A sea surface weak target detection method based on residual network and hypersphere constraint

Abstract

The application discloses a sea surface weak target detection method based on a residual network and a hypersphere constraint. The method can solve the problems of training underfitting or overfitting caused by the fact that the target sample is much smaller than the sea clutter sample under the sea clutter background, and the specific steps include: 1, performing short-time Fourier transform on each piece of data obtained after dividing the radar echo signal sample to obtain a time-frequency spectrum; 2, constructing an anomaly detection network model combined with a support vector data description based on a variational autoencoder of a residual network; 3, inputting the time-frequency spectrum into the anomaly detection network model for end-to-end training; 4, performing online detection of the target based on the trained model; and 5, evaluating the performance of the application from the constructed multi-dimensional index. The application can effectively enhance the separability between the target and the sea clutter, and improve the weak target detection precision under the condition of a real small sample.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing and deep learning technology, specifically relating to a method for detecting weak sea surface targets based on a combination of Residual Network (ResNet), Support Vector Data Description (SVDD), and Variational Auto-Encoder (VAE). This invention is applicable to anomaly detection of small sea surface targets against complex sea clutter backgrounds, particularly in tasks where target samples are extremely scarce or missing. This method performs deeper processing of radar echo signals, thereby improving the detection performance of weak sea surface targets against complex sea clutter backgrounds. Background Technology

[0002] Detecting weak targets on the sea surface has always been a key challenge in maritime defense early warning and maritime rescue. Sea clutter has nonlinear, nonstationary, and chaotic characteristics, and is often subject to discrete interference such as wave peaks and broken waves. Under high sea conditions such as strong winds and high waves, the signal-to-clutter ratio is extremely low. Traditional detection methods based on statistical models or texture analysis and chaotic prediction are prone to problems such as increased false alarm rate or decreased detection rate under complex sea conditions.

[0003] With the development of artificial intelligence, target detection methods based on deep learning have gradually become a research hotspot. Early studies mostly adopted supervised classification models based on convolutional neural networks, which require the construction of a balanced dataset containing a large number of "clutter" and "targets". However, in actual radar application scenarios, there is a serious "sample imbalance" problem: pure sea clutter data is easy to obtain and is huge in quantity, while data on weak targets on the sea surface (such as floating boats, periscopes, etc.) is extremely difficult to obtain, and samples are very scarce.

[0004] Existing supervised deep learning methods tend to overfit to the majority class (sea clutter) when the sample is severely imbalanced, while exhibiting poor generalization ability to the minority class (target). Furthermore, traditional autoencoders or simple feature extraction methods struggle to capture the deep temporal and frequency texture features of sea clutter. Therefore, how to utilize massive amounts of pure sea clutter data for classification learning, constructing a detection model that can effectively describe the distribution boundary of sea clutter and sensitively perceive targets deviating from that boundary, is a key technical problem urgently needing to be solved in this field.

[0005] The paper “Detection of low-velocity and floating small targets in seaclutter via income-reference particle filters” (Signal Processing, 2018, 148: 78-90) proposes an intelligent detection method by combining time-frequency features with convolutional neural networks. However, this method usually requires a large number of balanced samples for learning. The imbalance problem in sea surface echo data, where target samples are far fewer than clutter samples, can easily lead to poor generalization ability of the trained model, resulting in overfitting to sea clutter samples or underfitting to target samples. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing supervised learning methods that rely on a large number of target samples and are affected by sample imbalance, and to propose a method for detecting weak targets on the sea surface based on residual networks and hypersphere constraints.

[0007] The specific idea behind this invention is to treat the detection of weak targets on the sea surface as an "anomaly detection" problem. Utilizing the powerful generative reconstruction capabilities of a VAE built with ResNet and the spherical constraints of SVDD, the model is trained end-to-end using only pure sea clutter samples. The model learns a compact "clutter hypersphere" in the latent space, and during testing, samples deviating from the center of this hypersphere are identified as targets.

[0008] The specific steps to achieve the objective of this invention are as follows:

[0009] (1) Acquire radar echo signal data, select radar echo data under HH polarization mode as the analysis object, use sliding window method to cut one-dimensional time domain radar echo signal into several data segments, and perform short-time fourier transform (STFT) operation on each data segment to convert it into two-dimensional time spectrum, and construct a dataset containing pure sea clutter training samples and samples to be detected;

[0010] (2) An anomaly detection network model, ResNet-SVDD-VAE, is constructed, which combines a variational autoencoder (VAE) based on a residual network (ResNet) and a support vector data description (SVDD). The overall structure of this model consists of three main parts: a probabilistic encoder, a latent space sampling layer and an SVDD hypersphere constraint module, and a probabilistic decoder. In the ResNet-SVDD-VAE model, the VAE adopts a residual network architecture to extract high-dimensional nonlinear time-frequency features from the input two-dimensional radar time-frequency spectrum and compress and map them into probability distribution parameters of a low-dimensional latent space. The latent space sampling layer generates latent feature vectors using reparameterization techniques and constructs a feature hypersphere constraining the pure sea clutter distribution in the latent space based on the SVDD algorithm. The decoder adopts a transposed convolution structure symmetrical to the encoder to amplify and restore the latent feature vectors layer by layer, and finally reconstructs a time-frequency spectrum with the same size as the input data.

[0011] (3) The model is trained end-to-end using only the pure sea clutter samples constructed in step (1). A hypersphere center is initialized in the latent space. By optimizing the joint loss function, the projection of the pure sea clutter samples in the latent space is made to cluster as compactly as possible around the hypersphere center, while ensuring that the reconstruction error of the decoder is minimized.

[0012] (4) Based on the trained model, online target detection is performed. The distance between the feature of the sample to be tested in the latent space and the center of the hypersphere is calculated as the anomaly score. A binary classification decision is made based on the threshold determined by the validation set.

[0013] (5) Construct a multi-dimensional evaluation system to comprehensively evaluate the detection performance of the model under different false alarm rate conditions through indicators such as detection rate, false alarm rate, accuracy rate and F1 metric;

[0014] Compared with the prior art, the present invention has the following advantages:

[0015] This invention employs the concept of unsupervised learning for single-class classification. During the training phase, no target samples are required; only sea clutter samples are needed for training, perfectly solving the problem of scarce target data in real-world scenarios. Furthermore, the introduction of a ResNet residual structure addresses the training challenges of deep neural networks, resolving gradient vanishing and exploding issues in deep networks and preventing performance degradation. Through joint optimization using VAE reconstruction constraints and SVDD hypersphere compactness constraints, the model achieves a more robust fit to sea clutter distributions and effectively distinguishes between wave clutter spikes and weak targets. Attached Figure Description

[0016] Figure 1 This is a flowchart of the present invention;

[0017] Figure 2 The following is an example of the SVDD-VAE effect built using the Convolutional Neural Network (CNN) framework, based on data #280 of the IPIX radar signal dataset.

[0018] Figure 3 The following is an example of the SVDD-VAE built using the ResNet framework, based on data #280 of the IPIX radar signal dataset.

[0019] Figure 4 Taking data #26 of the IPIX radar signal dataset as an example, here is the effect of building a ResNet-SVDD-VAE with a spatial dimension parameter of 64.

[0020] Figure 5 Taking data #26 of the IPIX radar signal dataset as an example, here is the effect of building a ResNet-SVDD-VAE with a spatial dimension parameter of 16. Detailed Implementation

[0021] The present invention will now be further described with reference to the accompanying drawings.

[0022] See attached document Figure 1 The specific steps of this invention are as follows.

[0023] Step 1: Acquire radar echo signal data. This invention takes horizontal-horizontal (HH) polarized radar echo signals as the research object.

[0024] The specific implementation steps are as follows:

[0025] Perform signal slicing according to the following formula:

[0026]

[0027] Where x(n) is the original one-dimensional radar echo sequence with a total length of N. A sliding window technique is used to slice the original signal. The window length is L, and the sliding step size is S; by sliding truncation, the continuous echo data is divided into several local data segments. The total number of truncated windows is M.

[0028] According to the following formula, time-frequency transformation is performed on each truncated local data segment using STFT operation to extract the time-frequency characteristics of the signal:

[0029]

[0030] Where STFT(m,ω) represents the complex result in the time-frequency domain after the STFT operation, L is the set window length, n represents the discrete-time sampling index, m represents the time frame index, ω represents the angular frequency index, and W(n) is the window function used to truncate signal segments; x m (n) represents the m-th segment of data signal; the symbol ∑ represents the summation operation, that is, to accumulate all n that meet the conditions; the symbol e represents the base of the natural logarithm (Euler's number); the symbol j represents the imaginary unit.

[0031] Step 2: Construct the anomaly detection network model ResNet-SVDD-VAE.

[0032] The residual blocks contained in the encoder residual network architecture are calculated according to the following formula; the mathematical expression of each residual block is as follows:

[0033] y=σ(BN(Conv2(σ(BN(Conv1(x))))))+Shortcut(x)

[0034] Where x represents the input feature and y represents the output feature; Conv represents the convolution operation; BN represents batch normalization (BN), used to normalize the data distribution; σ represents the Leaky ReLU (Rectified LinearUnit) activation function, used to introduce non-linear characteristics; Shortcut(x) represents a skip connection, which directly performs the identity mapping when the dimensions of the input and output features are the same; when the dimensions are different, the input features are dimension-matched through linear projection before being added to the output of the convolutional layer to solve the degradation problem in deep networks;

[0035] After the encoder output is flattened, a one-dimensional depth feature vector h is first obtained. This h is then mapped through two parallel fully connected layers to obtain the mean vector μ and the log-variance vector log(σ). 2 The mean vector is calculated using the following formula after a linear transformation:

[0036] μ = W μ h+b μ

[0037] Among them, W μ It is a weight matrix specifically used to calculate the mean μ of the latent variables. μ It is specifically used to calculate the mean μ bias vector of latent variables.

[0038] Backpropagation is performed according to the following formula to calculate the latent variables z of the latent space sampling layer generated using the reparameterization technique:

[0039]

[0040] Where ∈ is a random noise vector sampled from the standard normal distribution; the symbol exp represents the exponential operation; the symbol This represents the square root operation;

[0041] The decoder employs a structure symmetrical to the encoder, using transposed convolutional layers to progressively enlarge the size of the low-dimensional latent variable z, ultimately reconstructing an image with the same size as the input image.

[0042] Step 3, Model initialization and end-to-end training.

[0043] A hypersphere center is initialized in the latent space. By optimizing the joint loss function, the projections of pure sea clutter samples in the latent space are made to cluster as compactly as possible around the hypersphere center, while minimizing the reconstruction error of the decoder.

[0044] Total loss function L total The reconstruction loss (denoted as L) recon KL divergence (KLD, denoted as L) KLD ) and SVDD hypersphere loss (denoted as L) SVDD The formula is as follows: (The formula is based on a weighted average of three parts.)

[0045] L total =L recon +L KLD +α·L SVDD

[0046] Where α is a hyperparameter used to balance the weight of SVDD loss in the total loss;

[0047] The reconstruction loss (L) is calculated according to the following formula. recon ):

[0048] Mean Squared Error (MSE) is used to measure the difference between the input time-frequency graph x and the reconstructed time-frequency graph. Differences between them:

[0049]

[0050] Where N represents the total number of time-frequency units in the sample; the symbol ||·|| 2 ∑ represents the square of the L2 norm, and ∑ represents the summation sign, which is to sum the squares of the differences between the corresponding time-frequency cell values ​​in the input time-frequency plot and the reconstructed time-frequency plot;

[0051] Calculate the KL divergence loss (L) according to the following formula. KLD ):

[0052]

[0053] Where K is the dimension of the latent vector; j is the index of the dimension; ∑ denotes the summation symbol; μ j This represents the component of the latent feature mean vector output by the encoder in the j-th dimension; This represents the component of the latent feature variance vector output by the encoder in the j-th dimension.

[0054] Calculate the SVDD hypersphere loss (L) according to the following formula. SVDD ):

[0055] To constrain the compactness of the pure sea clutter sample distribution in the potential space, the distance to the center c of the hypersphere is calculated:

[0056]

[0057] The hypersphere center c is determined through pre-training initialization. Specifically, before end-to-end joint training of the model, all pure sea clutter training set samples are first input into the encoder for forward computation to obtain multiple mean vectors corresponding to these samples in the latent space. Then, the arithmetic mean of these mean vectors is calculated and used as the initial position of the hypersphere center c. Here, k is the number of samples input into the latent space, B is the number of samples in the current training batch, and μ... k The potential mean of the k-th sample is pre-initialized to the hypersphere center vector c, which remains fixed or is dynamically updated during training;

[0058] Step 4, Target Online Detection and Judgment

[0059] The latent mean vector μ of the test sample is calculated according to the following formula. test The squared Euclidean distance from the center c of the hypersphere is used as the outlier score:

[0060] Score = ||μ test -c|| 2

[0061] Where μ test Denotes the potential mean vector, ||·|| 2 This represents the square operation of the Euclidean norm; the decision threshold is set to T, which is the decision threshold obtained after the false alarm rate is controlled; if Score≥T, it is determined to be a target; if Score<T, it is determined to be sea clutter.

[0062] Step 5: Use a multi-dimensional evaluation system to comprehensively assess performance.

[0063] The performance of this anomaly detection model is evaluated using the detection rate, false alarm rate, accuracy, and F1 score. The calculation formulas are as follows:

[0064]

[0065] Among them, the closer the detection rate is to 1, the better, indicating a lower false alarm rate; the false alarm rate represents the degree of misclassification of non-target class samples, and the lower the better; the accuracy rate represents the overall classification precision, and the larger the value, the stronger the model's comprehensive discrimination ability; the closer the F1 score is to 1, the better the model performance.

[0066] The invention will now be further described with reference to simulation diagrams.

[0067] 1. Simulation conditions:

[0068] The simulation system used in this invention features an Intel Core high-frequency CPU (3.92GHz) and 64GB of DDR5 memory. The graphics processing unit (GPU) is an NVIDIA GeForce RTX 5080 with 16GB of dedicated video memory, supporting TF32 and floating-point acceleration. The operating system is Windows 11 64-bit; the programming language is Python 3.10; the deep learning framework is PyTorch 2.4.0; and the development environment is PyCharm 2024.

[0069] The balance coefficient (regularization coefficient) α of the joint loss function is selected: this parameter is used to balance the reconstruction accuracy of the VAE with the compactness of the SVDD hypersphere. In this embodiment, α is specifically set to 1.5 to ensure that the latent space features can be compactly clustered around the center of the hypersphere; the learning rate η of the AdamW optimization algorithm is selected: to prevent gradient oscillations in the early stage of model training, the initial learning rate η = 0.0001 is set; the feature dimension d of the latent space can be flexibly adjusted, and after verification, d = 32 is sufficient for most datasets; at the same time, this parameter determines the compression ratio of the sea clutter time-frequency features. The maximum number of iterations E = 200 is set, and the tolerance period of the early stopping mechanism is set to 50 iterations.

[0070] 2. Simulation content and result analysis:

[0071] Simulation 1:

[0072] This paper uses publicly available data from the IPIX radar database website of McMaster University in Canada, primarily selecting data files collected in 1993 for simulation experiments.

[0073] Figure 2 and Figure 3Taking group #280 of the IPIX dataset as an example, with a sampling length of 1024 and a latent space dimension of d=32, the latent space feature distribution of CNN-SVDD-VAE and ResNet-SVDD-VAE built with CNN was compared under the hypersphere constraint.

[0074] The simulation results show that the ResNet-SVDD-VAE detection model can effectively enhance the distinguishability between the target and sea clutter.

[0075] Simulation 2:

[0076] We selected non-overlapping sampling with a sampling length of 1024 and continued using the ResNet-SVDD-VAE detection model. Figure 4 and Figure 5 Taking IPIX dataset #26 as an example, the difference from the former is that... Figure 4 Set the potential spatial dimension d = 64. Figure 5 By setting the latent space dimension d=16, the control variables are set to the same dataset, and the other parameters are set to different latent space dimensions.

[0077] The simulation results show that ResNet-SVDD-VAE can maintain high detection accuracy in potential spaces of different dimensions, proving that the model has good hyperparameter stability, can effectively extract the essential features of sea clutter in low-dimensional space, and is not overly sensitive to changes in network structure parameters.

[0078] Simulation 3:

[0079] To evaluate the performance of the proposed method, we selected group #311 of the IPIX radar dataset for validation experiments and compared it with several benchmark algorithms. These algorithms included En-OCSVM (Enhanced Single-Class Support Vector Machine), Residual Network (ResNet), Convolutional Neural Network (CNN), and Variational Autoencoder (VAE). Performance evaluation included detection rate, false alarm rate, and accuracy. As shown in Table 1, the proposed method achieves higher detection rate and accuracy with a lower false alarm rate, significantly outperforming other comparative methods.

[0080] Table 1 Comparison of detection performance for dataset #311

[0081] Detection rate False alarm rate accuracy En-OCSVM 0.6664 0.0528 0.8788 ResNet 0.8108 0.0012 0.9781 CNN 0.6409 0.0090 0.9617 VAE 0.9066 0.0012 0.9462 This article's method 0.9582 0.0012 0.9951

Claims

1. A method for detecting weak targets on the sea surface based on residual networks and hypersphere constraints, comprising the following steps: (1) Acquire radar echo signal data, select radar echo data under HH polarization mode as the analysis object, use sliding window method to cut one-dimensional time domain radar echo signal into several data segments, and perform short-time fourier transform (STFT) on each data segment to convert it into two-dimensional time spectrum, and construct a dataset containing pure sea clutter training samples and samples to be detected; (2) An anomaly detection network model, ResNet-SVDD-VAE, is constructed, which combines a variational autoencoder (VAE) based on a residual network (ResNet) and a support vector data description (SVDD). The overall structure of this model consists of three main parts: a probabilistic encoder, a latent space sampling layer and an SVDD hypersphere constraint module, and a probabilistic decoder. In the ResNet-SVDD-VAE model, the VAE adopts a residual network architecture to extract high-dimensional nonlinear time-frequency features from the input two-dimensional radar time-frequency spectrum and compress and map them into probability distribution parameters of a low-dimensional latent space. The latent space sampling layer generates latent feature vectors using reparameterization techniques and constructs a feature hypersphere constraining the pure sea clutter distribution in the latent space based on the SVDD algorithm. The decoder employs a transposed convolutional structure symmetrical to the encoder to amplify and restore the latent feature vectors layer by layer, ultimately reconstructing a time-spectrum map with the same size as the input data. (3) The model is trained end-to-end using only the pure sea clutter samples constructed in step (1). A hypersphere center is initialized in the latent space. By optimizing the joint loss function, the projection of the pure sea clutter samples in the latent space is made to cluster as compactly as possible around the hypersphere center, while ensuring that the reconstruction error of the decoder is minimized. (4) Based on the trained model, online target detection is performed. The distance between the feature of the sample to be tested in the latent space and the center of the hypersphere is calculated as the anomaly score. A binary classification decision is made based on the threshold determined by the validation set. (5) Construct a multi-dimensional evaluation system to comprehensively evaluate the detection performance of the model under different false alarm rate conditions through indicators such as detection rate, false alarm rate, accuracy rate and F1 metric.

2. The method according to claim 1, wherein the acquisition of the radar echo signal dataset in step (1) is characterized in that: The HH polarization refers to the use of horizontally polarized electromagnetic waves in both the transmission and reception processes of the radar system. That is, the transmitted wave is horizontally polarized, and the received wave is also horizontally polarized, thus forming an HH polarized radar signal. The radar echo signal data contains multiple range cells, each corresponding to echo information within a certain spatial range. Some range cells contain target echo signals, while the rest are pure sea clutter echo signals. After dividing the sampling points in each range cell, time-frequency transformation is performed to construct a sample set, forming a data set containing target samples and pure sea clutter samples. The data set also follows the principle of training using only sea clutter samples. The data preprocessing and time spectrum construction described in step (1) specifically include the following sub-steps: (1a) Let the original one-dimensional radar echo sequence be x(n), and its total length be N. The original signal is sliced ​​using a sliding window technique. The window length is set to L, and the sliding step size is S; by sliding truncation, the continuous echo data is divided into several local data segments. The total number of truncated windows M can be calculated using the following formula: (1b) Perform STFT on each truncated local data segment to extract the time-frequency characteristics of the signal. The transform formula is defined as follows: Where STFT(m, ω) represents the complex result in the time-frequency domain, L is the set window length, n represents the discrete-time sampling index, m represents the time frame index, ω represents the angular frequency index, and W(n) is the window function used to truncate signal segments; x m (n) represents the m-th segment of data signal; the symbol ∑ represents the summation operation, that is, to accumulate all n that meet the conditions; the symbol e represents the base of the natural logarithm (Euler's number); the symbol j represents the imaginary unit.

3. The method according to claim 1, characterized in that, The ResNet-SVDD-VAE detection model described in step (2) has the following network structure details: (2a) The encoder adopts a residual network architecture, which contains multiple residual blocks; the mathematical expression of each residual block is as follows: y=σ(BN(Conv2(σ(BN(Conv1(x))))))+Shortcut(x) Where x represents the input feature and y represents the output feature; Conv represents the convolution operation; BN represents batch normalization (BN), used to normalize the data distribution; σ represents the Leaky ReLU (Rectified Linear Unit) activation function, used to introduce non-linear characteristics; Shortcut(x) represents a skip connection, which directly performs the identity mapping when the dimensions of the input and output features are the same; when the dimensions are different, the input features are dimension-matched through linear projection and then added to the output of the convolutional layer to solve the degradation problem in deep networks; (2b) After the encoder output is flattened, a one-dimensional depth feature vector h is first obtained. This vector is then mapped through two parallel fully connected layers to obtain the mean vector μ and the log-variance vector log(σ). 2 The mean vector is calculated through a linear transformation, as shown in the following formula: μ=W μ h+b μ Wherein, the log-variance vector log(σ) 2 In ) σ represents the standard deviation; W μ It is a weight matrix specifically used to calculate the mean μ of the latent variables. μ This is specifically used to calculate the mean μ bias vector of the latent variables. To achieve backpropagation, a reparameterization technique is used to generate the latent variables z of the latent space sampling layer: Where ∈ is a random noise vector sampled from the standard normal distribution; the symbol exp represents the exponential operation; the symbol This represents the square root operation; (2c) The decoder adopts a structure symmetrical to the encoder, and uses transposed convolutional layers to gradually enlarge the size of the low-dimensional latent variable z, ultimately reconstructing an image with the same size as the input image.

4. The method according to claim 1, characterized in that, The joint loss function and its calculation method described in step (3) are as follows: (3a) Definition of joint loss function: Total loss function L total The reconstruction loss (denoted as L) recon KL divergence (KLD, denoted as L) KLD ) and SVDD hypersphere loss (denoted as L) SVDD The formula is as follows: (The formula is based on a weighted average of three parts.) L total L recon +L KLD +α·L SVDD Where α is a hyperparameter used to balance the weight of SVDD loss in the total loss; (3b) Reconstruction Loss (L) recon )calculate: Mean Squared Error (MSE) is used to measure the difference between the input time-frequency graph x and the reconstructed time-frequency graph. Differences between them: Where N represents the total number of time-frequency units in the sample; the symbol ||·|| 2 ∑ represents the square of the L2 norm, and ∑ represents the summation sign, which is to sum the squares of the differences between the corresponding time-frequency cell values ​​in the input time-frequency plot and the reconstructed time-frequency plot; (3c) KL divergence loss (L KLD )calculate: Where K is the dimension of the latent vector; j is the index of the dimension; ∑ denotes the summation symbol; μ j This represents the component of the latent feature mean vector output by the encoder in the j-th dimension; This represents the component of the latent feature variance vector output by the encoder in the j-th dimension. (3d)SVDD Hypersphere Loss (L) SVDD )calculate: To constrain the compactness of the pure sea clutter sample distribution in the potential space, the squared Euclidean distance to the center c of the hypersphere is calculated: Among them, ||·|| 2 The square operation represents the Euclidean norm; k is the number of samples input into the latent space, and B is the number of samples in the current training batch; μ k The potential mean of the k-th sample is given by a pre-initialized hypersphere center vector c, which remains fixed or is dynamically updated during training.

5. The method according to claim 1, characterized in that, The calculation formulas for the target detection decision and performance evaluation index mentioned in steps (4) and (5) are as follows: (4a) Calculate the latent mean vector μ of the test sample test The squared Euclidean distance from the center c of the hypersphere is used as the outlier score: Score=||μ test -c|| 2 Where μ test Denotes the potential mean vector, ||·|| 2 This represents the square operation of the Euclidean norm; the decision threshold is set to T, which is the decision threshold obtained after the false alarm rate is controlled; if Score≥T, it is determined to be a target; if Score<T, it is determined to be sea clutter. The calculation formulas for the detection rate, false alarm rate, accuracy, and F1 score mentioned in step (5) are as follows: Among them, a detection rate close to 1 indicates that the model has a strong ability to capture targets, corresponding to an extremely low risk of missed detection; the false alarm rate quantifies the probability that pure sea clutter is incorrectly identified as a target, and the smaller the value, the lower the false alarm level; the accuracy reflects the global precision of the classifier, and a high value represents stronger comprehensive recognition performance; the F1 measure, as a comprehensive evaluation index, has a value close to 1, which reflects the optimization of the overall detection efficiency of the model.

Images