Sea surface target re-detection method based on deep learning noise level estimation, storage medium and electronic equipment

Abstract

The invention provides a sea surface target re-detection method based on deep learning noise level estimation, a storage medium and electronic equipment, and the method comprises the following steps: firstly, calculating a background noise level and a detection threshold value through a VI-CFAR detector, and judging whether a target exists or not according to a binary hypothesis; on the basis of the existence of the target, further judging whether target missing detection exists or not by using a signal-to-noise ratio; secondly, if missing detection is carried out, generating more RD image samples by utilizing a generative adversarial network, and realizing expansion of a data set; inputting the expanded data set into a convolutional neural network, carrying out feature extraction, recognizing and replacing a salient target with environmental noise, and accurately estimating the noise level; and finally, the noise level estimation result is fed back to the VI-CFAR detector, and the detection threshold is recalculated, so that the dynamic adjustment of the detection threshold is realized, and the problem of missing detection in the environment that the weak target is submerged is solved.

Description

Technical Field

[0001] The present invention relates to the field of target detection in sea clutter environment, and in particular to a method for re-detecting sea surface targets based on deep learning noise level estimation, a storage medium and an electronic device. Background Art

[0002] In a complex sea clutter environment, radar target detection has long faced the dual challenges of strong interference and missed detection of weak targets. The VI-CFAR (constant false alarm rate) algorithm usually assumes that the background noise follows a certain specific statistical distribution (such as Gaussian distribution or exponential distribution), and designs a detection threshold based on this assumption. However, in practical applications, the sea clutter environment often exhibits strong non-stationarity and dynamic change characteristics, such as the fluctuations of sea waves, the changes of wind speed, and complex weather conditions such as storms. These factors make it difficult to uniformly describe the distribution characteristics of the background noise, resulting in a significant decrease in the accuracy of the threshold calculation of the traditional CFAR algorithm. This problem not only causes a rapid increase in the false alarm rate, but also makes the signals of weak targets submerged in the background noise, thus greatly increasing the missed detection rate and seriously affecting the accuracy and reliability of radar target detection. Therefore, there are two major problems in traditional CFAR target detection: one is that the traditional CFAR algorithm cannot effectively cope with the dynamic changes of the noise level. In a complex sea clutter environment, the background noise has significant non-stationary characteristics due to various environmental factors such as sea waves and wind speed. The statistical assumptions of the noise by traditional methods (such as Gaussian distribution or exponential distribution) often do not hold, resulting in inaccurate calculation of the detection threshold; the other is that the weak target signals are easily submerged in the strong clutter background, resulting in missed detection. Traditional algorithms are difficult to distinguish the subtle differences between weak targets and background noise. Especially in a background environment with dense or dynamic multi-targets, the missed detection problem is more serious, seriously affecting the accuracy and reliability of detection. And the accuracy of detection and reliability are important indicators of detection. How to improve it is an urgent technical problem to be solved. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for re-detecting sea surface targets based on deep learning noise level estimation, a storage medium and an electronic device, which can improve the detection accuracy and accuracy.

[0004] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0005] The method for re-detecting sea surface targets based on deep learning noise level estimation includes the following steps:

[0006] Step 1: Perform VI-CFAR detection on the echo signal received by the radar to determine whether a target exists. Specifically:

[0007] First, obtain the echo signal S(t) from the radar receiving device and use a low-pass filter to eliminate the high-frequency noise component to obtain a smoothed signal S filtered (t); secondly, estimate the background noise level N from the reference unit signals [R 1 ,R 2 ,...,R m ; finally, reduce the target detection to a binary hypothesis testing problem, and determine whether the target exists by comparing the signal strength S CUT of the unit under test and the detection threshold T.

[0008] Step 2: Determine whether there is a missed detection of a weak target according to the signal-to-noise ratio SNR for the echo signal. Specifically:

[0009] Calculate the signal-to-noise ratio SNR of the unit under test. If it is lower than the preset minimum threshold SNR min , it is determined that the signal is weak and there may be a missed detection.

[0010] Step 3: Expand the data samples in the missed detection area through a generative adversarial network (GAN) framework. Specifically:

[0011] Generate more range-Doppler (RD) map samples through a generative adversarial network (GAN) framework including a generator and a discriminator to expand the data set.

[0012] Step 4: Use a deep learning model to replace the target area in the missed detection area and then accurately estimate the noise level. Specifically:

[0013] The deep learning model extracts noise features from the range-Doppler (RD) map of the radar through a convolutional neural network (CNN) and replaces the target signal. The network consists of multiple convolutional layers (Conv), batch normalization layers, activation functions (PReLU), and fully connected layers. Based on the expanded data set, this model performs noise replacement processing on the target area and analyzes and estimates the noise level to obtain a new noise level N DL to adapt to the noise characteristics of the current environment;

[0014] Step 5: Reapply the VI-CFAR algorithm to the missed detection area for re-detection. Specifically:

[0015] According to the recalculated noise level value N DL , feedback it to the threshold T of the VI-CFAR detector for re-detection.

[0016] The method for performing VI-CFAR detection on the echo signal received by the radar in Step 1 to determine whether the target exists is as follows:

[0017] First, obtain the echo signal S(t) from the radar receiving device and use a low-pass filter to eliminate the high-frequency noise component to obtain the smoothed signal S filtered (t), and the formula is as follows:

[0018]

[0019] where h(t) is the impulse response of the filter. Then, divide the preprocessed signal according to a fixed window length w, and the window structure can be expressed as:

[0020] [R 1 ,R 2 ,...,R m ,G 1 ,...,G n ,CUT,G n+1 ,...G k ,R m+1 ,...,R 2m (2) where R i windows represent reference cells, CUT represents the cell to be measured, and G i represents the guard cell.

[0021] Secondly, estimate the background noise level N from the reference cell signals [R 1 ,R 2 ,...,R m , and the formula is:

[0022]

[0023] where R i is the signal strength of the i-th reference cell, and w i is the weight factor of the i-th reference cell, indicating the importance of this cell in noise estimation, satisfying w i ≥0, and m is the number of reference cells.

[0024] The weight factor w i in the formula is selected as follows:

[0025] (1) When the background noise is evenly distributed, equal weights can be used, that is, w i = 1;

[0026] (2) When the background noise is unevenly distributed, weights are assigned according to the distance between the reference cell and the cell to be measured. The closer the distance, the greater the weight. The formula is as follows:

[0027]

[0028] where d iis the distance between the i-th reference unit and the unit under test. Then, based on the background noise level N and the desired false alarm probability P FA , the detection threshold T is calculated, and its formula is as follows:

[0029] T = α·N (5)

[0030] where α is a threshold factor determined by the chi-square distribution function according to the desired false alarm probability P FA and the number of reference units m, and the derivation is as follows:

[0031]

[0032] The solution is as follows:

[0033]

[0034] Finally, the target detection is reduced to a binary hypothesis testing problem. By comparing the signal strength S of the unit under test CUT and the detection threshold T, it is determined whether the signal is a target:

[0035] (1) Null hypothesis (H 0 ): The unit under test is background noise, that is, H 0 : S CUT ≤T;

[0036] (2) Alternative hypothesis (H 1 ): The unit under test is a target signal, that is, H 1 : S CUT >T.

[0037] The method for judging whether a weak target is missed according to the signal-to-noise ratio SNR for the echo signal described in step 2 is as follows:

[0038] To measure the difference between the signal strength S of the unit under test CUT and the background noise level N, the concept of signal-to-noise ratio is introduced to characterize the contrast between the two, and its formula is:

[0039]

[0040] The judgment rule is as follows:

[0041] (1) If SNR < SNR min , it means that the signal is weaker than the background noise and may be judged as noise by the detection threshold, and there is a risk of missed detection;

[0042] (2) If SNR ≥ SNR min , it means that the signal strength is obvious enough and there is no risk of missed detection.

[0043] where SNR minIt is a preset value based on statistical analysis of historical data and actual application scenarios.

[0044] The data samples for the missed detection areas described in step 3 are augmented by a Generative Adversarial Network (GAN) framework using the following method:

[0045] First, generate a random noise vector and a class label:

[0046] z ∼ N(0, 1), c ∈ {0, 1,..., C - 1} (9)

[0047] Where z is the input noise vector of the generator, sampled from the standard normal distribution N(0, 1), used to provide randomness for the generator. c is the class label, representing the class of the target RD map, with a range from 0 to C - 1, where C is the total number of classes.

[0048] Second, the generator generates a conditional RD map based on the input noise z and class c:

[0049] G(z|c) = Generator(z, c) (10)

[0050] Where G is the generator, a deep neural network, learning to generate an RD map with real features from z and c. G(z|c) is the conditional RD map generated by the generator.

[0051] Next, to enhance the control effect of the class label on the generation process, the generator adopts a Conditional Channel Attention Module (CCAM), embedding the class label into the intermediate features of the generator:

[0052] CCAM(f, c) = S(MLP 3 (MLP 2 (f; MLP 1 (c)))) (11)

[0053]

[0054] Where f is the spatial average pooling result of the intermediate features of the generator or discriminator, used to extract the global information of the image, MLP i is a Multi - Layer Perceptron (MLP), used for non - linear mapping, converting the class label c into a more complex feature representation, S is the Sigmoid activation function, compressing the output value of CCAM to the range [0, 1], F is the intermediate feature map in the generator or discriminator, F′ is the feature map adjusted by CCAM, represents an element - wise multiplication operation, integrating the class - conditional information CCAM(f, c) into the feature map F.

[0055] Subsequently, the generated RD maps are normalized on a per-map basis, and the pixel values are normalized to the range [0, 1] to improve the consistency of the data distribution:

[0056]

[0057] Among them, x is the pixel value matrix of the RD map, max(x) is the maximum pixel value of the RD map, which is used to normalize the RD map to a unified range, and PIN(x) is the normalized RD map, whose pixel values are compressed to the range [0, 1].

[0058] Finally, the generated RD maps and the real RD maps are combined in a certain proportion to construct an enhanced dataset:

[0059] Dataset = {x real} ∪ {G(z|c)} (14)

[0060] Among them, {x real} is the real RD map dataset, and {G(z|c)} is the synthetic RD map dataset generated by the generator.

[0061] The method for accurately estimating the noise level by using a deep learning model to replace the target area in the missed detection area described in step 4 is as follows:

[0062] First, a truncation operation is performed on the RD image. By setting the upper threshold of the pixel value, the excessive signal intensity is restricted within a range, and the formula is:

[0063]

[0064] Among them, x i,j is the pixel value of the i, j-th pixel in the original RD image, and T is the preset truncation threshold. The preprocessed RD image is fed into the deep neural network, and the input RD image can be expressed as:

[0065] RDM = T + N (16)

[0066] Among them, T is the target signal image, and N is the pure noise image. The truncation operation transforms it into:

[0067]

[0068] Among them, is the truncated target signal part. The goal of the network is to learn the ability to restore N from RDM T through training, which requires the network to accurately identify and replace the feature patterns of the target signal.

[0069] Next, the RD map enters the first part of the network, that is, the multi-layer convolution module, and the formula for convolution is:

[0070]

[0071] Among them, w m,n is the weight of the convolution kernel, and b is the bias term. After each layer of convolution, batch normalization (Batch Normalization) is applied to accelerate training. The normalization formula is:

[0072]

[0073] Among them, μ and σ 2 are the mean and variance of the mini-batch data, and ε is a small positive number to prevent the denominator from being zero.

[0074] Secondly, the output of the convolutional layer is further processed through a residual block, which retains the original information of the input features, thereby gradually weakening the influence of the target signal. The design formula of the residual block is:

[0075] y residual = F(x, {W i}) + x (20)

[0076] Among them, F(x, {W i}) represents the output of the convolutional process, and x is the input feature.

[0077] Finally, enter the last part of the network, that is, the fully connected layer, which maps the input features to a pure noise image. The output noise map can be expressed as:

[0078] N i = f(RDM T , i; Θ) (21)

[0079] Among them, Θ are all the trainable parameters of the network, and f(·) is the non-linear mapping of the network.

[0080] After the above steps, the output of the network is a pure noise RD image, in which the target signal has been replaced, and the remaining noise information accurately reflects the distribution characteristics of the background noise.

[0081] The method of re-detection by re-applying the VI-CFAR algorithm to the missed detection area described in step 5 is as follows:

[0082] According to the recalculated noise level value N DL , adjust the threshold of the VI-CFAR detector. The calculation formula of the detection threshold T is as follows:

[0083] T = α·N DL (22)

[0084] By traversing the signal strength S in the RD map CUT, according to the binary hypothesis in step 1, determine which intensity values exceed the detection threshold T. In the case of possible missed detections, re - apply the VI - CFAR detector for re - detection such as in step 1 until all strong and weak targets are finally detected.

[0085] Advantages of the present invention:

[0086] Through the above - mentioned technical solution, the present invention aims at the problem of traditional CFAR target detection in sea clutter environment and proposes a method for re - detecting sea surface targets based on deep - learning noise level estimation.

[0087] First, use the VI - CFAR algorithm to initially detect the echo signal, quickly identify potential targets in the strong clutter background, and at the same time mark the areas where weak targets may be missed.

[0088] Secondly, introduce a noise level estimation module based on deep learning. Through a convolutional neural network (CNN), perform region filling and noise analysis on the pre - processed radar echo signal to dynamically estimate the noise level.

[0089] Finally, feedback the noise level output by the deep - learning model to the detector to dynamically optimize the detection threshold of the CFAR algorithm, and re - detect the radar echo signal, achieving accurate detection of multiple targets including strong and weak targets.

[0090] The present invention comprehensively considers the dynamic change characteristics of the noise level in a complex sea clutter environment and the difficulties of weak target detection. By introducing a deep - learning model to accurately estimate the noise level and feedback it to the detector, the adaptability of the traditional CFAR algorithm is effectively optimized. It improves the accuracy of target detection and the recognition ability of weak targets. Even in a strong clutter interference and non - stationary noise environment, it can achieve stable detection of multiple targets. At the same time, it reduces the missed detection rate and false alarm rate. Especially near strong targets or in high - interference regions, it significantly enhances the robustness and reliability of the system, providing an efficient and practical solution for radar target detection in complex environments. Description of the Drawings

[0091] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0092] Figure 1 It is a flowchart of the method of the present invention. Detailed Embodiments

[0093] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0094] As Figure 1 shown: In view of the problem of traditional CFAR target detection in the sea clutter environment, the present invention proposes a method for re-detecting sea surface targets based on deep learning noise level estimation. The method includes the following steps:

[0095] Step 1: Use the VI-CFAR algorithm to perform preliminary detection on the echo signal, quickly identify potential targets in the strong clutter background, and at the same time mark the areas where weak targets may be missed.

[0096] Step 1.1: Perform VI-CFAR detection on the echo signal received by the radar to determine whether a target exists. Specifically:

[0097] First, obtain the echo signal S(t) from the radar receiving device and use a low-pass filter to eliminate the high-frequency noise component to obtain the smoothed signal S filtered (t); secondly, estimate the background noise level N from the reference unit signals [R 1 , R 2 ,..., R m ; finally, reduce the target detection to a binary hypothesis testing problem, and determine whether a target exists by comparing the signal strength S CUT of the unit under test and the detection threshold T.

[0098] Among them, the method for performing VI-CFAR detection on the echo signal received by the radar in Step 1.1 to determine whether a target exists is as follows:

[0099] First, obtain the echo signal S(t) from the radar receiving device and use a low-pass filter to eliminate the high-frequency noise component to obtain the smoothed signal S filtered (t), and the formula is as follows:

[0100]

[0101] where h(t) is the impulse response of the filter. Then divide the preprocessed signal according to a fixed window length w, and the window structure can be expressed as:

[0102] [R 1 , R 2 ,..., R m , G 1 ,..., G n , CUT, G n+1 ,... Gk , R m+1 ,..., R 2m (2) Among them, R i represents the reference unit, CUT represents the unit under test, and G i represents the protection unit.

[0103] Secondly, estimate the background noise level N from the reference unit signals [R 1 , R 2 ,..., R m , and its formula is:

[0104]

[0105] Among them, R i is the signal strength of the i-th reference unit, and w i is the weight factor of the i-th reference unit, indicating the importance of this unit in noise estimation, and satisfies w i ≥ 0, and m is the number of reference units.

[0106] In the formula, the weight factor w i is selected as follows:

[0107] (1) When the background noise is evenly distributed, equal weights can be used, that is, w i = 1;

[0108] (2) When the background noise is unevenly distributed, weights are assigned according to the distance between the reference unit and the unit under test. The closer the distance, the greater the weight. The formula is as follows:

[0109]

[0110] Among them, d i is the distance between the i-th reference unit and the unit under test. Then, according to the background noise level N and the expected false alarm probability P FA , calculate the detection threshold T, and its formula is as follows:

[0111] T = α · N (5)

[0112] Among them, α is a threshold factor determined by the chi-square distribution function according to the expected false alarm probability P FA and the number of reference units m, and the derivation is as follows:

[0113]

[0114] Solving it gives:

[0115]

[0116] Finally, the target detection is reduced to a binary hypothesis testing problem by comparing the signal strength S of the unit under testCUT and the detection threshold T to determine whether the signal is a target:

[0117] (1) Null hypothesis (H 0 ): The unit under test is background noise, i.e., H 0 : S CUT ≤T;

[0118] (2) Alternative hypothesis (H 1 ): The unit under test is a target signal, i.e., H 1 : S CUT >T.

[0119] Step 1.2: Determine whether there is a missed detection of weak targets for the echo signal. Specifically:

[0120] Calculate the signal-to-noise ratio SNR of the unit under test. If it is lower than the preset minimum threshold SNR min , it is determined that the signal is weak and there may be a missed detection.

[0121] Among them, the method for determining whether there is a missed detection of weak targets for the echo signal described in Step 1.2 is as follows:

[0122] To measure the difference between the signal strength S CUT of the unit under test and the background noise level N, the concept of signal-to-noise ratio is introduced to characterize the contrast between the two. Its formula is:

[0123]

[0124] The judgment rule is as follows:

[0125] (1) If SNR < SNR min , it means that the signal is weak relative to the background noise and may be determined as noise by the detection threshold, and there is a risk of missed detection;

[0126] (2) If SNR ≥ SNR min , it means that the signal strength is obvious enough and there is no risk of missed detection.

[0127] Among them, SNR min is a preset value based on the statistical analysis of historical data and actual application scenarios.

[0128] Step 2: Introduce a noise level estimation module based on deep learning, and perform region filling and noise analysis on the preprocessed radar echo signal through a convolutional neural network (CNN) to dynamically estimate the noise level.

[0129] Step 2.1: Expand the data samples in the missed detection area through a generative adversarial network (GAN) framework. Specifically:

[0130] Generate more RD map samples through a generative adversarial network (GAN) framework that includes a generator and a discriminator to expand the dataset.

[0131] Among them, the method for expanding the data samples for the missed detection areas described in step 2.1 by using a generative adversarial network (GAN) framework is as follows:

[0132] First, generate a random noise vector and a class label:

[0133] z ∼ N(0, 1), c ∈ {0, 1,..., C - 1} (9)

[0134] Among them, z is the input noise vector of the generator, sampled from the standard normal distribution N(0, 1), used to provide randomness for the generator. c is the class label, representing the class of the target RD map, with a range from 0 to C - 1, where C is the total number of classes.

[0135] Second, the generator generates a conditional RD map according to the input noise z and class c:

[0136] G(z|c) = Generator(z, c) (10)

[0137] Among them, G is the generator, a deep neural network, learning to generate an RD map with real features from z and c. G(z|c) is the conditional RD map generated by the generator.

[0138] Next, in order to enhance the control effect of the class label on the generation process, the generator adopts a conditional channel attention module (CCAM) to embed the class label into the intermediate features of the generator:

[0139] CCAM(f, c) = S(MLP 3 (MLP 2 (f; MLP 1 (c)))) (11)

[0140]

[0141] Among them, f is the spatial average pooling result of the intermediate features of the generator or discriminator, used to extract the global information of the image, MLP i is a multi-layer perceptron (MLP), used for non-linear mapping, converting the class label c into a more complex feature representation, S is the Sigmoid activation function, compressing the output value of CCAM to the range [0, 1], F is the intermediate feature map in the generator or discriminator, F' is the feature map adjusted by CCAM, represents an element-wise multiplication operation, integrating the class conditional information CCAM(f, c) into the feature map F.

[0142] Subsequently, the generated RD images are normalized image by image, and the pixel values are normalized to the range of [0, 1] to improve the consistency of data distribution:

[0143]

[0144] where \(x\) is the pixel value matrix of the RD image, \(\max(x)\) is the maximum pixel value of the RD image, which is used to normalize the RD image to a unified range, and \(PIN(x)\) is the normalized RD image, whose pixel values are compressed to the range of [0, 1].

[0145] Finally, the generated RD images and the real RD images are combined in a certain proportion to construct an enhanced dataset:

[0146] Dataset = \{x real \} ∪ \{G(z|c)\} (14)

[0147] where \{x real \} is the real RD image dataset, and \{G(z|c)\} is the synthetic RD image dataset generated by the generator.

[0148] Step 2.2: Use a deep learning model to replace the target regions in the missed detection regions to accurately estimate the noise level. Specifically:

[0149] Based on the augmented dataset, the deep learning model performs noise replacement processing on the target regions and analyzes and estimates the noise level to obtain a new noise level \(N DL \), to adapt to the noise characteristics of the current environment;

[0150] Among them, the method of using a deep learning model to replace the target regions in the missed detection regions and accurately estimate the noise level described in Step 2.2 is as follows:

[0151] First, perform a truncation operation on the RD image. By setting the upper bound threshold of the pixel value, the excessive signal intensity is restricted within a range. The formula is:

[0152]

[0153] where \(x i,j \) is the \(i,j\)-th pixel value in the original RD image, and \(T\) is the preset truncation threshold. The preprocessed RD image is fed into the deep neural network, and the input RD image can be expressed as:

[0154] RDM = T + N (16)

[0155] where \(T\) is the target signal image and \(N\) is the pure noise image. The truncation operation converts it to:

[0156]

[0157] Among them, is the truncated part of the target signal. The goal of the network is to learn the ability to restore N from the RDM T through training, which requires the network to accurately identify and replace the characteristic pattern of the target signal.

[0158] Next, the RD diagram enters the first part of the network, namely the multi-layer convolutional module. The calculation formula for convolution is:

[0159]

[0160] where w m,n is the weight of the convolution kernel, and b is the bias term. Batch normalization is applied after each layer of convolution to accelerate training. The normalization formula is:

[0161]

[0162] where μ and σ 2 are the mean and variance of the mini-batch data, and ε is a small positive number to prevent the denominator from being zero.

[0163] Secondly, the output of the convolutional layer is further processed by the residual block, which retains the original information of the input features, thereby gradually weakening the influence of the target signal. The design formula of the residual block is:

[0164] y residual = F(x,{W i}) + x (20)

[0165] where F(x,{W i}) represents the output of the convolutional process, and x is the input feature.

[0166] Finally, it enters the last part of the network, namely the fully connected layer, which maps the input features to a pure noise image. The output noise map can be expressed as:

[0167] N i = f(RDM T , i; Θ) (21)

[0168] where Θ are all the trainable parameters of the network, and f(·) is the non-linear mapping of the network.

[0169] After the above steps, the output of the network is a pure noise RD image, in which the target signal has been replaced, and the remaining noise information accurately reflects the distribution characteristics of the background noise.

[0170] Step 3: Feed back the noise level output by the deep learning model to the detector to dynamically optimize the detection threshold of the CFAR algorithm, and re-detect the radar echo signal. Specifically:

[0171] According to the recalculated noise level value N DL , feed it back to the threshold T of the VI-CFAR detector for re-detection.

[0172] Among them, the method of re-applying the VI-CFAR algorithm for re-detection to the missed detection area described in step 3 is as follows:

[0173] According to the recalculated noise level value N DL , adjust the threshold of the VI-CFAR detector. The calculation formula of the detection threshold T is as follows:

[0174] T = α·N DL (22)

[0175] By traversing the signal intensity S in the RD map CUT , judge which intensity values exceed the detection threshold T according to the binary hypothesis in step 1. In case of possible missed detection, re-apply the VI-CFAR detector for re-detection such as in step 1 to finally detect all strong and weak targets.

[0176] In recent years, the rapid development of deep learning technology has provided a brand-new solution idea for radar signal processing, especially showing excellent performance in fields such as feature extraction and noise level estimation. The core advantage of deep learning lies in its ability to learn complex non-linear features through large-scale data training, which makes its adaptability in dynamic backgrounds far exceed that of traditional algorithms. As a powerful data generation model, the Generative Adversarial Network (GAN) has been widely used in sample augmentation and enhancement in sparse data scenarios in recent years. The samples generated by GAN can not only simulate the complex distributions of noise and targets in real scenarios, but also significantly improve the generalization ability of deep learning models in weak target detection tasks, providing strong technical support for radar target detection in low signal-to-noise ratio environments. In addition, due to its unique advantages in multi-layer feature extraction and efficient non-linear mapping, the Convolutional Neural Network (CNN) has become one of the mainstream methods for solving complex sea clutter environment problems. CNN extracts high-dimensional features of sea clutter through multi-layer convolution operations, can effectively distinguish target signals from background noise, and significantly improve the robustness of detection algorithms by dynamically adjusting detection thresholds. Compared with traditional methods, deep learning algorithms can adaptively adjust model parameters and dynamically optimize detection strategies according to the changes in noise distribution in the real-time environment, thus showing higher detection accuracy in scenarios with dense multi-target distributions and strong interference environments. Based on this background, an innovative technology combining deep learning and the VI-CFAR algorithm is proposed. This method significantly optimizes the accuracy of noise level estimation by introducing a Convolutional Neural Network (CNN) module and dynamically adjusts the detection threshold to adapt to the changes in complex environments. Compared with traditional CFAR algorithms, this method can significantly reduce the false alarm rate and miss detection rate in high-interference backgrounds, while improving the reliability of weak target signal detection. Generally speaking, this detection method that combines deep learning and traditional algorithms provides a brand-new technical path for solving target detection problems in complex sea clutter environments, not only having theoretical innovation, but also providing an effective reference scheme for the optimization of radar systems in practical applications.

[0177] The present invention comprehensively considers the dynamic change characteristics of the noise level in complex sea clutter environments and the difficulties of weak target detection. By introducing a deep learning model to accurately estimate the noise level and feedback it to the detector, the adaptability of traditional CFAR algorithms is effectively optimized. The accuracy of target detection and the recognition ability of weak targets are improved, and stable detection of multi-targets can be achieved even in strong clutter interference and non-stationary noise environments. At the same time, the miss detection rate and false alarm rate are reduced. Especially in the vicinity of strong targets or high-interference regions, the robustness and reliability of the system are significantly enhanced, providing an efficient and practical solution for radar target detection in complex environments.

[0178] The present invention is different from the traditional CFAR detector target detection method. The present invention takes into account the dynamic change characteristics of the noise level in a complex sea clutter environment and the difficulties of weak target detection, and addresses the problems of missed detection and false alarm rate in target detection. On this basis, a target detection method in a sea clutter environment based on deep learning noise level estimation is proposed, realizing accurate detection of multiple targets and reliable recognition of weak targets in a strong clutter background.

[0179] A computer-readable storage medium stores a computer program thereon. When the computer program is executed by a processor, it causes the device where the computer-readable storage medium is located to execute the above-mentioned sea surface target re-detection method based on deep learning noise level estimation. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory and other memories, etc.

[0180] An electronic device includes: a memory and a processor. A program that can run on the processor is stored on the memory. When the processor executes the program, it implements the above-mentioned sea surface target re-detection method based on deep learning noise level estimation.

[0181] If the modules / units integrated in the electronic device described in this application are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above-mentioned embodiment methods of this application, it can also be completed by a computer program instructing relevant hardware devices. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above-mentioned various method embodiments.

[0182] Furthermore, the computer-readable storage medium mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function, etc.; the data storage area can store data created according to the use of blockchain nodes, etc.

[0183] Computer-readable instructions are stored in the computer-readable storage medium. The computer-readable instructions are executed by a processor in the electronic device to implement the above-mentioned sea surface target re-detection method based on deep learning noise level estimation in any one of the embodiments.

[0184] In several embodiments provided by the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation.

[0185] The technical features of the above-described embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0186] The above-described embodiments only represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

[0187] It should be noted that the terms "including" and "having" in the specification and claims of the present application, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products, or devices.

[0188] Note that the above is only the preferred embodiment of the present invention and the application of technical principles. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described here. Various obvious changes, re-adjustments, and substitutions can be made by those skilled in the art without departing from the protection scope of the present invention. Therefore, although the present invention is described in more detail through the above embodiments, the present invention is not limited to the specific embodiments described here. Without departing from the concept of the present invention, more other effective embodiments can also be included, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A sea surface target re-detection method based on deep learning noise level estimation, characterized in that: The specific steps include the following: Step 1: Preprocess the radar echo signal: Use the VI-CFAR algorithm to perform initial detection on the echo signal, quickly identify potential targets in a strong clutter background, and mark weak target areas that may be missed; Step 2: Introduce a noise level estimation module based on deep learning, perform region filling and noise analysis on the preprocessed radar echo signal through the convolutional neural network (CNN), and dynamically estimate the noise level; Step 3: The noise level output by the deep learning model is fed back to the VI-CFAR detector to dynamically optimize the detection threshold of the CFAR algorithm, and the radar echo signal is re-detected to achieve accurate detection of multiple targets including strong and weak targets.

2. According to the sea surface target re-detection method based on deep learning noise level estimation, it is characterized in that: The step 1 specifically includes the following steps: Step 1.1: Perform VI-CFAR detection on the echo signal received by the radar to determine whether the target exists. Specifically: First, the echo signal S(t) is obtained from the radar receiving device and the high-frequency noise component is eliminated using a low-pass filter to obtain a smoothed signal S filtered (t); Secondly, from the reference unit signal [R1, R2, ..., R m ] to estimate the background noise level N; finally, target detection is reduced to a binary hypothesis testing problem, by comparing the signal strength S of the unit to be tested CUT and detection threshold T, to determine whether the target exists; Step 1.2: Determine whether a weak target is missed based on the signal-to-noise ratio (SNR) of the echo signal. The set of sample data where a weak target is missed is recorded as a weak target area. Specifically: Calculate the signal-to-noise ratio (SNR) of the unit under test. If it is lower than the preset minimum threshold SNR, min , the signal is judged to be weak, and there may be missed detection, which is recorded as the sample data of weak target missed detection. All the units to be tested are traversed, and the set of sample data of weak target missed detection is the weak target area.

3. According to the sea surface target re-detection method based on deep learning noise level estimation, it is characterized in that: The step 2 specifically includes the following steps: Step 2.1: For the data samples in the weak target area, the samples are expanded by generating adversarial network (GAN) framework. Specifically: Generate more RD graph samples through the Generative Adversarial Network (GAN) framework including generators and discriminators to expand the data set; Step 2.2: Use a deep learning model to replace the weak target area to accurately estimate the noise level. Specifically: The deep learning model extracts noise features from the radar's range-Doppler map through a convolutional neural network and replaces the target signal. The network consists of multiple convolutional layers Conv, batch normalization layers Batch Normalization, activation functions PReLU, and fully connected layers Fully Connected Layers; based on the expanded data set, this model performs noise replacement processing on the target area and analyzes and estimates the noise level to obtain a new noise level N DL , to adapt to the noise characteristics of the current environment; Step 5: Re-apply the VI-CFAR algorithm to re-detect the missed areas. Specifically: According to the recalculated noise level value N DL , fed back to the threshold T of the VI-CFAR detector for re-detection.

4. According to the sea surface target re-detection method based on deep learning noise level estimation, it is characterized in that: The step 1.1 specifically includes the following steps: First, the echo signal S(t) is obtained from the radar receiving device and the high-frequency noise component is eliminated using a low-pass filter to obtain a smoothed signal S filtered (t), the formula is as follows: Where h(t) is the impulse response of the filter; then the preprocessed signal is divided according to the fixed window length w, and the window structure can be expressed as: [R1,R2,...,R m ,G1,...,G n ,CUT,G n+1 ,...G k ,R m+1 ,...,R 2m ](2) Among them, R i Window represents the reference unit, CUT represents the unit under test, G i Indicates protection unit; Secondly, from the reference unit signal [R1, R2, ..., R m ] to estimate the background noise level N, the formula is: Among them, R i is the signal strength of the ith reference unit, w i is the weight factor of the i-th reference unit, indicating the importance of the unit in noise estimation, satisfying w i ≥0, m is the number of reference units; The weight factor w i The choices are as follows: (1) When the background noise is evenly distributed, equal weighting can be used, that is, w i =1; (2) When the background noise distribution is non-uniform, a weight is assigned according to the distance between the reference unit and the unit under test. The closer the distance, the greater the weight. The formula is as follows: where d i is the distance between the i-th reference unit and the unit to be tested; then according to the background noise level N and the expected false alarm probability P FA , calculate the detection threshold T, the formula is as follows: T=α·N (5) Among them, α is the expected false alarm probability P FA and the number of reference cells m, the threshold factor determined by the chi-square distribution function, is derived as follows: The solution is: Finally, by comparing the signal strength S of the unit under test CUT And the detection threshold T, determine whether the signal is the target: (1) Null hypothesis H0: The unit under test is background noise, that is, H0:S CUT ≤T; (2) Alternative hypothesis H1: The unit under test is the target signal, that is, H1:S CUT >T.

5. According to the sea surface target re-detection method based on deep learning noise level estimation, it is characterized in that: The step 1.2 specifically includes the following steps: In order to measure the signal strength S of the unit under test CUT The difference between the signal-to-noise ratio and the background noise level N introduces the concept of signal-to-noise ratio, which is used to characterize the contrast between the two. The formula is: The judgment rules are as follows: (1) If SNR < SNR min , indicating that the signal is weak relative to the background noise and may be judged as noise by the detection threshold, resulting in the risk of missed detection; (2) If SNR ≥ SNR min , indicating that the signal strength is obvious enough and there is no risk of missed detection; Among them, SNR min It is a preset value based on statistical analysis of historical data and actual application scenarios.

6. According to the sea surface target re-detection method based on deep learning noise level estimation, it is characterized in that: The step 2.1 specifically includes the following steps: First, generate random noise vectors and class labels: z~N(0,1),c∈{0,1,...,C-1} (9) Among them, z is the input noise vector of the generator, sampled from the standard normal distribution N(0,1), used to provide randomness for the generator; c is the category label, indicating the category of the target RD map, ranging from 0 to C-1, where C is the total number of categories; Second, the generator generates a conditioned RD graph based on the input noise z and category c: G(z|c)=Generator(z,c) (10) Where G is the generator, a deep neural network that learns to generate RD graphs with real features from z and c; G(z|c) is the conditional RD graph generated by the generator; Next, in order to enhance the control effect of category labels on the generation process, the generator adopts the conditional channel attention module (CCAM) to embed the category labels into the intermediate features of the generator: CCAM(f,c)=S(MLP3(MLP2(f;MLP1(c)))) (11) Among them, f is the spatial average pooling result of the intermediate features of the generator or discriminator, which is used to extract the global information of the image. i is a multilayer perceptron used for nonlinear mapping to transform the category label c into a more complex feature representation. S is a Sigmoid activation function that compresses the output value of CCAM to the range of [0,1]. F is an intermediate feature map in the generator or discriminator. F′ is the feature map adjusted by CCAM. represents the element-by-element multiplication operation, which integrates the category condition information CCAM(f,c) into the feature map F; Subsequently, the generated RD graphs are normalized one by one to normalize the pixel values ​​to the range [0,1] to improve the consistency of data distribution: Where x is the pixel value matrix of the RD map, max(x) is the maximum pixel value of the RD map, which is used to normalize the RD map to a uniform range, and PIN(x) is the normalized RD map, whose pixel values ​​are compressed to the range of [0,1]; Finally, the generated RD graph and the real RD graph are combined in a certain proportion to form an enhanced dataset: Dataset={x real }∪{G(z|c)} (14) Among them, {x real } is a real RD graph dataset, and {G(z|c)} is a synthetic RD graph dataset generated by the generator.

7. According to the sea surface target re-detection method based on deep learning noise level estimation, it is characterized in that: The step 2.2 specifically includes the following steps: First, the RD image is truncated: by setting the upper threshold of the pixel value, the excessive signal intensity is limited to a range. The formula is: Among them, x i,j is the i,jth pixel value in the original RD image, T is the preset truncation threshold; the preprocessed RD image is fed into the deep neural network, and the input RD image can be expressed as: RDM=T+N (16) Among them, T is the target signal image, N is the pure noise image; the truncation operation transforms it into: in, is the truncated target signal part; the goal of the network is to learn from RDM through training T The ability to restore N in the network requires the network to accurately identify and replace the characteristic patterns of the target signal; Next, the RD graph enters the first part of the network, the multi-layer convolution module. The convolution calculation formula is: Among them, w m,n is the weight of the convolution kernel, b is the bias term; BatchNormalization is applied after each layer of convolution to speed up training. The normalization formula is: Among them, μ and σ 2 is the mean and variance of the mini-batch data, ε is a small positive number to prevent the denominator from being zero; Secondly, the output of the convolutional layer is further processed by the residual block, which retains the original information of the input features and gradually weakens the influence of the target signal. The design formula of the residual block is: y residual =F(x,{W i })+x (20) Among them, F(x,{W i }) represents the output of the convolution process, and x is the input feature; Finally, we enter the last part of the network, the fully connected layer, which maps the input features to a pure noise image. The output noise map can be expressed as: N i =f(RDM T ,i;Θ) (21) Among them, Θ is all the trainable parameters of the network, f(·) is the nonlinear mapping of the network; After the above steps, the output of the network is a pure noise RD image, in which the target signal has been replaced and the remaining noise information accurately reflects the distribution characteristics of the background noise.

8. According to the sea surface target re-detection method based on deep learning noise level estimation, it is characterized in that: The step 3 specifically includes the following steps: According to the recalculated noise level value N DL , adjust the threshold of the VI-CFAR detector: The calculation formula of the detection threshold T is as follows: T=α·N DL (22) By traversing the signal strength S in the RD graph CUT , according to the binary hypothesis in step 1, determine which intensity values ​​exceed the detection threshold T; in the case of possible missed detection, reapply the VI-CFAR detector to re-detect as in step 1, and finally detect all strong and weak targets.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, the device where the computer-readable storage medium is located executes the sea surface target re-detection method based on deep learning noise level estimation as described in any one of claims 1 to 8.

10. An electronic device, characterized in that: include: A memory and a processor, wherein the memory stores a program that can be run on the processor, and when the processor executes the program, the sea surface target re-detection method based on deep learning noise level estimation as described in any one of claims 1-8 is implemented.

Images