High-resolution target bright spot extraction method under reverberation background

Abstract

The invention provides a high-resolution target bright spot extraction method under a reverberation background, belongs to the field of underwater acoustic signal processing, and particularly relates to the field of active sonar identification research. In order to improve the performance of bright spot extraction, time-frequency domain reverberation and target echo correlation are modeled into a low-rank matrix and a non-low-rank matrix respectively according to the difference between time-frequency domain reverberation and target echo correlation. Modeling is carried out on a low-rank matrix under a Bayesian framework, and Laplacian and generalized inverse Gaussian mixed distribution priori is given to a non-low-rank part so as to enhance robustness of matrix recovery. After Bayesian inference is completed through MCMC sampling, RLSTFRFT processing is carried out on a time-frequency matrix formed by target echoes after reverberation suppression, so that the time-frequency resolution is improved, and the bright spot extraction performance is comprehensively improved. Simulation and experiment results verify that the method is effective, and compared with an existing method, the performance of the method is obviously improved. According to the method, the bright spot extraction performance is improved, and a technical basis is provided for underwater target recognition.

Description

Technical Field

[0001] The invention belongs to the field of underwater acoustic signal processing, more precisely to the category of underwater active sonar recognition, and is mainly a method for extracting high-resolution target bright spots under a reverberation background. Background Art

[0002] The recognition technology of underwater targets is one of the main technologies to be studied in sonar technology. The correct recognition of underwater targets is based on effective feature extraction. The target bright spot model is a practical target model based on the generation mechanism of echoes. The target bright spot structure is also a widely used underwater target feature. Therefore, the target echo bright spot and the extraction of bright spots have become the key to underwater target recognition. At present, the research on target echo bright spots can be mainly divided into three categories: theoretical and prediction research on underwater target echo bright spots, research on target echo bright spots through measured data, and target echo simulation and recognition research based on bright spot models. The target echo bright spot contains the information of the target. The better the effect of bright spot extraction, the more beneficial it is to the recognition of the target. The research on target echo bright spot extraction can currently be divided into one-dimensional target echo (i.e., instantaneous signal) bright spot extraction and two-dimensional target echo (i.e., beam domain signal) bright spot extraction. The research on one-dimensional target echo bright spot extraction is represented by methods such as matched filtering and time-frequency analysis. Two-dimensional target echo bright spot extraction can be effectively completed by existing image sonar combined with array processing technology. In the past decade, the low-rank matrix recovery theory has developed rapidly and has now been widely used in the field of image processing. In recent years, it has also been gradually applied to underwater acoustic engineering. In terms of target highlight extraction, Zhu Guangping et al. separated the target echo and reverberation in the time-frequency domain based on the different correlations between the energy distribution of the target echo and reverberation in the time-frequency domain. This work provides a new idea for underwater target highlight extraction. However, this work still has shortcomings such as insufficient robustness, performance is greatly affected by parameter adjustment, does not consider the clustering of highlights, and does not improve the performance of highlights from the perspective of resolution. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for extracting high-resolution target highlights in a reverberation background, further improve the highlight extraction performance, and provide a technical basis for active sonar recognition.

[0004] The object of the present invention is achieved through the following technical solution: A method for extracting high-resolution target highlights in a reverberation background, comprising the following steps:

[0005] Step 1: Perform time-frequency analysis on the active sonar receiving signal to obtain its time-frequency distribution;

[0006] Step 2: Perform Bayesian modeling on the low-rank matrix L representing the reverberation time-frequency matrix;

[0007] Step 3: Perform Bayesian modeling on the received signal time-frequency matrix Y, and assign Laplace and generalized inverse Gaussian mixture distribution priors to the non-low-rank part representing the target echo;

[0008] Step 4: Place a first-order Markov random field on the non-low-rank part of the time-frequency matrix to ensure the clustering effect of the target echo highlights in the time-frequency domain;

[0009] Step 5: Implement the Bayesian inference of the model through MCMC sampling, and output the low-rank part representing the reverberation and the non-low-rank part representing the target echo;

[0010] Step 6: After improving the time-frequency resolution of the non-low-rank part of the output through deconvolution short-time fractional Fourier transform, the target echo highlight extraction result is output.

[0011] The time-frequency analysis of the active sonar receiving signal is as follows: the short-time fractional Fourier transform STFRFT of the active sonar receiving signal is performed to obtain its time-frequency matrix. The expression of STFRFT is:

[0012]

[0013] Among them, s(t) is the transmitted signal, g(t) is the window function, α is the transformation angle, K α (t,u) is the transformation kernel, u is the parameter of fractional Fourier transform,

[0014] Furthermore, the obtained time-frequency matrix Y is expressed as

[0015] Y=L+S

[0016] Among them, L is a low-rank matrix representing the time-frequency matrix of reverberation, and S is a non-low-rank matrix representing the time-frequency matrix of target echo;

[0017] The time-frequency matrix of the received signal Bayesian modeling is performed as

[0018]

[0019] η ij ~GIG(p,a,b)

[0020] in Indicates that the mean is a i b j T , with a scale of η ij Laplace distribution, GIG represents the generalized inverse Gaussian distribution, and p, a, and b are hyperparameters.

[0021] In the step 4, in order to achieve the clustering effect of non-low-rank matrix pixels (target echoes in the time-frequency matrix), Place a first-order Markov random field in and define a potential function for the adjacent pixels (i, j) and (p, q)

[0022]

[0023] ψ(Λ ij ,Λ pq )=exp{-α|logΛ ij -logΛ pq |}

[0024] The strength of the prior is controlled by the parameter α.

[0025] In step 6, the output matrix S is subjected to deconvolution short-time fractional Fourier transform RLSTFRFT to improve the time-frequency resolution, and the target bright spot extraction result is output. The iterative update expression of RLSTFRFT is:

[0026]

[0027] in Represents the correlation operation, k is the number of iterations, and the initial value RLFRFT0 is set to FRFT.

[0028] The beneficial effects of the present invention are as follows: the present invention first performs a short-time fractional Fourier transform (STFRFT) on the received signal of the active sonar to convert the received signal into the time-frequency domain. When the transmitted signal is a linear frequency modulation signal, the echo signal is a small number of oblique lines in the time-frequency domain, showing sparsity, while the reverberation within the bandwidth has correlation and shows low rank in the time-frequency domain. Under the Markov Bayesian robust matrix decomposition (MBRMF) model, the low-rank matrix composed of the reverberation is Bayesian modeled, and the sparse matrix composed of the target echo is given Laplace and generalized inverse Gaussian mixture distribution priors. And in order to ensure the clustering of the time-frequency matrix elements of the target echo bright spot, a first-order Markov random field is placed between the sparse matrix elements. Then, the inference of the parameters in the Bayesian model is realized through MCMC sampling, thereby realizing low-rank sparse decomposition, that is, the separation of reverberation and target bright spots. The sparse matrix composed of target bright spots is deconvolved in the fractional domain through the deconvolution short-time fractional Fourier transform (RLSTFRFT), which further improves the resolution of target bright spots and comprehensively improves the performance of bright spots. Simulation and experimental results show that this method is effective for extracting target echo bright spots. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art or ordinary technicians, other drawings can be obtained based on these drawings without paying creative work.

[0030] Figure 1 This is a flow chart of the method for extracting high-resolution target highlights in a reverberant background.

[0031] Figure 2 The target model used for simulation is a rigid cylinder with a truncated cone at one end and a hemisphere at the other end.

[0032] Figure 3 This is a schematic diagram of the simulation results when the signal-to-mix ratio is -5dB.

[0033] Figure 4 Schematic diagram of the curve of the third-order Rayleigh entropy obtained by different methods as the signal-to-mixing ratio changes.

[0034] Figure 5 Schematic diagram of experimental processing results for real targets. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of the present invention.

[0036] like Figure 1 As shown, the present invention proposes a high-resolution target bright spot extraction method under reverberation background, completes the separation of target echo and reverberation in time and frequency domain under the Bayesian framework, and uses fractional deconvolution on this basis to further improve the bright spot resolution, which specifically includes the following steps:

[0037] 1. First, perform short-time fractional Fourier transform (STFRFT) on the active sonar receiving signal to obtain its time-frequency matrix. The expression of STFRFT is:

[0038]

[0039] Among them, s(t) is the transmitted signal, g(t) is the window function, α is the transformation angle, K α (t,u) is the transformation kernel, u is the parameter of fractional Fourier transform,

[0040] 2. The obtained time-frequency matrix Y is expressed as

[0041] Y=L+S

[0042] Among them, L is a low-rank matrix representing the time-frequency matrix of reverberation, and S is a sparse (non-low-rank) matrix representing the time-frequency matrix of target echo.

[0043] 3. Bayesian modeling of the low-rank matrix L is

[0044] L=AB T

[0045]

[0046] in, r<<m,n;a i and b j denote the i-th row of A and the j-th row of B respectively, is a normal distribution with mean μ and variance τ; μ0, v0, β0, W0 are hyperparameters, represents a Wishart distribution with scaling matrix W and v degrees of freedom.

[0047] 4. Time-frequency matrix Bayesian modeling is performed as

[0048]

[0049] η ij ~GIG(p,a,b)

[0050] in Indicates that the mean is a i b j T , with a scale of η ij Laplace distribution, GIG represents the generalized inverse Gaussian distribution, and p, a, and b are hyperparameters.

[0051] 5. In order to achieve the clustering effect of non-low-rank matrix pixels (target echoes in the time-frequency matrix), Place a first-order Markov random field in. Define a potential function for the adjacent pixels (i, j) and (p, q)

[0052]

[0053] ψ(Λ ij ,Λ pq )=exp{-α|logΛ ij -logΛ pq |}

[0054] The strength of the prior is controlled by the parameter α.

[0055] 6. The established Bayesian model can be approximated by MCMC sampling, thereby realizing the recovery of the low-rank matrix L. The basic logic of MCMC sampling is to obtain a series of independent samples from the target posterior distribution, thereby realizing the approximation of the target distribution. Due to space limitations, the inference results are no longer listed.

[0056] 7. The target echo matrix S after reverberation suppression is output through S=YL.

[0057] 8. The output matrix S is subjected to deconvolution short-time fractional Fourier transform (RLSTFRFT) to improve the time-frequency resolution and output the target highlight extraction result. The iterative update expression of RLSTFRFT is:

[0058]

[0059] in Represents the correlation operation, k is the number of iterations, and the initial value RLFRFT0 is generally set to FRFT.

[0060] like Figure 3 The simulation results are shown when the signal-to-mix ratio is -5dB. Figure 3 (a) is the result after STFRFT. Figure 3 (b) for Figure 3 (a) The result after using the existing reverberation suppression method based on principal component tracking (PCP). Figure 3 (c) is the result after reverberation suppression by this method. Figure 3 (d) for Figure 3 (c) The result of using fractional deconvolution to improve resolution. In the simulation, the distance between the receiver and the target is 100m. The timing of each bright spot of the target is obtained by the pulse theory, and the amplitude of each bright spot is calculated by physical acoustics. The transmitted signal is a linear frequency modulation signal with a frequency band of 30kHz-40kHz, a pulse width of 1ms, and a sampling rate of 200kHz. It can be seen from the result graph that compared with the existing PCP method, this method has better reverberation suppression ability, improves the time-frequency resolution, and comprehensively improves the performance of the bright spot.

[0061] like Figure 4 The third-order Rayleigh entropy of the results obtained by different methods is shown as the signal-to-mixture ratio changes. The smaller the third-order Rayleigh entropy, the more concentrated the energy in the time-frequency domain, and the better the bright spot extraction effect. From the result graph, it can be seen that the matrix decomposition method can effectively improve the bright spot extraction performance, and the performance of this method is better than the existing methods.

[0062] like Figure 5 The experimental processing results for real targets are shown. Figure 5 (a) is the result of STFRFT of the received data. Figure 5(b) is the result after being processed by this method. The target in the experiment is a hollow spherical crown cylinder with a steel shell made according to a scaled model. The top is a hemisphere with a diameter of 6 cm, and the bottom is a cylinder with the same diameter as the hemisphere and a height of 18 cm. The transmission signal uses a linear frequency modulation signal with a frequency band of 340kHz-440kHz, a pulse width of 0.5ms, a sampling rate of 20MHz, and is transmitted vertically to the target. The result graph shows that this method is still effective in the experimental data.

[0063] In summary, in order to improve the performance of bright spot extraction, the present invention models the two as low-rank matrices and non-low-rank matrices respectively according to the difference in correlation between reverberation and target echo in the time-frequency domain. The low-rank matrix is ​​modeled under the Bayesian framework, and the non-low-rank part is given Laplace and generalized inverse Gaussian mixture distribution priors to enhance the robustness of matrix recovery. After completing Bayesian inference through MCMC sampling, the time-frequency matrix composed of the target echo after reverberation suppression is subjected to RLSTFRFT processing to improve the time-frequency resolution and comprehensively improve the performance of bright spot extraction. Simulation and experimental results verify the effectiveness of this method, and the performance is significantly improved compared to the existing methods. The present invention achieves the improvement of bright spot extraction performance and also provides a technical basis for underwater target recognition.

[0064] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for extracting high-resolution target highlights in a reverberation background, characterized by: The steps include: Step 1: Perform time-frequency analysis on the active sonar receiving signal to obtain its time-frequency distribution; Step 2: Perform Bayesian modeling on the low-rank matrix L representing the reverberation time-frequency matrix; Step 3: Perform Bayesian modeling on the received signal time-frequency matrix Y, and assign Laplace and generalized inverse Gaussian mixture distribution priors to the non-low-rank part representing the target echo; Step 4: Place a first-order Markov random field on the non-low-rank part of the time-frequency matrix to ensure the clustering effect of the target echo highlights in the time-frequency domain; Step 5: Implement the Bayesian inference of the model through MCMC sampling, and output the low-rank part representing the reverberation and the non-low-rank part representing the target echo; Step 6: After improving the time-frequency resolution of the non-low-rank part of the output through deconvolution short-time fractional Fourier transform, the target echo highlight extraction result is output.

2. The method for extracting high-resolution target highlights under reverberation background according to claim 1, characterized in that: The time-frequency analysis of the active sonar receiving signal is as follows: the short-time fractional Fourier transform STFRFT of the active sonar receiving signal is performed to obtain its time-frequency matrix. The expression of STFRFT is: Among them, s(t) is the transmitted signal, g(t) is the window function, α is the transformation angle, K α (t,u) is the transformation kernel, u is the parameter of fractional Fourier transform, 3. The method for extracting high-resolution target highlights under reverberation background according to claim 2, characterized in that: The obtained time-frequency matrix Y is expressed as Y = L + S Among them, L is a low-rank matrix representing the time-frequency matrix of reverberation, and S is a non-low-rank matrix representing the time-frequency matrix of target echo; The time-frequency matrix of the received signal Bayesian modeling is performed as η ij ~GIG(p,a,b) in Indicates that the mean is a i b j T , with a scale of η ij Laplace distribution, GIG represents the generalized inverse Gaussian distribution, and p, a, and b are hyperparameters.

4. The method for extracting high-resolution target highlights under reverberation background according to claim 3, characterized in that: In the step 4, Place a first-order Markov random field in and define a potential function for the adjacent pixels (i, j) and (p, q) ψ(Λ ij ,L pq )=exp{-α|logΛ ij -logΛ pq |} The strength of the prior is controlled by the parameter α.

5. The method for extracting high-resolution target highlights under reverberation background according to claim 4, characterized in that: In step 6, the output matrix S is subjected to deconvolution short-time fractional Fourier transform RLSTFRFT to improve the time-frequency resolution, and the target bright spot extraction result is output. The iterative update expression of RLSTFRFT is: in Represents the correlation operation, k is the number of iterations, and the initial value RLFRFT0 is set to FRFT.

Images