprocessed validation data into the network for validation and calculate the network loss. The curves showing the changes in classification loss and classification accuracy with the number of iterations are shown below. Figure 4 , Figure 5 As shown, after the 7th iteration, the network's classification loss is as low as 0.0261, and the classification accuracy reaches 99.33%. At this point, the network is optimal, and the subdomain adaptive residual network parameters are saved.
[0069] Step 4: Import the saved network parameters into the subdomain adaptive residual network. Create a network test dataset by applying a sliding window (32×32) to the synthetic aperture radar (SAR) data to be tested. Preprocess the data in the test dataset and input it into the subdomain adaptive residual network. Combine the network's output predictions to obtain the SAR ground moving target detection results.
[0070] To verify the performance of the subdomain adaptive residual network, its detection results are compared with those of the traditional method CA-CFAR and deep learning methods. The first SAR image is shown below. Figure 6 As shown in (a), where T1-T8 are moving targets. The detection results of CA-CFAR are as follows: Figure 6 As shown in (b), solid boxes represent correct detections, dashed boxes represent false alarms, and the edge graph represents the magnitude within the bounded region before CA-CFAR detection. The detection results of the deep learning method and the subdomain adaptive residual network are shown below. Figure 6 (c) Figure 6 As shown in (d), solid and dashed boxes represent correct detections and false alarms, respectively. As shown in Table 2, even when both other methods have false alarms and missed alarms, the subdomain adaptive residual network can still achieve correct detections.
[0071] Table 2 shows the detection results of the first set of measured data.
[0072]
[0073] The second SAR image is as follows Figure 7 As shown in (a), where T1-T4 are moving targets, the detection results of CA-CFAR, deep learning, and subdomain adaptive residual network are respectively as follows: Figure 7 As shown in (b)-7(d).
[0074] Table 3 shows the detection results of the second set of measured data.
[0075]
[0076] As shown in Table 3, while all three methods can detect all real moving targets, the CA-CFAR method has 6 false alarms, the deep learning method has 1 false alarm, and the subdomain adaptive residual network has no false alarms. In summary, the subdomain adaptive multi-channel synthetic aperture radar moving target detection method of this invention significantly improves detection performance.
Claims
1. A moving target detection method for multi-channel synthetic aperture radar based on subdomain adaptation, characterized in that, Includes the following steps: (1) Obtain simulation and measured data of n-channel synthetic aperture radar, and make training dataset and verification dataset from the simulation and measured data; the training dataset includes simulated moving targets, measured clutter and measured data of unknown categories; the verification dataset includes measured clutter and measured moving targets, and adds real labels to clutter and moving targets; (2) Preprocess the data in the training and validation datasets, treating the real and imaginary parts as independent channels and standardizing them; (3) Based on ResNet18 and subdomain adaptation, a subdomain adaptive residual network for ground moving target detection is constructed. The preprocessed data is input into the subdomain adaptive residual network for training and validation, and the optimal subdomain adaptive residual network parameters are saved. (4) Import the saved network parameters into the subdomain adaptive residual network, make the synthetic aperture radar data to be tested into a test dataset, preprocess the data of the test dataset and input it into the subdomain adaptive residual network, and window the output prediction results to obtain the ground moving target detection results.
2. The multi-channel synthetic aperture radar moving target detection method based on subdomain adaptation according to claim 1, characterized in that, The implementation process of step (1) is as follows: (11) Set the radar system and moving target parameters, acquire n-channel radar simulation echo data, and obtain clutter-free simulated moving targets after channel alignment, imaging and image registration. (12) Obtain the measured echo data of the n-channel radar. After channel alignment, imaging and image registration processing, the measured moving target, measured clutter and measured data of unknown type are obtained. (13) The simulated moving target without clutter is superimposed with the measured clutter at a certain signal-to-clutter ratio to obtain the simulated moving target; (14) Add labels 0 and 1 to clutter and moving targets respectively; (15) The simulated moving target and some measured clutter constitute the labeled data of the source domain, and the measured data of unknown category is the unlabeled data of the target domain. Some measured clutter and measured moving target constitute the labeled data of the target domain.
3. The multi-channel synthetic aperture radar moving target detection method based on subdomain adaptation according to claim 1, characterized in that, The implementation process of step (2) is as follows: Treating the real and imaginary parts of the data in the training and validation datasets as independent channels, the n-channel complex data is transformed into 2n-channel real data and standardized based on standard scores.
4. The multi-channel synthetic aperture radar moving target detection method based on subdomain adaptation according to claim 1, characterized in that, The implementation process of step (3) is as follows: (31) Retain the convolutional blocks in ResNet18 and then connect them with three fully connected layers to introduce subdomain adaptation and construct a subdomain adaptive residual network for ground moving target detection. The network input is 2n channels of data and the network output is 2 channels of predicted labels. (32) During network training, the input includes labeled data and its true labels from the source domain and unlabeled data from the target domain; during network validation, the input includes labeled data and its true labels from the target domain, and the output is the predicted label. (33) Transform the real labels of the data into one-hot encoded form, use the cross-entropy between the real labels and the predicted labels to measure the classification loss, and use the sum of the differences between the multi-kernel local maximum means of the three fully connected layers to measure the domain confusion loss. (34) During training, the network loss is the classification loss of the source domain labeled data and the domain confusion loss regularization term; during validation, the network loss is the classification loss of the target domain labeled data.
5. The multi-channel synthetic aperture radar moving target detection method based on subdomain adaptation according to claim 4, characterized in that, Step (34) is achieved by the following formula: in, Loss due to domain obfuscation For classifying losses, It is a regularization parameter that represents the degree of domain confusion. Let cross-entropy be the loss function. , For the true labels and predicted labels of the source domain data, and These are the outputs after the source domain data and target domain data have passed through the fully connected layer, respectively. The number of labeled data. This represents the local maximum mean difference between the two domains.