A method for detecting small targets on the sea surface based on improved 1D-Lenet network

Abstract

The present invention proposes a method for detecting small targets on the sea surface based on an improved 1D-Lenet network. First, multidimensional features are extracted from the time domain, frequency domain, time-frequency domain, and polarization domain, and the energy, geometry, and other differences of sea clutter and target echoes in different domains are fully exploited. Second, an improved 1D-Lenet network is designed to achieve controllable classification design of false alarms in high-dimensional feature space. Deep networks usually use a large number of convolution kernels and have a deeper structure. While improving performance, they also lead to problems such as excessive computational complexity and long training time during network training. However, the improved 1D-Lenet network has fewer parameters and enables the network to have the potential to achieve rapid detection.

Description

Technical Field

[0001] The invention belongs to the technical field of radar signal processing, and in particular relates to a method for detecting small sea targets based on an improved 1D-Lenet network. Background Art

[0002] At present, small targets on the sea surface are the main targets for high-resolution surveillance radar detection, such as boats, periscopes, lifeboats, etc. Due to the small scattering cross-sectional area of ​​such targets and the complexity and variability of sea clutter, the signal-to-clutter ratio (SCR) of small targets is small and they are often in a critical detectable state. In radar signal processing, long-term observation is an effective means to improve the detection performance of small targets on the sea surface. Under long-term observation, the differences between sea clutter and echoes containing targets are found and condensed into features, and multiple features are combined for detection, that is, feature-based detection methods. In the time domain, some scholars have proposed a detector based on the Hurst index, which describes the difference in amplitude between sea clutter and echoes containing targets. In the Doppler domain, some scholars have proposed a detector based on three features, which extracts one time domain feature and two frequency domain features to describe energy and geometric differences, namely relative amplitude ratio, relative Doppler peak height, and relative vector entropy. In the time-frequency domain, some scholars have proposed a detector based on three time-frequency features, which extracts energy and geometric features in the time-frequency diagram. In order to further improve the detection performance, more feature joint detection is an inevitable trend. Then, how to extract effective features from multiple domains is the first difficulty in the research. In addition, the combination of multiple features transforms the traditional detection problem into a binary classification problem in a high-dimensional feature space, which brings up the problem of designing two classifiers with controllable false alarms, which is the second difficulty in the research. Summary of the invention

[0003] In order to solve the above problems, the present invention provides a method for detecting small floating targets on the sea surface, which can be used for shore-based or ship-borne high-resolution radars to comprehensively improve the sea detection performance. The present invention provides a method for detecting small targets on the sea surface based on an improved 1D-Lenet network, comprising the following steps:

[0004] (1) Obtain the reference unit observation matrix Z of P reference units around the unit to be detected p ,p=1,2,...,P, where the unit to be detected receives N continuous pulses under four polarizations respectively, forming a CUT observation matrix Z=[z HH ,z HV ,z VH ,z VV ],z HH ,z HV ,z VH ,z VV They represent the observation vectors under HH, HV, VH, and VV polarizations, respectively, with a dimension of N;

[0005] (2) Convert the CUT measurement matrix and the reference unit measurement matrix into the frequency domain, time-frequency domain, and polarization domain;

[0006] (3) Extract the features of the unit to be detected from the time domain, frequency domain, time-frequency domain, and polarization domain respectively, and record the total number of features as D;

[0007] (4) Take each feature as a dimension and construct a D-dimensional feature space. Then, convert the D features in step (3) into a feature vector ξ = [ξ 1 ,ξ 2 ,…,ξ D ] T ;

[0008] (5) Take the feature vector of the unit to be detected in step (4) as input and enter the 1D-Lenet network with the optimal model parameters to obtain the input belonging to H 1 The probability value of the hypothesis β, H 1 Assume that the unit to be detected contains a target echo, wherein the optimal model parameters of the 1D-Lenet network are extracted according to the pure clutter and the simulated target echo according to steps (2) to (4), as a training set, based on the 1D-Lenet network, the model parameters of the 1D-Lenet network are optimized based on the training set to obtain the optimal model parameters;

[0009] (6) Use the probability value β in step (5) as the final test statistic

[0010] (7) H of the training set 0 Assume that the sample under is used as the input of the 1D-Lenet network with the optimal model parameters, and obtain M inputs belonging to H 1 The probability of the hypothesis β 1 ,β 2 ,β 3 ,....,β M , H 0 Assume that the unit to be detected contains only sea clutter; according to the set false alarm rate P fa , and obtain the corresponding decision threshold γ; according to the test statistic The relationship between the detection threshold γ and the detection threshold γ is used to determine whether there is a target in the unit to be detected:

[0011] if It indicates that the unit to be detected contains target echo. If This indicates that the unit to be detected does not contain the target echo.

[0012] Furthermore, H 0 Assumptions and H1 The assumed formula is:

[0013]

[0014] Among them, c HH ,c HV ,c VH ,c VV denote the sea clutter vectors under HH, HV, VH, and VV polarizations, respectively, and s HH ,s HV ,s VH ,s VV Represent the target echo vectors under four polarizations, They respectively represent the clutter vectors of the p-th reference unit under four polarizations. The reference unit is pure clutter data and is used to provide sea clutter information for the unit to be detected.

[0015] Furthermore, in step (2), for each column of data in the measurement matrix of the reference unit to be detected, the frequency domain amplitude spectrum AS is calculated respectively, and the formula is:

[0016]

[0017] Where z represents the N-dimensional echo vector, f represents the frequency variable, fr represents the radar pulse repetition frequency, |·| represents the amplitude, n represents the time dimension, and z(n) represents the amplitude of the echo at the nth time point;

[0018] For each column of data in the measurement matrix of the reference unit to be detected, the two-dimensional time-frequency diagram SPWVD of the echo vector z is calculated respectively, and the formula is as follows:

[0019]

[0020] Among them, n represents the time dimension, l represents the frequency dimension, Δf d is the sampling interval of the normalized Doppler frequency, Q and q represent the number of samples in the time dimension and the corresponding variable, J and j represent the number of samples in the frequency dimension and the corresponding variable, g(q) and h(p) are the smoothing windows of the time dimension and frequency dimension, respectively.

[0021] Furthermore, in step (3), in the time domain feature extraction, the Hurst exponent calculation formula is as follows:

[0022]

[0023] Where k and b represent the slope and intercept of the fitted line, m represents the sequence interval, n represents the time dimension, and z(n) represents the amplitude of the echo at the nth time point.

[0024] The relative average amplitude RAA is calculated as:

[0025]

[0026] Furthermore, in step (3), in the frequency domain feature extraction, the calculation formula of the relative Doppler peak height RDPH is:

[0027]

[0028] Among them, DPH represents the Doppler peak height position, Peak represents the maximum value of the Doppler peak, represents the position of the Doppler peak in the Doppler spectrum, the symbol #Δ represents the total number falling into the number set Δ, and Δ represents the Doppler interval around the Doppler peak;

[0029] Relative vector entropy RVEξ 4 The calculation formula is:

[0030]

[0031] ξ 4 represents the relative vector entropy RVE, represents the normalized frequency domain amplitude spectrum, represents the normalized frequency domain amplitude spectrum of the reference unit, P is the total number of reference units, and p represents the pth reference unit.

[0032] Furthermore, in frequency domain feature extraction, the ridge energy RI calculation formula is as follows:

[0033]

[0034] At each time scale of the time-frequency graph, take the first 5 bright pixels and mark them as 1 and the remaining pixels as 0 to generate a binary graph. On the binary graph, get the connected region Ω 1 ,Ω 2 ,...,Ω w , the number of connected regions CR and the maximum connected region size MS are obtained by the following formulas:

[0035] CR(z)=W

[0036]

[0037] Where W represents the number of connected regions, #Ω w It means taking the non-zero element of the w-th connected region, and W represents the number of connected regions.

[0038] Furthermore, in the polarization domain feature extraction, the Hurst exponent and RAA are extracted for each polarization vector of the CUT observation matrix Z, namely:

[0039] ξ 1 = Hurst(zHH ),ξ 2 = Hurst(z HV ),ξ 3 = Hurst(z VH ),ξ 4 = Hurst(z VV )

[0040] ξ 5 =RAA(z HH ),ξ 6 =RAA(z HV ),ξ 7 =RAA(z VH ),ξ 8 =RAA(z VV )

[0041] For each polarization vector of the CUT observation matrix Z, extract RDPH and RVE, that is:

[0042] ξ 9 =RDPH(z HH ),ξ 10 =RDPH(z HV ),ξ 11 =RDPH(z VH ),ξ 12 =RDPH(z VV )

[0043] ξ 13 =RVE(z HH ),ξ 14 =RVE(z HV ),ξ 15 =RVE(z VH ),ξ 16 =RVE(z VV )

[0044] For each polarization vector of the CUT observation matrix Z, RI, CR and MS are extracted, that is:

[0045] ξ 17 =RI(z HH ),ξ 18 =RI(z HV ),ξ 19 =RI(z VH ),ξ 20 =RI(z VV )

[0046] ξ 21 =CR(z HH ),ξ 22 =CR(z HV ),ξ23 =CR(z VH ),ξ 24 =CR(z VV )

[0047] ξ 25 =MS(z HH ),ξ 26 =MS(z HV ),ξ 27 =MS(z VH ),ξ 28 =MS(z VV ).

[0048] Furthermore, in step (5), the 1D-Lenet network includes a C1 convolution layer, an S2 downsampling layer, a C3 convolution layer, an S4 downsampling layer, a C5 convolution layer, an F6 fully connected layer and a final output Output layer which are connected in sequence. The C1 convolution layer, the C3 convolution layer and the C5 convolution layer are convolution kernels with a size of 3 and a step size of 1. The final output Output layer has a PRelu nonlinear activation function. A Dropout operation unit is provided between the C1 convolution layer and the S2 downsampling layer, between the C3 convolution layer and the S4 downsampling layer, and between the C5 convolution layer and the F6 fully connected layer. The 1D-Lenet network is trained in batches.

[0049] Furthermore, H 0 Assumptions and H 1 Assume that two types of sample sets Ω are constructed respectively 0 and Ω 1 , forming a training set Ω=Ω 0 ∪Ω 1 .

[0050] Furthermore, in step (7), the training set H 0 Assume that M samples under the optimal model parameters are used as inputs to the 1D-Lenet network, and obtain M inputs belonging to H 1 The probability of the hypothesis β 1 ,β 2 ,β 3 ,....,β M ;

[0051] The statistic p 1 ,p 2 ,p 3 ,....,p M Arrange in descending order from large to small, that is, p 1 ≥p 2 ≥p 3 ≥....≥p M , at the false alarm rate P fa Next, the discrimination threshold

[0052]

[0053] Here, [x] means rounding x to the nearest integer.

[0054] The present invention also provides a sea surface small target detection device based on an improved 1D-Lenet network, comprising:

[0055] The first acquisition module is used to obtain the reference unit observation matrix Z of P reference units around the unit to be detected. p ,p=1,2,...,P, where the unit to be detected receives N continuous pulses under four polarizations respectively, forming a CUT observation matrix Z=[z HH ,z HV ,z VH ,z VV ],z HH ,z HV ,z VH ,z VV They represent the observation vectors under HH, HV, VH, and VV polarizations, respectively, with a dimension of N;

[0056] A conversion module, used to convert the CUT measurement matrix and the reference unit measurement matrix into the frequency domain, time-frequency domain, and polarization domain;

[0057] The feature extraction module is used to extract the features of the unit to be detected from the time domain, frequency domain, time-frequency domain, and polarization domain respectively, and the total number of features is recorded as D;

[0058] The feature conversion module is used to construct a D-dimensional feature space by taking each feature as a dimension and converting the D features extracted in the feature extraction module into a feature vector ξ=[ξ 1 ,ξ 2 ,…,ξ D ] T ;

[0059] The second acquisition module is used to take the feature vector ξ of the unit to be detected after the feature conversion module as input, enter the 1D-Lenet network with the optimal model parameters, and obtain the input belonging to H 1 The probability value of the hypothesis β, H 1 Assume that the unit to be detected contains a target echo, wherein the optimal model parameters of the 1D-Lenet network are extracted according to the pure clutter and the simulated target echo according to steps (2) to (4), as a training set, based on the 1D-Lenet network, the model parameters of the 1D-Lenet network are optimized based on the training set to obtain the optimal model parameters;

[0060] The assignment module is used to use the probability value β obtained by the second acquisition module as the final test statistic

[0061] The judgment module is used to convert the H 0 Assume that the sample under is used as the input of the 1D-Lenet network with the optimal model parameters, and obtain M inputs belonging to H 1 The probability of the hypothesis β 1 ,β 2 ,β 3 ,....,β M , H 0 Assume that the unit to be detected contains only sea clutter; according to the set false alarm rate P fa , and obtain the corresponding decision threshold γ; according to the test statistic The relationship between the detection threshold γ and the detection threshold γ is used to determine whether there is a target in the unit to be detected:

[0062] if It indicates that the unit to be detected contains target echo. If This indicates that the unit to be detected does not contain the target echo.

[0063] The present invention designs a Lenet network binary classifier, which solves the problem that the false alarm rate of high-dimensional feature classifiers is difficult to control through deep network learning of multi-dimensional features, and improves the detection performance of small floating targets under complex sea clutter background.

[0064] The present invention proposes a multi-dimensional feature detection based on a 1D-Lenet network. The main innovations are as follows. First, multi-dimensional features are extracted from the time domain, frequency domain, time-frequency domain, and polarization domain, and the energy, geometry, and other differences of sea clutter and target echoes in different domains are fully explored. Second, an improved 1D-Lenet network is designed to achieve controllable classification design of false alarms in high-dimensional feature space. Deep networks usually use a large number of convolution kernels and have a deeper structure. While improving performance, they also lead to the problem of excessive computational complexity and long training time during network training. However, the improved 1D-Lenet network has fewer parameters and enables the network to have the potential to achieve rapid detection.

[0065] Compared with the prior art, the present invention has the following advantages:

[0066] The present invention comprehensively extracts characteristic information from multiple domains such as time domain, frequency domain, time-frequency domain, polarization domain, etc., thereby increasing the distinction between sea clutter and target echoes.

[0067] The invention proposes an improved 1D-Lenet network, designs and optimizes the network structure, and solves the problem that the false alarm rate of the classifier is difficult to control in a high-dimensional feature space. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 Flow chart of the detection method of the present invention.

[0069] Figure 2 It is a 1D-Lenet network structure diagram of the detection method of the present invention.

[0070] Figure 3 This is a comparison result of the time consumption of three network trainings of the detection method of the present invention.

[0071] Figure 4 It is a result diagram of the detector based on the Hurst index of the detection method of the present invention.

[0072] Figure 5 It is a result diagram of the detector based on three features of the detection method of the present invention.

[0073] Figure 6 This is a result diagram of the feature detector based on the improved 1D-Lenet of the detection method of the present invention.

[0074] Figure 7 It is a structural diagram of the detection device of the invention. DETAILED DESCRIPTION

[0075] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention; it is obvious that the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments, and all other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making creative work are within the scope of protection of the present invention.

[0076] The present invention will be further described below in conjunction with the accompanying drawings.

[0077] Reference Figure 1 , the implementation steps of the present invention are as follows:

[0078] Step 1: Get the echo matrix

[0079] 1.1) For the full polarization radar in dwell mode, assume that the cell under test (CUT) receives N consecutive pulses under four polarizations, forming an observation matrix Z = [z HH ,z HV ,z VH ,z VV ], where z HH ,z HV ,z VH ,z VVRespectively represent the observation vectors under HH, HV, VH, and VV polarizations, with a dimension of N. Then, the measurement matrix Z of the P reference units around the CUT is obtained p ,p=1,2,...,P.

[0080] 1.2) The essence of detection is to determine whether the CUT contains the target echo. Therefore, the detection problem is classified as the following binary hypothesis testing problem

[0081]

[0082] Among them, c HH ,c HV ,c VH ,c VV denote the sea clutter vectors under HH, HV, VH, and VV polarizations, respectively, and s HH ,s HV ,s VH ,s VV Represent the target echo vectors under four polarizations, H represent the clutter vector of the pth reference unit under four polarizations. 0 Assume that the CUT contains only sea clutter, H 1 Assume that the CUT contains target echoes. The reference unit is pure clutter data, providing sea clutter information for the CUT.

[0083] Step 2: Calculate multi-domain data

[0084] 2.1) For each column of data in the measurement matrix of the CUT and the reference unit, the frequency domain amplitude spectrum (AS) is calculated respectively, and the formula is:

[0085]

[0086] Among them, z represents an N-dimensional vector, f represents a frequency variable, and f r represents the radar pulse repetition frequency, and |·| represents the amplitude.

[0087] 2.2) For each column of data in the measurement matrix of the CUT and the reference unit, the two-dimensional time-frequency diagram SPWVD of the echo vector z is calculated respectively, and the formula is as follows:

[0088]

[0089] Among them, n represents the time dimension, l represents the frequency dimension, Δf d is the sampling interval of the normalized Doppler frequency, Q and q represent the number of samples in the time dimension and the corresponding variable, J and j represent the number of samples in the frequency dimension and the corresponding variable, g(q) and h(p) are the smoothing windows of the time dimension and frequency dimension, respectively.

[0090] Step 3: Extract features

[0091] 3.1) Time Domain Characteristics

[0092] The Hurst index calculation formula is as follows

[0093]

[0094] Where k and b represent the slope and intercept of the fitted line, and m represents the sequence interval.

[0095] The relative average amplitude (RAA) calculation formula is:

[0096]

[0097] 3.2) Frequency domain characteristics

[0098] The calculation formula of relative Doppler peak height (RDPH) is:

[0099]

[0100] Among them, DPH represents the Doppler peak height position, Peak represents the maximum value of the Doppler peak, represents the position of the Doppler peak in the Doppler spectrum, the symbol #Δ represents the total number falling into the number set Δ, and Δ represents the Doppler interval around the Doppler peak.

[0101] The calculation formula of relative vector entropy (RVE) is:

[0102]

[0103] 3.3) Time-frequency domain features

[0104] The ridge intensity (RI) is calculated as follows:

[0105]

[0106] At each time scale of the time-frequency graph, take the first 5 bright pixels and mark them as 1 and the remaining pixels as 0 to generate a binary image. On the binary image, get the connected region Ω 1 ,Ω 2 ,...,Ω w The number of connected regions (CR) and the maximum size of connected regions (MS) are obtained by the following formulas:

[0107] CR(z)=W

[0108]

[0109] Where W represents the number of connected regions, #Ω w It means taking the non-zero element of the w-th connected region.

[0110] 3.4) Polarization domain characteristics

[0111] For each polarization vector of the observation matrix Z in the CUT, the Hurst index and RAA are extracted, that is,

[0112] ξ 1 = Hurst(z HH ),ξ 2 = Hurst(z HV ),ξ 3 = Hurst(z VH ),ξ 4 = Hurst(z VV )

[0113] ξ 5 =RAA(z HH ),ξ 6 =RAA(z HV ),ξ 7 =RAA(z VH ),ξ 8 =RAA(z VV )

[0114] For each polarization vector of the observation matrix Z in the CUT, RDPH and RVE are extracted, that is,

[0115] ξ 9 =RDPH(z HH ),ξ 10 =RDPH(z HV ),ξ 11 =RDPH(z VH ),ξ 12 =RDPH(z VV )

[0116] ξ 13 =RVE(z HH ),ξ 14 =RVE(z HV ),ξ 15 =RVE(z VH ),ξ 16 =RVE(z VV )

[0117] For each polarization vector of the observation matrix Z in the CUT, RI, CR and MS are extracted, that is,

[0118] ξ 17 =RI(z HH ),ξ 18 =RI(z HV ),ξ 19 =RI(z VH ),ξ 20 =RI(z VV )

[0119] ξ 21 =CR(z HH ),ξ 22 =CR(z HV ),ξ 23 =CR(z VH ),ξ 24 =CR(z VV )

[0120] ξ 25 =MS(z HH ),ξ 26 =MS(z HV ),ξ 27 =MS(z VH ),ξ 28 =MS(z VV )

[0121] Step 4: Construct feature vector

[0122] Take each feature as a dimension and construct a D-dimensional feature space. Then, convert the D features in step (3) into a feature vector

[0123] ξ=[ξ 1 ,ξ 2 ,…,ξ D ] T

[0124] Step 5: Build a network model

[0125] 5.1) Based on a large number of pure clutter and simulated target echoes, extract the feature vector samples under two assumptions according to steps (2) to (4). Assume that H 0 Assume that and H 1 Assume that there are M feature vectors in total, and construct two types of sample sets Ω respectively. 0 and Ω 1 Therefore, 2M feature vector samples constitute the training set Ω = Ω 0 ∪Ω 1

[0126] 5.2) According to Figure 2As shown in the figure, a 1D-Lenet network is built, including C1 convolution layer, S2 downsampling layer, C3 convolution layer, S4 downsampling layer, C5 convolution layer, F6 fully connected layer and the final output Output layer. The original 2D-Lenet network is improved from four aspects: first, the three convolution layers are changed to convolution kernels with a size of 3 and a step size of 1 to obtain a 1D-Lenet network; second, the PRelu nonlinear activation function is used to replace the sigmoid linear activation function to optimize the network parameter training process; third, the Dropout operation is added to alleviate the overfitting problem; fourth, batch training is added to improve the network classification accuracy. Therefore, the improved 1D-Lenet network can speed up the training speed while improving the classification performance of the network.

[0127] 5.3) Take the training set Ω in step 5.1) as input and train the 1D-Lenet network built in step 5.2). Through multiple iterations of learning and updating, the optimal model parameters can be finally obtained.

[0128] 5.4) Take the feature vector of CUT in step (4) as input and enter the 1D-Lenet network with the optimal model parameters to obtain the input belonging to H 1 The probability of the hypothesis is β.

[0129] Step 6: Get the test statistic

[0130] Use the probability value β in step 5.4) as the final test statistic

[0131]

[0132] Step 7: Determine the test result

[0133] 7.1) Set the training set H 0 Assume that M samples under the optimal model parameters are used as inputs to the 1D-Lenet network, and obtain M inputs belonging to H 1 The probability of the hypothesis β 1 ,β 2 ,β 3 ,....,β M .

[0134] 7.2) The statistic p 1 ,p 2 ,p 3 ,....,p M Arrange in descending order from large to small, that is, p 1 ≥p 2 ≥p 3 ≥....≥p M Then, at the false alarm rate P faNext, the discrimination threshold

[0135]

[0136] Here, [x] means rounding x to the nearest integer.

[0137] 7.3) Determine the test statistic The relationship with the decision threshold γ determines whether there is a target in the CUT:

[0138] if This indicates that the CUT contains target echo.

[0139] if This indicates that there is no target echo in the CUT.

[0140] The above method is implemented based on a small target detection device on the sea surface based on an improved 1D-Lenet network. Figure 7 As shown, the device comprises:

[0141] The first acquisition module is used to obtain the reference unit observation matrix Z of P reference units around the unit to be detected. p ,p=1,2,...,P, where the unit to be detected receives N continuous pulses under four polarizations respectively, forming a CUT observation matrix Z=[z HH ,z HV ,z VH ,z VV ],z HH ,z HV ,z VH ,z VV They represent the observation vectors under HH, HV, VH, and VV polarizations, respectively, with a dimension of N;

[0142] A conversion module, used to convert the CUT measurement matrix and the reference unit measurement matrix into the frequency domain, time-frequency domain, and polarization domain;

[0143] The feature extraction module is used to extract the features of the unit to be detected from the time domain, frequency domain, time-frequency domain, and polarization domain respectively, and the total number of features is recorded as D;

[0144] The feature conversion module is used to construct a D-dimensional feature space by taking each feature as a dimension and converting the D features in step 3 into a feature vector ξ = [ξ 1 ,ξ 2 ,…,ξ D ] T ;

[0145] The second acquisition module is used to take the feature vector of the unit to be detected in step 4 as input, enter the 1D-Lenet network with the optimal model parameters, and obtain the input belonging to H1 The probability value of the hypothesis β, H 1 Assume that the unit to be detected contains a target echo, wherein the optimal model parameters of the 1D-Lenet network are extracted according to the pure clutter and the simulated target echo according to steps (2) to 4, as a training set, based on the 1D-Lenet network, the model parameters of the 1D-Lenet network are optimized based on the training set to obtain the optimal model parameters;

[0146] The assignment module is used to use the probability value β in step 5 as the final test statistic

[0147] The judgment module is used to convert the H 0 Assume that the sample under is used as the input of the 1D-Lenet network with the optimal model parameters, and obtain M inputs belonging to H 1 The probability of the hypothesis β 1 ,β 2 ,β 3 ,....,β M , H 0 Assume that the unit to be detected contains only sea clutter; according to the set false alarm rate P fa , and obtain the corresponding decision threshold γ; according to the test statistic The relationship between the detection threshold γ and the detection threshold γ is used to determine whether there is a target in the unit to be detected:

[0148] if It indicates that the unit to be detected contains target echo. If This indicates that the unit to be detected does not contain the target echo.

[0149] The effects of the present invention are further described below in conjunction with the experimental results of measured data.

[0150] 1. Experimental Data

[0151] This example uses a database collected by the IPIX radar in 1993, which is publicly available on the Internet. This database is currently internationally recognized as small target measurement data. In the experiment, #310 data, HH polarization, X-band radar pulse repetition frequency of 1000Hz, distance resolution of 30m, and the test target is a small ball with a diameter of 1m, floating with the waves.

[0152] 2. Simulation Experiment

[0153] Simulation 1 compares the time consumption during the training process of the classic Lenet network, the Lenet network using a one-dimensional convolution kernel, the Alexnet network, and the improved 1D-Lenet network proposed in the present invention. The results are as follows Figure 3 As shown, where:

[0154] In one round of training, the Lenet network using a one-dimensional convolution kernel only takes 5 to 6 seconds to complete the training, while the classic Lenet network with different convolution kernels takes 15 seconds to complete a round of training, which shows that the use of one-dimensional convolution kernels greatly reduces the time for network training initialization. In multiple rounds of training, the 1D-Lenet network takes much less time to train than the Alexnet network and the classic Lenet, which shows that the use of the PRelu activation function with less computational effort and the batch training method suitable for one-dimensional vector input greatly improves performance and reduces computational time to a certain extent.

[0155] Simulation 2 verifies the detection algorithm proposed in this invention. The cumulative number of pulses is 512 and the false alarm rate is set to 10. -3 , the target is in the 9th distance unit, and the reference unit P = 10. The result is as follows Figure 4 , Figure 5 , Figure 6 As shown, where:

[0156] exist Figure 4 In the example, the detection probability of the detector based on the Hurst exponent is 0.133, which is a poor performance. This is because the detector based on the Hurst exponent only relies on a single feature and cannot obtain satisfactory results under low signal-to-clutter ratio conditions.

[0157] exist Figure 5 In the figure, the detection probability of the detector based on three features is 0.598. After combining the time domain and frequency domain features, the performance is significantly improved, which shows the effectiveness of multi-feature joint detection. However, the convex hull algorithm used in the detector cannot achieve false alarm control in high-dimensional feature space, and thus cannot combine more features to further improve the detection performance.

[0158] exist Figure 6 In the experiment, the detector based on the improved 1D-Lenet network proposed by the present invention has a detection probability of 0.990, which is the best performance. This is because the detector uses 28 features in the time domain, frequency domain, time-frequency domain, and polarization domain, which greatly improves the performance of small detection under low signal-to-noise ratio. In addition, the improved 1D-Lenet network realizes the design of binary classifiers in high-dimensional features, has a controllable false alarm rate, and meets the detection requirements of full polarization radar in resident mode.

[0159] In summary, the detection performance of the present invention for small targets on the sea surface is better than that of existing detection methods. It integrates the advantages of deep network autonomous learning and multi-dimensional feature combination, and proposes new ideas for traditional detection methods.

[0160] The above is only a preferred specific implementation of the present invention; however, the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solution and its improved conception within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for detecting small targets on the sea surface based on an improved 1D-Lenet network, characterized in that: The following steps are involved: (1) Obtain the reference unit observation matrix of P reference units around the unit to be detected , where the unit to be detected receives N continuous pulses under four polarizations respectively, forming a CUT observation matrix , They represent the observation vectors under HH, HV, VH, and VV polarizations, respectively, with a dimension of N; (2) Convert the CUT measurement matrix and the reference unit measurement matrix into the frequency domain, time-frequency domain, and polarization domain; (3) Extract the features of the unit to be detected from the time domain, frequency domain, time-frequency domain, and polarization domain respectively, and record the total number of features as D; (4) Take each feature as a dimension and construct a D-dimensional feature space; then, convert the D features in step (3) into a feature vector of the unit to be detected ; (5) Taking the feature vector of the unit to be detected in step (4) as input, entering the 1D-Lenet network with optimal model parameters, and obtaining the probability value β that the input belongs to the H1 hypothesis, where the H1 hypothesis indicates that the unit to be detected contains a target echo. The optimal model parameters of the 1D-Lenet network are extracted according to steps (2) to (4) based on pure clutter and simulated target echoes. The feature vector samples are used as a training set. Based on the 1D-Lenet network, the model parameters of the 1D-Lenet network are optimized based on the training set to obtain the optimal model parameters. (6) Use the probability value β in step (5) as the final test statistic φ; (7) Take the samples under the training set H0 hypothesis as the input of the 1D-Lenet network with the optimal model parameters, and obtain the probability that the M inputs belong to the H1 hypothesis , H0 hypothesis indicates that the unit to be detected contains only sea clutter; according to the set false alarm rate P fa , and obtain the corresponding decision threshold γ; according to the relationship between the test statistic φ and the decision threshold γ, it is determined whether there is a target in the unit to be detected: if , it indicates that the unit to be detected contains target echo. If , it indicates that the unit to be detected does not contain the target echo; Among them, the formulas of H0 hypothesis and H1 hypothesis are: in, denote the sea clutter vectors under HH, HV, VH, and VV polarizations, respectively, Represent the target echo vectors under four polarizations, They respectively represent the clutter vectors of the p-th reference unit under four polarizations. The reference unit is pure clutter data and is used to provide sea clutter information for the unit to be detected.

2. According to claim 1, a method for detecting small targets on the sea surface based on an improved 1D-Lenet network is characterized in that: In step (2), for each column of data in the measurement matrix of the reference unit to be detected, the frequency domain amplitude spectrum AS is calculated respectively, and the formula is: in, represents the N-dimensional echo vector, f represents the frequency variable, fr represents the radar pulse repetition frequency, represents the amplitude, n represents the nth time dimension, and z(n) represents the amplitude of the echo in the nth time dimension; For each column of data in the measurement matrix of the reference unit to be detected, the echo vector is calculated respectively. The two-dimensional time-frequency diagram SPWVD is as follows: Among them, n represents the nth time dimension, l represents the frequency dimension, is the sampling interval of the normalized Doppler frequency, Q and q represent the number of samples in the time dimension and the corresponding variable, J and j represent the number of samples in the frequency dimension and the corresponding variable, g(q) and h(p) are the smoothing windows in the time dimension and frequency dimension, respectively.

3. The method for detecting small targets on the sea surface based on an improved 1D-Lenet network according to claim 2, characterized in that: In step (3), in the time domain feature extraction, the Hurst exponent calculation formula is as follows: Among them, k and b represent the slope and intercept of the fitting line; represents the sequence interval, n represents the nth time dimension, and z(n) represents the amplitude of the echo in the nth time dimension; The relative average amplitude RAA is calculated as: Wherein, P is the total number of reference units, and p represents the pth reference unit.

4. A device for implementing the sea surface small target detection method based on improved 1D-Lenet network as claimed in claim 1, characterized in that: include: The first acquisition module is used to obtain the reference unit observation matrix of P reference units around the unit to be detected , where the unit to be detected receives N continuous pulses under four polarizations respectively, forming a CUT observation matrix , They represent the observation vectors under HH, HV, VH, and VV polarizations, respectively, with a dimension of N; A conversion module, used to convert the CUT measurement matrix and the reference unit measurement matrix into the frequency domain, time-frequency domain, and polarization domain; The feature extraction module is used to extract the features of the unit to be detected from the time domain, frequency domain, time-frequency domain, and polarization domain respectively, and the total number of features is recorded as D; The feature conversion module is used to construct a D-dimensional feature space by taking each feature as a dimension and converting the D features extracted in the feature extraction module into a feature vector of a unit to be detected. ; The second acquisition module is used to convert the feature vector of the unit to be detected after the feature conversion module As input, enter the 1D-Lenet network with optimal model parameters, obtain the probability value β that the input belongs to the H1 hypothesis, the H1 hypothesis indicates that the unit to be detected contains the target echo, wherein the optimal model parameters of the 1D-Lenet network are extracted according to the pure clutter and the simulated target echo according to steps (2) to (4), as the training set, based on the 1D-Lenet network, the model parameters of the 1D-Lenet network are optimized based on the training set to obtain the optimal model parameters; An assignment module, used for taking the probability value β obtained by the second acquisition module as the final test statistic φ; The judgment module is used to use the samples under the H0 hypothesis of the training set as the input of the 1D-Lenet network with the optimal model parameters to obtain the probability that the M input belongs to the H1 hypothesis. , H0 hypothesis indicates that the unit to be detected contains only sea clutter; according to the set false alarm rate P fa , and obtain the corresponding decision threshold γ; according to the relationship between the test statistic φ and the decision threshold γ, it is determined whether there is a target in the unit to be detected: If φ , it indicates that the unit to be detected contains the target echo. If φ , it indicates that the unit to be detected does not contain target echo.

Images