False alarm controllable radar target detection method based on differentiable neumann pearson criterion

Abstract

The application discloses a false alarm controllable radar target detection method based on a differential Neumann Pearson criterion and belongs to the technical field of radar target detection. The application aims at the problem that the NP criterion cannot be directly used for neural network training due to non-differentiability in radar target detection based on a neural network. The application comprises the following steps: a traditional CFAR detector is used to process radar AR spectrum original data to extract candidate targets, and sample image blocks centered on the candidate targets are generated; all sample image blocks are used to construct a training data set, and labels are configured for all training data; a small classification network model with controllable false alarm is constructed, network parameters are iteratively optimized and trained by using the training data, and a trained classification network model is obtained; and real-time radar AR spectrum data are used to obtain image blocks to be detected, and the trained classification network model is used for target detection. The application realizes controllable false alarm radar target detection.

Description

Technical Field

[0001] The invention relates to a false alarm controllable radar target detection method based on a differentiable Newman-Pearson criterion, and belongs to the technical field of radar target detection. Background Art

[0002] Compared with the traditional constant false alarm rate (CFAR) detector, the neural network-based detection has the data-driven feature representation learning capability, which can improve the weak target detection performance in non-uniform clutter environments.

[0003] Under conditions of extreme sample imbalance, neural network detectors trained using a cross-entropy loss function struggle to control the false alarm probability by setting a variable threshold. In traditional radar theory, the Neyman-Pearson (NP) criterion is used to find the optimal detector under constraints on the false alarm probability. This criterion could potentially be applied to neural networks to achieve controllable false alarm detection. However, this criterion is non-differentiable and cannot be directly used for neural network training.

[0004] Patent CN114594463A ("A Small Target Feature Detection Method Based on a Combined Convex Hull") introduces a feature-based detection method that uses a combined convex hull to achieve controllable false alarm detection. By extracting three features from the radar observation vector in the sum and frequency domains, a combined convex hull algorithm is used to achieve a 3D non-convex decision region under a specific false alarm probability. Compared to the convex hull decision region of the original convex hull algorithm, this method has a smaller volume and higher detection performance. Due to the dimensionality limitation of the convex hull algorithm, this method is only applicable to features with a dimension of less than or equal to three. Furthermore, its performance relies primarily on manually extracted features, and its discriminability is poor compared to the features automatically extracted by data-driven neural networks.

[0005] Patent CN111580064A (“A method for detecting small targets on the sea surface based on multi-domain and multi-dimensional feature fusion”) introduces a method for compressing seven-dimensional features into a three-dimensional feature space and then using a fast convex hull algorithm to obtain controllable false alarm detection. The echo data under the four polarizations of HH, HV, VH, and VV are jointly processed to extract seven-dimensional feature information in the time domain, frequency domain, time-frequency domain, and polarization domain. Considering the need to ensure high-dimensional information volume while reducing the computational cost, the seven-dimensional features are compressed into a three-dimensional feature space, and a fast convex hull algorithm is used to obtain a convex decision area with controllable false alarms to achieve target detection. This method circumvents the dimensionality limitation of the convex hull algorithm, but there is a problem of information loss when linear dimensionality reduction is performed on high-dimensional features. At the same time, its performance mainly depends on manually extracted features, and it also has poor discriminability compared to the features automatically extracted by data-driven neural networks.

[0006] Patent CN112147601A ("A Method for Detecting Small Targets on the Sea Surface Based on Random Forests") introduces a method for controlled false alarm detection based on random forests. This method extracts multidimensional feature vectors from the time, frequency, and time-frequency domains and feeds them into the random forest input. A mathematical relationship between the splitting factor and the false alarm rate is established to obtain a two-class random forest classifier for controlled false alarm detection, thereby completing target detection. This method separates the neural network feature extraction step from the subsequent controlled false alarm detection step, resulting in poor discriminability of the automatically extracted features.

[0007] Patent CN113064133A (“A method for detecting small targets on the sea surface based on a deep network in the time-frequency domain”) introduces a method for obtaining controllable false alarm detection using a guided cubic spline curve in a two-dimensional feature space. The method uses the Inception-ResNet V2 deep network as a feature extractor to extract the two-dimensional feature vector of the normalized time-frequency graph, and then uses a guided cubic sample curve in the two-dimensional feature space to obtain a non-convex decision region where the false alarm is controllable, thereby achieving anomaly detection or target detection. This method combines neural networks and feature detection technology, and has the advantages of autonomous learning feature extraction and multi-dimensional feature combination. However, the direct use of large deep networks has a large computational overhead. At the same time, the feature extraction step of the neural network and the subsequent controllable false alarm detection step are separated, resulting in poor discriminability of the automatically extracted features.

[0008] Patent CN110363151A ("Radar Target Detection Method with Controllable False Alarms Based on a Dual-Channel Convolutional Neural Network") introduces a method for achieving controllable false alarm detection using a variable threshold and a support vector machine (SVM). The method first preprocesses the radar echo signal to obtain signal time-frequency and amplitude information to construct a training dataset. A dual-channel convolutional neural network model is then constructed to extract deep features. Finally, a controllable false alarm classifier is used to detect targets. This method separates the neural network feature extraction step from the subsequent controllable false alarm detection step based on the SVM, resulting in poor discriminability of the automatically extracted features. Furthermore, the variable threshold-based controllable false alarm detection method is not adaptable to extreme sample imbalance situations.

[0009] In summary, the problems existing in existing radar target detection include:

[0010] (1) Traditional CFAR detectors in radar detection are based on the statistical model of radar echo signals. Their performance is optimal only under ideal conditions and they are difficult to adapt to the changing radar detection background and targets with complex characteristics.

[0011] (2) Neural network is a data-driven method that can simultaneously realize automatic feature learning and end-to-end training of classifiers. The extreme sample imbalance phenomenon often existing in radar measured data greatly affects the performance of the neural network model. It is difficult to achieve high-performance classification using the commonly used cross entropy loss function and minimum mean square error loss function.

[0012] (3) Existing image classification models and image segmentation models often use classification accuracy, recall rate, Dice coefficient, F1 score, etc. as performance evaluation indicators, without fully considering the impact of false alarm rate, and it is difficult to meet the actual application needs of actual radar. Summary of the Invention

[0013] Aiming at the problem that the NP criterion cannot be directly used for neural network training in radar target detection based on neural network due to its non-differentiability, the present invention provides a false alarm controllable radar target detection method based on the differentiable Newman-Pearson criterion.

[0014] The present invention provides a false alarm controllable radar target detection method based on a differentiable Newman-Pearson criterion, comprising:

[0015] Step 1: Use a traditional CFAR detector to process the radar AR spectrum raw data to extract candidate targets and generate sample image patches centered on the candidate targets. Use all sample image patches to construct a training dataset and assign labels to all training data.

[0016] Step 2: Build a small classification network model with controllable false alarms, and use the training data to iteratively optimize the network parameters to obtain a trained classification network model;

[0017] The small classification network model is constructed based on a small convolutional neural network, and a differentiable Newman-Pearson loss function L(θ) with respect to network parameters is set based on the differentiable Newman-Pearson criterion:

[0018]

[0019] Where L S (θ) is the supervised loss function term, L R (θ) is the regular loss function term, λ is the regular term coefficient, is the differentiable target detection probability, is the differentiable false alarm probability, α is the expected false alarm probability level;

[0020] Based on the supervised loss function term L S (θ) Use training data to pre-train the network parameters; then based on the supervised loss function term L S (θ) and the regularized loss function term L R(θ) Fine-tune the network parameters using the training data to obtain the trained classification network model, where the decision threshold of the trained classification network model is fixed at 0.5;

[0021] Step 3: The real-time radar AR spectrum data is passed through the traditional CFAR detector to extract a local image block centered on the candidate target as the image block to be detected, and then the trained classification network model is used to perform target detection on the image block to be detected.

[0022] According to the false alarm controllable radar target detection method based on the differentiable Newman-Pearson criterion of the present invention, the method of configuring labels for all training data in step 1 is:

[0023] The real target echo area of ​​the radar AR spectrum raw data is annotated with pixels, and then the real target position is obtained by point condensation operation. The label of the training data is determined according to the distance between the spatial position of the training data and the real target position. If the distance between the spatial position of the training data and the real target position is less than or equal to the distance threshold, the training data is assigned the target label, otherwise the training data is assigned the clutter label.

[0024] According to the false alarm controllable radar target detection method based on the differentiable Newman-Pearson criterion of the present invention, the differentiable target detection probability is set and the differentiable false alarm probability as follows:

[0025]

[0026] Where N s is the number of training data in mini-batch training; is the neural network for the i-th training data z i The predicted value of the class label; Represents the i-th training data z i The corresponding true label; ε is a constant.

[0027] According to the false alarm controllable radar target detection method based on the differentiable Newman-Pearson criterion of the present invention, the method for determining the regularization term coefficient λ includes:

[0028] The value of the regularization coefficient λ is set so that the network parameter fine-tuning training phase is initialized to supervise the loss function term L S (θ) and λ·L R The (θ) values ​​are the same or of the same order of magnitude.

[0029] According to the false alarm controllable radar target detection method based on the differentiable Newman-Pearson criterion of the present invention, the expected false alarm probability level α is set to:

[0030]

[0031] In the formula is the true expected false alarm probability of each frame of training data, N t is the average number of candidate targets extracted by the traditional CFAR detector for each frame of training data; 2 / (1×10 -4 ) indicates that the false alarm probability of each frame of training data is 1×10 -4 There are 2 clutter false alarms.

[0032] According to the false alarm controllable radar target detection method based on the differentiable Newman-Pearson criterion of the present invention, each frame of training data corresponds to 200 to 1000 image block training data.

[0033] According to the false alarm controllable radar target detection method based on the differentiable Newman-Pearson criterion of the present invention, the value of ε is 1.

[0034] The beneficial effects of the present invention are as follows: the method of the present invention provides a method for realizing controllable detection of false alarms based on a neural network, adopts a neural network to realize automatic feature extraction under data drive, designs a differentiable Newman-Pearson criterion loss function suitable for extreme sample imbalance scenarios to approximately realize the NP criterion, and realizes the training of the classification network.

[0035] Compared to methods combining neural networks and feature detection, the method of the present invention does not require the design of an additional feature-based false alarm controllable classifier to achieve false alarm controllable detection. Compared to false alarm controllable detection methods based on variable thresholds, it is particularly suitable for actual radar detection scenarios under extreme sample imbalance. It has an end-to-end neural network training process, a simple implementation method, high false alarm control accuracy, and high detection performance at low false alarm levels. It is applicable to various neural network architectures.

[0036] The method of the present invention breaks through the limitations of traditional CFAR detector methods and can realize weak target detection in complex and non-uniform radar clutter environments. It not only has good false alarm control capabilities under extreme sample imbalance of radar data, but also has good detection performance at low false alarm levels. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flowchart of the false alarm controllable radar target detection method based on the differentiable Newman-Pearson criterion of the present invention;

[0038] Figure 2 This is the structural block diagram of a small classification network model. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0040] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0041] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0042] Specific implementation method 1. Combination Figure 1 and Figure 2 As shown, the present invention provides a false alarm controllable radar target detection method based on the differentiable Newman-Pearson criterion, comprising:

[0043] Step 1: Use a traditional CFAR detector to process the radar AR spectrum raw data to extract candidate targets, generating sample image patches centered on the candidate targets. All sample image patches are used to construct a training dataset, and all training data is labeled. The label of each image patch is determined based on its distance from the true target location, thereby simultaneously constructing target training samples and clutter training samples. The samples and corresponding labels are fed into the network multiple times for processing to fully train the network model parameters.

[0044] Step 2: Build a small classification network model (CNN) with controllable false alarms, and use the training data to iteratively optimize the network parameters to obtain the trained classification network model;

[0045] The small classification network model is constructed based on a small convolutional neural network, and a differentiable Newman-Pearson loss function L(θ) with respect to network parameters is set based on the differentiable Newman-Pearson criterion:

[0046]

[0047] Where L S (θ) is the supervised loss function term, which is used to constrain the network’s detection performance to be optimal; L R (θ) is the regularization loss function term, which is used to constrain the false alarm probability of the network to meet the expected value; λ is the regularization term coefficient, is the differentiable target detection probability, is the differentiable false alarm probability, α is the expected false alarm probability level;

[0048] Based on the supervised loss function term L S(θ) Use training data to pre-train the network parameters; then based on the supervised loss function term L S (θ) and the regularized loss function term L R (θ) uses the training data to fine-tune the network parameters; obtain the trained classification network model; after the neural network is trained by the loss function L(θ), the false alarm control ability of the network is fixed at the desired level, and the decision threshold of the network output is set to 0.5 by default and does not need to be adjusted.

[0049] Step 3: The real-time radar AR spectrum data is passed through the traditional CFAR detector to extract a local image block centered on the candidate target as the image block to be detected. The trained classification network model is then used to perform target detection on the image block to determine the target and the clutter background, thereby achieving controllable false alarm detection of the target.

[0050] In step 2 of this embodiment, the detection network model is iteratively optimized and trained. The supervised loss term in the differentiable Newman-Pearson loss function is first used for pre-training, and then the supervised loss term and the regularized loss term are fully used to fine-tune the network, which can reduce the difficulty of selecting the coefficient of the regularized loss term.

[0051] The radar AR spectrum data in this embodiment is angle-range spectrum data.

[0052] Furthermore, the method for configuring labels for all training data in step 1 is:

[0053] The real target echo area of ​​the radar AR spectrum raw data is annotated using a pixel labeling tool to obtain the echo area occupied by the real target. The real target position is then obtained using a point condensation operation. The label of the training data is determined based on the distance between the spatial position of the training data in the AR spectrum raw data and the real target position. If the distance between the spatial position of the training data and the real target position is less than or equal to the distance threshold, the training data is assigned a target label; otherwise, the training data is assigned a clutter label.

[0054] Going further, set the differentiable target detection probability and the differentiable false alarm probability as follows:

[0055]

[0056] Where N s is the number of training data in mini-batch training; is the neural network for the i-th training data z i The predicted value of the class label; Represents the i-th training data z i The corresponding true label; ε is a small constant used to ensure the stability of network training; it can be set to 1 by default.

[0057] Training data z i It is a small-sized image block. Considering the pixel size occupied by the target echo and the fact that the image block contains enough background echoes, the image block size can be set to 31×31.

[0058] Based on the loss function L(θ), the classification network can be trained. In order to facilitate the setting of the regularization term coefficient λ, this embodiment adopts a two-stage network training method. First, the supervised loss function term L is used. S (θ) is used for pre-training of the network until the network basically converges, and then the regular loss function term L is used R (θ) is used to constrain the network to meet the desired false alarm probability.

[0059] The method for determining the regularization term coefficient λ includes:

[0060] The value of the regularization coefficient λ is set so that the network parameter fine-tuning training phase is initialized to supervise the loss function term L S (θ) and λ·L R The values ​​of (θ) are the same or of the same order of magnitude, thereby avoiding the problem of blind selection of the regularization term coefficient λ.

[0061] In addition, since we need to directly control the false alarm probability corresponding to each frame of radar AR spectrum training data, rather than the number of current batch training samples N corresponding to the loss function, s Therefore, for the setting of parameter α, it is necessary to set the true expected false alarm probability corresponding to each frame of radar AR spectrum training data. Transformed into the expected false alarm probability α of the current batch of training samples, that is, the expected false alarm probability level α is set to:

[0062]

[0063] In the formula is the true expected false alarm probability of each frame of training data, N t is the average number of candidate targets extracted by the traditional CFAR detector for each frame of training data; 2 / (1×10 -4 ) indicates that the false alarm probability of each frame of training data is 1×10 -4 There are two clutter false alarms when the target is detected, and the ratio varies with the different forms of radar echo data. After the network training is completed, the default detection threshold of 0.5 is used for the output of the classification network to directly achieve controllable false alarm detection of the target.

[0064] Each frame of training data corresponds to 200 to 1000 image block training data, which includes more than 40 target image blocks and the rest are clutter image blocks. The specific number of image block training data can be determined based on computing resources, network training overhead, radar AR spectrum size, etc. Each frame of training data is a radar angle-range spectrum (AR spectrum) data with a size of 1024×480, and the data in each pixel (or unit) is quantized into the interval [0,255].

[0065] In this embodiment, the specific structure of the constructed false alarm controllable small classification network model is as follows: Figure 2 As shown in the figure, the network architecture consists of three convolutional layers, three max pooling layers, one global pooling layer, one fully connected layer, and a final sigmoid activation function. All three convolutional layers use 3×3 kernels. The first convolutional layer extracts features from a 1-channel input image patch to produce an 8-channel feature map. The second convolutional layer further processes this 8-channel feature map to produce a 16-channel feature map. The third convolutional layer performs a similar process, extracting deeper features from the data. All three max pooling layers use a stride of 2×2 for maximum sampling to preserve the local maximum eigenvalues. This reduces the spatial size of the processed feature maps by a factor of four, reducing the overall computational overhead and ensuring rotational invariance. The global pooling layer compresses deep features in tensor form into 32×1 vector form. The fully connected layer maps the 32×1 high-dimensional vector features into a one-dimensional space. The sigmoid activation function converts the feature values ​​in the one-dimensional space into probabilistic outputs.

[0066] In summary, the design of the differentiable detection probability in the method of the present invention can adapt to extreme sample imbalance scenarios; the design of the parameter α can directly control the false alarm probability at the point level of each frame of radar data; the training process of pre-training plus fine-tuning training facilitates network training and improves the false alarm control accuracy.

[0067] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely illustrative of the principles and applications of the invention. It should be understood that many modifications may be made to the illustrative embodiments, and that other arrangements may be devised, without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in ways other than those described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be used in conjunction with other described embodiments.

Claims

1. A false alarm controllable radar target detection method based on the differentiable Newman-Pearson criterion, characterized in that include, Step 1: Use a traditional CFAR detector to process the radar AR spectrum raw data to extract candidate targets and generate sample image patches centered on the candidate targets. Use all sample image patches to construct a training dataset and assign labels to all training data. Step 2: Build a small classification network model with controllable false alarms, and use the training data to iteratively optimize the network parameters to obtain a trained classification network model; The small classification network model is constructed based on a small convolutional neural network, and a differentiable Newman-Pearson loss function L(θ) with respect to network parameters is set based on the differentiable Newman-Pearson criterion: Where L S (θ) is the supervised loss function term, L R (θ) is the regularized loss function term, λ is the regularized term coefficient, is the differentiable target detection probability, is the differentiable false alarm probability, α is the expected false alarm probability level; Based on the supervised loss function term L S (θ) Use training data to pre-train the network parameters; then based on the supervised loss function term L S (θ) and the regularized loss function term L R (θ) Use training data to fine-tune the network parameters; The trained classification network model is obtained, and the decision threshold of the trained classification network model is 0.5; Step 3: The real-time radar AR spectrum data is passed through the traditional CFAR detector to extract a local image block centered on the candidate target as the image block to be detected, and then the trained classification network model is used to perform target detection on the image block to be detected.

2. The false alarm controllable radar target detection method based on the differentiable Newman-Pearson criterion according to claim 1 is characterized in that: The method for configuring labels for all training data in step 1 is: The real target echo area of ​​the radar AR spectrum raw data is annotated with pixels, and then the real target position is obtained by point condensation operation. The label of the training data is determined according to the distance between the spatial position of the training data and the real target position. If the distance between the spatial position of the training data and the real target position is less than or equal to the distance threshold, the training data is assigned the target label, otherwise the training data is assigned the clutter label.

3. The false alarm controllable radar target detection method based on the differentiable Newman-Pearson criterion according to claim 2 is characterized in that: Setting differentiable target detection probability and the differentiable false alarm probability as follows: Where N s is the number of training data in mini-batch training; is the neural network for the i-th training data z i The predicted value of the class label; Represents the i-th training data z i The corresponding true label; ε is a constant.

4. The false alarm controllable radar target detection method based on the differentiable Newman-Pearson criterion according to claim 3 is characterized in that: The method for determining the regularization term coefficient λ includes: The value of the regularization coefficient λ is set so that the network parameter fine-tuning training phase is initialized to supervise the loss function term L S (θ) and λ·L R The (θ) values ​​are the same or of the same order of magnitude.

5. The false alarm controllable radar target detection method based on the differentiable Newman-Pearson criterion according to claim 4 is characterized in that: The desired false alarm probability level α is set as: In the formula is the true expected false alarm probability of each frame of training data, N t is the average number of candidate targets extracted by the traditional CFAR detector for each frame of training data; 2 / (1×10 -4 ) indicates that the false alarm probability of each frame of training data is 1×10 -4 There are 2 clutter false alarms.

6. The false alarm controllable radar target detection method based on the differentiable Newman-Pearson criterion according to claim 5, characterized in that: Each frame of training data corresponds to 200 to 1000 image block training data.

7. The false alarm controllable radar target detection method based on the differentiable Newman-Pearson criterion according to claim 3, characterized in that: The value of ε is 1.

Images