Reverberation suppression method based on time-frequency patch tensor model

Abstract

The invention belongs to the technical field of underwater acoustic engineering. The invention provides a reverberation suppression method based on a time-frequency patch tensor model. According to the embodiment of the invention, the reverberation suppression problem is modeled as the optimization problem of low-rank sparse decomposition by utilizing the low-rank sparse characteristic of the echo signal in the time-frequency domain. And decomposing the time-frequency patch tensor into a low-rank patch tensor and a sparse patch by using an alternating direction multi-multiplier method. The sparse patch tensor mainly comprises a target echo component and a small part of reverberation components; the low-rank patch tensor is mainly a reverberation component. The low-rank patch tensor comprises most reverberation components, so that the signal-to-mix ratio of the obtained sparse patch tensor is remarkably improved, and reverberation suppression is realized.

Description

Technical Field

[0001] The disclosed embodiments relate to the field of underwater acoustic engineering technology, and in particular to a reverberation suppression method based on a time-frequency patch tensor model. Background Art

[0002] When using active sonar to detect underwater targets, reverberation will interfere with the target echo signal. Reverberation is a random signal, which is a signal generated by the random scattering and superposition of sound waves when they encounter a large number of uneven scatterers underwater. Reverberation is one of the important factors affecting sonar detection performance, especially in shallow waters. How to effectively reduce the interference of reverberation on target detection has always been one of the research hotspots in the field of underwater acoustic signal processing.

[0003] Reverberation is strongly correlated with the transmitted signal and is colored and non-stationary, which limits the performance of conventional matched filters. In the related art, a one-dimensional autoregressive pre-whitening processing method is proposed. This method is a sub-optimal signal detection method. It assumes the temporal local stationarity of the reverberation background and uses the AR model coefficients of the current data segment to "whiten" the next data segment. However, due to the complex underwater environment, the reverberation pre-whitening method may fail in an environment with poor reverberation stationarity. Based on the power difference between reverberation and signal, a reverberation suppression algorithm based on principal component inversion is also proposed. This algorithm can effectively separate the reverberation in the echo signal. How to estimate the reverberation and signal power is the key to the success of this method. However, the time-frequency characteristics of reverberation are related to the transmitted signal. In the time domain, the reverberation overlaps with the target echo, and in the frequency domain, the reverberation is coherent with the target echo. It can be seen that it is difficult to remove reverberation from the time domain or frequency domain alone, and it is necessary to study a reverberation suppression method that combines time domain and frequency domain information.

[0004] Therefore, it is necessary to improve one or more problems existing in the above-mentioned related technical solutions.

[0005] It should be noted that this section is intended to provide background or context for the technical solutions of the present disclosure stated in the claims. The description herein is not admitted to be prior art by virtue of being included in this section. Summary of the invention

[0006] The purpose of the embodiments of the present disclosure is to provide a reverberation suppression method based on a time-frequency patch tensor model, thereby overcoming one or more problems caused by limitations and defects of related technologies at least to a certain extent.

[0007] According to an embodiment of the present disclosure, a reverberation suppression method based on a time-frequency patch tensor model is provided, the method comprising:

[0008] Performing short-time Fourier transform on the echo data collected once, and processing the echo data after the short-time Fourier transform by using a modulo operation to obtain a time-frequency matrix and phase information;

[0009] constructing a time-frequency patch tensor based on the time-frequency matrix;

[0010] Decomposing the time-frequency patch tensor using an alternating direction multi-multiplier method to obtain a sparse patch tensor;

[0011] Reconstructing the sparse patch tensor to obtain a sparse time-frequency matrix;

[0012] An inverse short-time Fourier transform is performed on the phase information and the sparse time-frequency matrix to obtain time domain data after reverberation suppression.

[0013] Further, the step of performing short-time Fourier transform on the echo data collected once, and processing the echo data after the short-time Fourier transform by using a modulo operation to obtain a time-frequency matrix and phase information includes:

[0014] Perform a short-time Fourier transform on the echo signal m of length L, and obtain the time-frequency matrix by using a modulo operation. and the phase information in,

[0015] The expression for performing short-time Fourier transform on the discretized echo signal m is:

[0016]

[0017] The window function uses the Hanning window, which is defined as follows:

[0018]

[0019] Wherein, w[n] is the window function, the length of the window function is M, n is the position of the current window, N is the number of DFT points, R is the number of overlapping points, a is the frequency index, b is the time index, M(a,b)=M is ​​the time-frequency matrix, k is the frequency point of the short-time Fourier transform, and j is the imaginary unit;

[0020] A is the number of frequency resolution units, A = N / 2, B is the number of time resolution units, Indicates rounding down.

[0021] Furthermore, the step of constructing a time-frequency patch tensor based on the time-frequency matrix includes:

[0022] Using a sliding window from top left to bottom right Slide with a preset sliding step size to obtain a local patch matrix; where size is the dimension of the sliding window;

[0023] Stack the patch matrices into the time-frequency patch tensor

[0024] Furthermore, the number n1 of the sliding windows is:

[0025]

[0026] In the formula, step h is the sliding step length in the vertical direction, step v is the sliding step length in the horizontal direction.

[0027] Furthermore, the step of decomposing the time-frequency patch tensor by using an alternating direction multi-multiplier method to obtain a sparse patch tensor includes:

[0028] Let the low-rank patch tensor be The sparse patch tensor is The Lagrange multiplier is The initial values ​​of the low-rank patch tensor, the sparse patch tensor and the Lagrange multiplier are set to and Let k be the number of iterations, and take k = 0; the initial value of the penalty factor is

[0029] Calculate the low-rank patch tensor of the k-th iteration; wherein, at the k-th iteration, the calculation expression of the low-rank patch tensor is:

[0030]

[0031] In the formula, fold i To restore the matrix expanded in mode-i to the original tensor, is the threshold shrinkage operator for the matrix singular values, X=Udiag(σ)V T is the SVD decomposition of matrix X,

[0032] According to the low-rank patch tensor of the k-th iteration, the sparse patch tensor of the k-th iteration is calculated; wherein, in the k-th iteration, the calculation expression of the sparse patch tensor is:

[0033]

[0034] Among them, λ is the weight coefficient, is the element-wise threshold shrinkage operator, sign is the sign function.

[0035] The Lagrangian multiplier of the kth iteration is calculated according to the low-rank patch tensor and the sparse patch tensor of the kth iteration; at the kth iteration, the calculation expression of the Lagrangian multiplier is:

[0036]

[0037] Determine whether the iteration stopping criteria are met:

[0038]

[0039] Among them, ε is a positive constant;

[0040] If the iteration stopping criterion is not met, update k=k+1 and update μ k+1 =μ k / ρ, repeat the iteration; where ρ is a constant;

[0041] If the stopping criteria of the iteration are met, the output is and stops the iteration process.

[0042] Furthermore, the step of reconstructing the sparse patch tensor to obtain a sparse time-frequency matrix includes:

[0043] The mean filter function is used to determine the value of each resolution unit, and its expression is:

[0044] v = mean(x)

[0045] in, is the vector of corresponding values ​​of p patch tensors corresponding to a resolution unit;

[0046] According to the value of each resolution unit, the sparse time-frequency matrix T is obtained.

[0047] Furthermore, the step of performing an inverse short-time Fourier transform on the phase information and the sparse time-frequency matrix to obtain time domain data after reverberation suppression includes:

[0048] The phase information and the sparse time-frequency matrix are subjected to an inverse short-time Fourier transform, and the inverse short-time Fourier transform uses a synthesis window s to obtain the time domain data.

[0049] Among them, the expression of the inverse short-time Fourier transform is:

[0050]

[0051] The synthesis window s satisfies the energy complementarity constraint:

[0052]

[0053] The technical solution provided by the embodiments of the present disclosure may have the following beneficial effects:

[0054] In the embodiments of the present disclosure, the reverberation suppression method based on the time-frequency patch tensor model is used, and the echo signal has a low-rank sparse characteristic in the time-frequency domain, so as to model the reverberation suppression problem as an optimization problem of low-rank sparse decomposition. The time-frequency patch tensor is decomposed into a low-rank patch tensor and a sparse patch using the alternating direction multi-multiplier method. The sparse patch tensor mainly includes the target echo component and a small part of the reverberation component; the low-rank patch tensor is mainly the reverberation component. Since the low-rank patch tensor contains most of the reverberation components, the signal-to-mixing ratio of the obtained sparse patch tensor is significantly improved, thereby achieving reverberation suppression. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The accompanying drawings herein are incorporated into the specification and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the specification are used to explain the principles of the present disclosure. Obviously, the accompanying drawings described below are only some embodiments of the present disclosure, and for ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without creative work.

[0056] Figure 1 A step diagram of a reverberation suppression method based on a time-frequency patch tensor model in an exemplary embodiment of the present disclosure is shown;

[0057] Figure 2 A schematic diagram showing an experimental scenario in an exemplary embodiment of the present disclosure;

[0058] Figure 3 shows beam domain echo data in an exemplary embodiment of the present disclosure;

[0059] Figure 4 A specific flow chart of a reverberation suppression method based on a time-frequency patch tensor model in an exemplary embodiment of the present disclosure is shown;

[0060] Figure 5 The time-frequency matrix obtained by short-time Fourier transform and modulo operation of beam-domain echo data in the exemplary embodiment of the present disclosure is shown;

[0061] Figure 6 The time-frequency matrix after reverberation suppression in the exemplary embodiment of the present disclosure is shown;

[0062] Figure 7 The signal after reverberation suppression by the reverberation suppression method based on the time-frequency patch tensor model in the exemplary embodiment of the present disclosure is shown;

[0063] Figure 8 In the exemplary embodiment of the present disclosure, Figure 3 The beam domain echo data in is the signal after matched filtering;

[0064] Fig. 9 In the exemplary embodiment of the present disclosure, Figure 5The signal after reverberation suppression and matched filtering;

[0065] Fig.10 The receiving operating characteristic curves of the active sonar detection system corresponding to different reverberation suppression methods in the exemplary embodiments of the present disclosure are shown. DETAILED DESCRIPTION

[0066] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that the disclosure will be more comprehensive and complete and to fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0067] In addition, the accompanying drawings are only schematic illustrations of the embodiments of the present disclosure and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated descriptions will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0068] This example embodiment provides a reverberation suppression method based on a time-frequency patch tensor model. Figure 1 As shown in , the reverberation suppression method based on the time-frequency patch tensor model may include: steps S101 to S105.

[0069] Step S101: performing short-time Fourier transform on the echo data collected once, and processing the echo data after the short-time Fourier transform by using a modulo operation to obtain a time-frequency matrix and phase information;

[0070] Step S102: constructing a time-frequency patch tensor based on the time-frequency matrix;

[0071] Step S103: Decomposing the time-frequency patch tensor using an alternating direction multi-multiplier method to obtain a sparse patch tensor;

[0072] Step S104: reconstructing the sparse patch tensor to obtain a sparse time-frequency matrix;

[0073] Step S105: performing an inverse short-time Fourier transform on the phase information and the sparse time-frequency matrix to obtain time domain data after reverberation suppression.

[0074] Through the above-mentioned reverberation suppression method based on the time-frequency patch tensor model, the reverberation suppression problem is modeled as an optimization problem of low-rank sparse decomposition by utilizing the low-rank sparse characteristics of the echo signal in the time-frequency domain. The time-frequency patch tensor is decomposed into a low-rank patch tensor and a sparse patch using the alternating direction multi-multiplier method. The sparse patch tensor mainly includes the target echo component and a small part of the reverberation component; the low-rank patch tensor is mainly the reverberation component. Since the low-rank patch tensor contains most of the reverberation components, the signal-to-mixing ratio of the obtained sparse patch tensor is significantly improved, thereby achieving reverberation suppression.

[0075] Next, we will refer to Figures 1 to 10 Each step of the above-mentioned reverberation suppression method based on the time-frequency patch tensor model in this example implementation is described in more detail.

[0076] In step S101, a short-time Fourier transform (STFT) is performed on the single-collected echo data m, and then a time-frequency matrix M is constructed using a modulo operation, and the phase information PI is saved;

[0077] For the discretized echo signal m, its time spectrum is:

[0078]

[0079] Wherein, w is the window function, the length is M, N is the number of DFT points, R is the number of overlapping points, a is the frequency index, and b is the time index. In the method proposed in this application, the Hanning window is used, which is defined as follows:

[0080]

[0081] For the echo signal m of length L, STFT is performed and the time-frequency matrix is ​​obtained by modulo operation and Phase Where A is the number of frequency resolution units, A = N / 2; B is the number of time resolution units, Indicates rounding down.

[0082] In step S102, the time-frequency patch tensor is constructed using the time-frequency matrix M Construct a time-frequency patch tensor from the time-frequency matrix M The specific steps are as follows:

[0083] (1) Using a sliding window Get a series of local patch matrices, where size is the dimension of the sliding window, and the sliding window w starts from the upper left corner of the time-frequency matrix M with a sliding step size step v Slide horizontally until it can no longer move horizontally; then the window returns to the far left and moves in steps of step h Move vertically downward. Then, repeat the process until the entire time-frequency matrix is ​​traversed.

[0084] (2) Stack these patches into a three-dimensional cube, which is called the patch tensor in this application.

[0085] Where n1 is the number of sliding windows,

[0086]

[0087] In this way, the time-frequency matrix M can be converted into a time-frequency patch tensor

[0088] In step S103, the time-frequency patch tensor M is decomposed into a low-rank patch tensor using an alternating direction multi-multiplier method. and sparse patch tensors The specific decomposition process is as follows:

[0089] (1) Variable initialization;

[0090] Let tensor is a low-rank patch tensor, tensor is a sparse patch tensor, a tensor is a Lagrange multiplier; a tensor Tensor as well as The initial values ​​are and Let k be the number of iterations, and take k = 0; the initial value of the penalty factor is

[0091] (2) Calculate the k-th iteration low-rank patch tensor; at the k-th iteration, the calculation expression of the low-rank patch tensor is:

[0092]

[0093] in, is the threshold shrinkage operator for the matrix singular values, Where X = Udiag(σ)V T is the SVD decomposition of matrix X, fold i It is an operation to restore the matrix expanded according to mode-i to the original tensor.

[0094] (3) Calculate the k-th iteration sparse patch tensor; at the k-th iteration, the calculation expression of the sparse patch tensor is:

[0095]

[0096] Among them, λ is the weight coefficient, which controls the sparsity of the target patch tensor; is an element-by-element threshold shrinkage operator. For a given shrinkage threshold μ, sign is the sign function.

[0097] (4) Calculate the Lagrange multiplier of the kth iteration. The calculation expression of the Lagrange multiplier of the kth iteration is:

[0098]

[0099] (5) Determine whether the conditions for stopping iteration have been reached; determine whether the algorithm meets the iteration stopping criteria:

[0100]

[0101] where ε is a small positive constant that depends on the accuracy required by the algorithm.

[0102] If the iteration stopping criterion is not met, update k=k+1, update

[0103] μ k+1 =μ k / ρ(8)

[0104] Then repeat the above process (2)-(4) until the algorithm iteration stopping criterion is met.

[0105] If the iteration stopping criterion is met, the output is and stops the iteration process.

[0106] In step S104, from the sparse patch tensor Reconstruct the sparse time-frequency matrix T from the sparse patch tensor The specific steps to reconstruct the sparse time-frequency matrix T are as follows:

[0107] The mean filter function is used to determine the value of each resolution unit of the time-frequency matrix T. The value of each resolution unit in the time-frequency matrix T corresponds to the sparse patch tensor Therefore, a mean filter function is used to determine the value v of the resolution unit:

[0108] v = mean(x)(9)

[0109] in Is a vector of the corresponding values ​​of p patch tensors corresponding to a resolution unit. By using the mean filter function to determine each resolution unit value of the time-frequency matrix T, a sparse time-frequency matrix T can be obtained.

[0110] In step S105, using the phase information After inverse STFT with the sparse time-frequency matrix T, the time domain data after reverberation suppression is obtained This is the result after reverberation suppression.

[0111] The inversion of STFT is done by the synthetic operator STFT -1 Implementation, use the synthetic window s to get the signal

[0112]

[0113] The synthesis window s satisfies the energy complementarity constraint:

[0114]

[0115] In a specific embodiment, the present invention is further described below in conjunction with the accompanying drawings and embodiments. The present invention includes but is not limited to the following embodiments.

[0116] 1. Experimental configuration.

[0117] Experimental scenarios, such as Figure 2 As shown, the active sonar is fixed at a depth of 40m, and the target is a small ball. The active detection system transmits at an interval of 1s, and the transmitted signal is an LFM signal with a normalized frequency range of 0.56-0.65. The receiving array uses a uniform linear array of 64 elements, with a unit spacing of 3cm, a sampling rate of 100kHz, and an effective horizontal detection field of view of 180 degrees. The experiment was carried out in a lake environment, and the main interference was reverberation.

[0118] 2. Experimental data.

[0119] The active detection system repeatedly transmits linear frequency modulation signals, and the echo data collected by the linear array is beamformed to obtain beam domain data such as Figure 3 shown.

[0120] like Figure 4 As shown, the specific implementation process of the reverberation suppression method based on the time-frequency patch tensor model proposed in the present invention is as follows:

[0121] Perform short-time Fourier transform (STFT) on the single-acquired beam domain echo data m, then use the modulo operation to construct the time-frequency matrix M and save the phase information PI;

[0122] For the discretized echo signal m, its time spectrum is:

[0123]

[0124] Where w is the Hanning window with a length of 64, the number of overlap points is 32, the number of DFT points is 128, a is the frequency index, and b is the time index.

[0125] For the echo signal m with a length of 10000, STFT is performed and the time-frequency matrix is ​​obtained by modulo operation. like Figure 5 shown.

[0126] Using the time-frequency matrix Constructing a time-frequency patch tensor Construct a time-frequency patch tensor from the time-frequency matrix M The specific steps are as follows:

[0127] (1) Using a sliding window from top left to bottom right Slide horizontally with step length step v = 8 until it can no longer move horizontally; then the window returns to the far left and moves with a step size of step h =8 and move vertically downward. Then, repeat the process until the entire time-frequency matrix is ​​traversed. A series of local patch matrices are obtained by sliding with a certain sliding step size;

[0128] (2) Stack these patches into a three-dimensional cube, which is called the patch tensor in this application.

[0129] In this way, the time-frequency matrix M can be converted into a time-frequency patch tensor

[0130] The time-frequency patch tensor is transformed using the alternating direction multi-multiplier method. Decompose into low-rank patch tensors and sparse patch tensors The specific decomposition process is as follows:

[0131] (1) Variable initialization;

[0132] Let tensor is a low-rank patch tensor, tensor is a sparse patch tensor, a tensor is a Lagrange multiplier; a tensor Tensor as well as The initial values ​​are and Let k be the number of iterations, and take k = 0; the initial value of the penalty factor is

[0133] (2) Calculate the k-th iteration low-rank patch tensor; at the k-th iteration, the calculation expression of the low-rank patch tensor is:

[0134]

[0135] in is the threshold shrinkage operator for the matrix singular values, Where X = Udiag(σ)V T is the SVD decomposition of matrix X, fold i It is an operation to restore the matrix expanded according to mode-i to the original tensor.

[0136] (3) Calculate the k-th iteration sparse patch tensor; at the k-th iteration, the calculation expression of the sparse patch tensor is:

[0137]

[0138] Where λ is the weight coefficient, is an element-by-element threshold shrinkage operator. For a given shrinkage threshold μ, S μ (x)=sign(x)max(|x|-μ,0), where sign is the sign function.

[0139] (4) Calculate the Lagrange multiplier of the kth iteration. The calculation expression of the Lagrange multiplier of the kth iteration is:

[0140]

[0141] (5) Determine whether the conditions for stopping iteration have been reached; determine whether the algorithm meets the iteration stopping criteria:

[0142] and

[0143] If the iteration stopping criterion is not met, update k=k+1, update

[0144] μ k+1 =μ k / 1.5

[0145] Then repeat the above process (2)-(4) until the algorithm iteration stopping criterion is met.

[0146] If the iteration stopping criterion is met, the output is and stops the iteration process.

[0147] Reconstruct the sparse time-frequency matrix T from the sparse patch tensor T; reconstruct the sparse time-frequency matrix T from the sparse patch tensor The steps to reconstruct the sparse time-frequency matrix T are as follows:

[0148] The mean filter function is used to determine the value of each resolution unit of the time-frequency matrix T, and the value of each resolution unit in the time-frequency matrix T corresponds to several values ​​from the sparse patch tensor T. Therefore, the mean filter function is used to determine the value v of the resolution unit:

[0149] v = mean(x)

[0150] in Is a vector of p patch tensor values ​​corresponding to a resolution unit. Using the mean filter function to determine each resolution unit value of the time-frequency matrix T, the time-frequency matrix T can be obtained, such as Figure 6 shown.

[0151] Using phase information After inverse STFT with the sparse time-frequency matrix T, the time domain data after reverberation suppression is obtained This is the result after reverberation suppression.

[0152] The final result after reverberation suppression is as follows Figure 7 In order to compare the reverberation suppression effect, matched filtering is selected as the detector. Figure 3 The result of matched filtering of the beam domain data shown is as follows Figure 8 As shown, Figure 7 The result of the reverberation suppression signal after matched filtering is shown in Fig. 9 As shown. Figure 8 and Fig. 9 It can be seen that after being processed by the method proposed in the present invention, the amplitude of reverberation interference can be significantly suppressed.

[0153] In order to verify the innovativeness of the method proposed in this invention, simulation data were used to compare with the reverberation pre-whitening method and the principal component inversion method. The results are as follows: Fig.10 It can be seen that under different signal-to-mixture ratio conditions, the method proposed in the present invention achieves better reverberation suppression effect than the reverberation pre-whitening method and the principal component inversion method, and improves the detection performance of active sonar.

[0154] Through the above-mentioned reverberation suppression method based on the time-frequency patch tensor model, the reverberation suppression problem is modeled as an optimization problem of low-rank sparse decomposition by utilizing the low-rank sparse characteristics of the echo signal in the time-frequency domain. The time-frequency patch tensor is decomposed into a low-rank patch tensor and a sparse patch using the alternating direction multi-multiplier method. The sparse patch tensor mainly includes the target echo component and a small part of the reverberation component; the low-rank patch tensor is mainly the reverberation component. Since the low-rank patch tensor contains most of the reverberation components, the signal-to-mixing ratio of the obtained sparse patch tensor is significantly improved, thereby achieving reverberation suppression.

[0155] It should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like in the above description indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the embodiments of the present disclosure and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the embodiments of the present disclosure.

[0156] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of the present disclosure, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0157] In the embodiments of the present disclosure, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present disclosure can be understood according to specific circumstances.

[0158] In the embodiments of the present disclosure, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may include the first and second features being in direct contact, or may include the first and second features not being in direct contact but being in contact through another feature between them. Moreover, a first feature being "above", "above" and "above" a second feature includes the first feature being directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature includes the first feature being directly below and obliquely below the second feature, or simply indicates that the first feature is lower in level than the second feature.

[0159] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification.

[0160] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any modification, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the appended claims.

Claims

1. A reverberation suppression method based on a time-frequency patch tensor model, characterized in that: The method includes: Performing short-time Fourier transform on the echo data collected once, and processing the echo data after the short-time Fourier transform by using a modulo operation to obtain a time-frequency matrix and phase information; constructing a time-frequency patch tensor based on the time-frequency matrix; Decomposing the time-frequency patch tensor using an alternating direction multi-multiplier method to obtain a sparse patch tensor; Reconstructing the sparse patch tensor to obtain a sparse time-frequency matrix; An inverse short-time Fourier transform is performed on the phase information and the sparse time-frequency matrix to obtain time domain data after reverberation suppression.

2. The reverberation suppression method based on the time-frequency patch tensor model according to claim 1 is characterized in that: The step of performing short-time Fourier transform on the echo data collected once, and processing the echo data after the short-time Fourier transform by using a modulo operation to obtain a time-frequency matrix and phase information includes: Perform a short-time Fourier transform on the echo signal m of length L, and obtain the time-frequency matrix by using a modulo operation. and the phase information in, The expression for performing short-time Fourier transform on the discretized echo signal m is: The window function uses the Hanning window, which is defined as follows: Wherein, w[n] is the window function, the length of the window function is M, n is the position of the current window, N is the number of DFT points, R is the number of overlapping points, a is the frequency index, b is the time index, M(a,b)=M is ​​the time-frequency matrix, k is the frequency point of the short-time Fourier transform, and j is the imaginary unit; A is the number of frequency resolution units, A=N / 2, B is the number of time resolution units, Indicates rounding down.

3. The reverberation suppression method based on the time-frequency patch tensor model according to claim 2 is characterized in that: The step of constructing a time-frequency patch tensor based on the time-frequency matrix includes: Using a sliding window from top left to bottom right Slide with a preset sliding step size to obtain a local patch matrix; where size is the dimension of the sliding window; Stack the patch matrices into the time-frequency patch tensor 4. The reverberation suppression method based on the time-frequency patch tensor model according to claim 3 is characterized in that: The number n1 of the sliding windows is: In the formula, step h is the sliding step length in the vertical direction, step v is the sliding step length in the horizontal direction.

5. The reverberation suppression method based on the time-frequency patch tensor model according to claim 4 is characterized in that: The step of decomposing the time-frequency patch tensor by using an alternating direction multi-multiplier method to obtain a sparse patch tensor comprises: Let the low-rank patch tensor be The sparse patch tensor is The Lagrange multiplier is The initial values ​​of the low-rank patch tensor, the sparse patch tensor and the Lagrange multiplier are set to and i = 1, 2, 3; let k be the number of iterations, and take k = 0; the initial value of the penalty factor is Calculate the low-rank patch tensor of the k-th iteration; wherein, at the k-th iteration, the calculation expression of the low-rank patch tensor is: In the formula, fold i To restore the matrix expanded in mode-i to the original tensor, is the threshold shrinkage operator for the matrix singular values, X=Udiag(σ)V T is the SVD decomposition of matrix X, According to the low-rank patch tensor of the k-th iteration, the sparse patch tensor of the k-th iteration is calculated; wherein, in the k-th iteration, the calculation expression of the sparse patch tensor is: Among them, λ is the weight coefficient, is the element-wise threshold shrinkage operator, sign is the sign function. The Lagrangian multiplier of the kth iteration is calculated according to the low-rank patch tensor and the sparse patch tensor of the kth iteration; at the kth iteration, the calculation expression of the Lagrangian multiplier is: Determine whether the iteration stopping criteria are met: Among them, ε is a positive constant; If the iteration stopping criterion is not met, update k=k+1 and update μ k+1 =μ k / ρ, repeat the iteration; where ρ is a constant; If the stopping criteria of the iteration are met, the output is and stops the iteration process.

6. The reverberation suppression method based on the time-frequency patch tensor model according to claim 5, characterized in that: The step of reconstructing the sparse patch tensor to obtain a sparse time-frequency matrix includes: The mean filter function is used to determine the value of each resolution unit, and its expression is: v = mean(x) in, is the vector of corresponding values ​​of p patch tensors corresponding to a resolution unit; According to the value of each resolution unit, the sparse time-frequency matrix T is obtained.

7. The reverberation suppression method based on the time-frequency patch tensor model according to claim 6 is characterized in that: The step of performing an inverse short-time Fourier transform on the phase information and the sparse time-frequency matrix to obtain time domain data after reverberation suppression includes: The phase information and the sparse time-frequency matrix are subjected to an inverse short-time Fourier transform, and the inverse short-time Fourier transform uses a synthesis window s to obtain the time domain data. Among them, the expression of the inverse short-time Fourier transform is: The synthesis window s satisfies the energy complementarity constraint:

Images