A multi-beam sonar compression imaging method and system based on depth unfolding

Abstract

The present application relates to the technical field of sonar imaging, and particularly relates to a multi-beam sonar compressed imaging method and system based on depth unfolding, which comprises the following steps: a multi-beam sonar system is used to emit sound waves to the seabed, and a receiving device is used to receive reflected echo signals to obtain multi-channel digital time sequence echo signals; an observation matrix is constructed to compress and sample the multi-channel digital time sequence echo signals to obtain a low-dimensional measurement matrix; the low-dimensional measurement matrix is input into a trained depth unfolding beamforming network to output a reconstructed beamforming spectrum, and a sonar sounding image is generated according to the beamforming spectrum. The present application can greatly improve the imaging speed and support real-time applications such as underwater obstacle avoidance and online monitoring.

Description

Technical Field

[0001] This invention relates to the technical field of sonar imaging, and in particular to a multi-beam sonar compression imaging method and system based on depth unfolding. Background Technology

[0002] Multibeam sonar is widely used in critical fields such as seabed topography mapping, underwater target detection, shipwreck search, marine resource exploration, and underwater security. In these applications, achieving high-resolution imaging typically requires the acquisition of a large amount of echo data, resulting in massive data volumes and heavy storage and transmission burdens. This is especially true on resource-constrained platforms such as unmanned underwater vehicles or long-term deployed sonar nodes, where efficient data compression and high-quality reconstruction capabilities are urgently needed. Therefore, compressed imaging technology has become crucial for improving system efficiency and practicality.

[0003] However, traditional multibeam sonar imaging methods based on compressed sensing heavily rely on pre-defined sparse dictionaries, making it difficult to accurately characterize the non-ideal sparse echo signal characteristics in complex underwater environments. Under conditions of low signal-to-noise ratio, strong reverberation, or multipath interference, these methods are prone to significant reconstruction errors, leading to blurred images, target location misalignment, or even missed detections, severely impacting subsequent target detection and topographic mapping. More seriously, due to the lack of sufficient consideration of the mismatch between the actual array manifold and the theoretical model, traditional methods have insufficient sidelobe suppression capabilities during beamforming, resulting in false targets and reduced angular resolution, potentially causing misjudgments or significant economic losses in target detection or detailed mapping.

[0004] In recent years, although some studies have attempted to introduce deep learning to improve reconstruction performance, most methods treat neural networks as "black box" post-processing modules, failing to jointly optimize learnable parameters with the sonar physical imaging model. This fragmented design not only weakens the model's interpretability and generalization ability but also fails to effectively compensate for systematic errors caused by array manifold mismatch, limiting its practicality in real and complex underwater environments. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a multi-beam sonar compressed imaging method and system based on depth unfolding, which solves the technical problems of dependence on preset sparse dictionary and insufficient robustness of multi-beam sonar imaging methods based on compressed sensing.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the main technical solutions adopted by the present invention include:

[0009] In a first aspect, embodiments of the present invention provide a multi-beam sonar compression imaging method based on depth unfolding, comprising the following steps:

[0010] S1, using a multibeam sonar system to transmit sound waves to the seabed and receiving the reflected echo signals through a receiving device to obtain multi-channel digital time-series echo signals;

[0011] S2, construct the observation matrix, compress and sample the multi-channel digital time-series echo signal to obtain a low-dimensional measurement matrix;

[0012] S4 inputs the low-dimensional measurement matrix into the trained depth unwrap beamforming network, outputs the reconstructed beamforming spectrum, and generates a sonar depth sounding image based on the beamforming spectrum.

[0013] The multi-beam sonar compression imaging method based on depth unfolding proposed in this invention can complete reconstruction with a single feedforward after training, greatly improving imaging speed and supporting real-time applications such as underwater obstacle avoidance and online monitoring.

[0014] Optionally, the multibeam sonar compression imaging method may further include:

[0015] S3, Offline training phase of the deep unfolded beamforming network. The offline training phase is used to construct and train the deep unfolded beamforming network. Optionally, the deep unfolded beamforming network includes an initial reconstruction module, K iterative reconstruction modules, and a differentiable beamforming module, where K ≥ 1.

[0016] Optionally, the deep-unfolded beamforming network is trained through the following steps:

[0017] S31, Constructing a training dataset: Acquire multiple sets of original multibeam sonar echo signals as raw signals, compress and sample each set of raw signals to obtain a low-dimensional measurement matrix; perform beamforming processing on the raw signals to generate a target beamforming spectrum, pair the low-dimensional measurement matrix corresponding to each set of raw signals with the target beamforming spectrum to form training samples, multiple sets of training samples form a dataset, and divide the dataset into a training set and a test set;

[0018] S32, Initial Reconstruction Module: Converts the low-dimensional measurement matrix using the transpose Φ of the observation matrix. T A preliminary reconstruction is performed to obtain the initial signal X(0);

[0019] S33, Iterative Reconstruction Module: For k=1, 2, ..., K iterations, the data consistency sub-step, complex gain correction sub-step, and soft threshold denoising sub-step are executed sequentially, and the reconstructed signal X(K) of the Kth iteration is output;

[0020] S34, Differentiable beamforming module: Input the reconstructed signal X(K) of the Kth iteration into the differentiable beamforming module and output the reconstructed beamforming spectrum;

[0021] S35, Network Training: The deep unfolded beamforming network is trained end-to-end using training set samples. The reconstruction error between the reconstructed beamforming spectrum output by the deep unfolded beamforming network and the target beamforming spectrum is used as the loss function. The minimization of the loss function is used as the optimization objective. Other learnable parameters in the deep unfolded beamforming network are jointly optimized for training until the model converges.

[0022] Optionally, the data consistency sub-step includes: inputting the initial signal X(0) or the reconstructed signal X(k-1) output by the (k-1)th iteration into the data consistency sub-module; calculating the forward projection of the current reconstructed signal in the measurement domain using the observation matrix, and comparing it with the original low-dimensional echo measurement matrix to obtain the measurement domain residual; backprojecting the residual back to the signal domain through the conjugate transpose of the observation matrix, multiplying it by the learnable step size parameter, and then weighting and fusing it into the current reconstructed signal to output the consistency-calibrated signal Z(k).

[0023] Optionally, the complex gain correction sub-step specifically involves: inputting the uniformly calibrated signal Z(k) into the complex gain correction sub-module; constructing an ideal steering vector set as an array manifold prior based on the geometry and operating frequency of the sonar receiving array; introducing a set of learnable complex gain correction factors corresponding one-to-one with the receiving channels, each learnable complex gain correction factor containing trainable amplitude and phase components; using the complex gain correction factors to perform amplitude and phase joint correction on each channel of the uniformly calibrated signal Z(k), so that the phase structure between channels of the corrected signal approximates the ideal array manifold, and outputting the complex gain corrected signal P(k). This invention embeds a learnable complex gain correction module into the network, and through end-to-end training, adaptively compensates for channel errors, thereby improving angular resolution, suppressing sidelobes, and achieving accurate discrimination of nearby targets and subtle terrain features.

[0024] Optionally, the soft-threshold denoising sub-step is as follows: input the complex gain-corrected signal P(k) into the soft-threshold denoising sub-module; construct a guide vector based on the array geometry and a preset azimuth angle set, and perform matrix operations on the guide vector and signal P(k) to obtain the angle domain beamforming spectrum s(k); introduce a learnable threshold parameter, perform element-wise soft-thresholding on the angle domain beamforming spectrum s(k) to obtain the denoised angle domain beamforming spectrum S(k); transform the denoised angle domain beamforming spectrum S(k) back to the array element domain by performing matrix operations with the transpose of the guide vector to obtain the reconstructed signal X(k) for the kth iteration.

[0025] Optionally, the differentiable beamforming module is specifically used for: calculating the regularized covariance matrix based on the reconstructed signal X(K) output by the network; calculating the MVDR weight vector based on the array steering vector and the regularized covariance matrix; and performing beamforming operations on the MVDR weights and the reconstructed signal X(K) to obtain the beamforming spectrum S(θ) as the final output of the network.

[0026] Optionally, the loss function is the mean square error (MSE) between the reconstructed beamforming spectrum and the ground truth beamforming spectrum; the gradient of the loss with respect to all learnable parameters in the network is calculated using the backpropagation algorithm, and the learnable parameters include the data consistency step size parameter, the complex gain correction factor, and the soft threshold parameter in each iteration stage; the Adam optimizer is used to update the parameters until the loss function converges.

[0027] Optionally, generating a sonar depth image includes the following steps:

[0028] The low-dimensional measurement matrix to be processed is input into a trained depth unwrap beamforming network, which outputs a reconstructed beamforming spectrum. The time sampling points in the beamforming spectrum are converted into slant ranges using the two-way propagation distance formula, and the data is remapped to an angular slant range image grid. The spectral intensity at each grid point is mapped to grayscale or pseudo-color to generate a high-resolution sonar depth sounding image in the angular range domain. The two-way propagation distance formula is: R = (C × T) / (2 × Fs), where R is the slant range, C is the speed of sound in the medium, T is the time sampling point, and Fs is the time sampling frequency.

[0029] In a second aspect, embodiments of the present invention provide a multibeam sonar compression imaging system based on depth unfolding, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the above methods.

[0030] (III) Beneficial Effects

[0031] The beneficial effects of this invention are:

[0032] The multibeam sonar compressed imaging method and system based on depth unfolding of the present invention effectively overcomes the problems of dependence on preset sparse dictionaries and sensitivity to array manifold mismatch in traditional compressed sensing methods. It achieves high-precision and high-real-time imaging while significantly reducing the amount of data, and is suitable for seabed topographic mapping and weak target detection in resource-constrained scenarios such as autonomous underwater platforms.

[0033] Compared to multibeam sonar imaging methods based on compressed sensing, the depth-unfolded multibeam sonar compressed imaging method and system of the present invention, in the preferred embodiment, performs better in the following two aspects:

[0034] Enhanced Real-Time Imaging Performance: Overcoming the limitations of traditional iterative algorithms which involve high computational demands and struggle to meet the requirements of real-time mapping and dynamic target tracking, this invention employs a fixed-layer forward computation. Reconstruction can be completed with a single feedforward operation after training, significantly improving imaging speed and supporting real-time applications such as underwater obstacle avoidance and online monitoring.

[0035] Superior beamforming performance: To address the issues of manifold mismatch and excessive sidelobes caused by amplitude and phase errors in actual arrays, this invention embeds a learnable complex gain correction module into the network. Through end-to-end training, it adaptively compensates for channel errors, thereby improving angular resolution, suppressing sidelobes, and achieving accurate differentiation of nearby targets and subtle terrain features. Attached Figure Description

[0036] Figure 1 This is a flowchart of a multibeam sonar compression imaging method based on depth unfolding according to an embodiment of the present invention;

[0037] Figure 2 This is a diagram of the deep unfolded beamforming network structure according to an embodiment of the present invention. Detailed Implementation

[0038] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] The multi-beam sonar compressed imaging method and system based on depth unfolding proposed in this invention effectively overcomes the problems of traditional compressed sensing methods, such as dependence on a preset sparse dictionary and sensitivity to array manifold mismatch. Reconstruction can be completed with a single feedforward after training, significantly improving imaging speed and supporting real-time applications such as underwater obstacle avoidance and online monitoring. It achieves high-precision, high-real-time imaging while significantly reducing data volume, making it suitable for seabed topographic mapping and weak target detection in resource-constrained scenarios such as autonomous underwater platforms.

[0040] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0041] Example 1

[0042] Figure 1 This is a flowchart of a depth-unfolded multibeam sonar compression imaging method according to an embodiment of the present invention. The depth-unfolded multibeam sonar compression imaging method of this embodiment includes the following steps:

[0043] S1 uses a multibeam sonar system to transmit sound waves to the seabed and receives the reflected echo signals through a receiving device to obtain multi-channel digital time-series echo signals.

[0044] S2. Construct the observation matrix and compress and sample the multi-channel digital time-series echo signal to obtain a low-dimensional measurement matrix.

[0045] S4 inputs the low-dimensional measurement matrix into the trained depth-unfolded beamforming network, outputs the reconstructed beamforming spectrum, and generates a sonar depth sounding image based on the beamforming spectrum. In implementation, the depth-unfolded beamforming network may include an initial reconstruction module, K iterative reconstruction modules, and a differentiable beamforming module, where K ≥ 1. That is, a fixed number of layers are used for forward computation.

[0046] The multi-beam sonar compression imaging method based on depth unfolding proposed in this invention can complete reconstruction with a single feedforward after training, greatly improving imaging speed and supporting real-time applications such as underwater obstacle avoidance and online monitoring.

[0047] The above steps constitute the real-time imaging stage of the depth-unfolded multi-beam sonar compressed imaging method according to this embodiment of the invention. The real-time imaging stage inputs the compressed, sampled low-dimensional measurement matrix into the depth-unfolded beamforming network and directly outputs the reconstructed high-resolution beamforming spectrum to generate a sonar depth sounding image. In practice, the depth-unfolded multi-beam sonar compressed imaging method provided in this embodiment may further include:

[0048] S3, Offline Training Phase of the Deep Expanded Beamforming Network. The offline training phase is used to construct and train the deep expanded beamforming network. The offline training phase includes: constructing and training an end-to-end deep expanded beamforming network, which consists of an initial reconstruction module, K iterative reconstruction modules, and a differentiable beamforming module. Each iterative reconstruction module sequentially includes a data consistency submodule, a complex gain correction submodule, and a soft thresholding denoising submodule; using a training set consisting of the compressed measurement matrix and its corresponding beamforming spectrum of the original signal, the network is trained end-to-end, where the beamforming spectrum of the original signal is obtained by beamforming processing the original multi-beam sonar signal; by minimizing the reconstruction error between the network's output beamforming spectrum and the original signal beamforming spectrum, the learnable parameters in the network are optimized to obtain the trained deep expanded beamforming network. The specific process is as follows:

[0049] S31, Construct a training dataset: Obtain multiple sets of original multibeam sonar echo signals as original signals, compress and sample each set of original signals to obtain a low-dimensional measurement matrix; perform beamforming processing on the original signals to generate a target beamforming spectrum, pair the low-dimensional measurement matrix corresponding to each set of original signals with the target beamforming spectrum to form training samples, multiple sets of training samples form a dataset, and divide the dataset into a training set and a test set.

[0050] Then, a deep unfolded beamforming network is constructed, which consists of an initial reconstruction module, K iterative reconstruction modules, and a differentiable beamforming module. The network structure is as follows:

[0051] S32, Initial Reconstruction Module: Performs preliminary reconstruction on the low-dimensional measurement matrix Y obtained by compressed sampling, performs matrix operation on it with the transpose of the observation matrix Φ to obtain the initial estimated signal X1, and uses this signal as the input of the iterative reconstruction module.

[0052] S33, Iterative Reconstruction Module: Set the total number of iterations to K (K ≥ 1), see [link to documentation]. Figure 2 Each iterative refactoring module is linearly connected to the following sub-modules:

[0053] Data Consistency Submodule: In the k-th iteration (k = 1, 2, …, K), the initial estimated signal X1 (k=1) or the output signal X(k-1) of the previous layer is first fed into the data consistency module as the input of this layer. The forward projection of the current reconstructed signal in the measurement domain is calculated using the observation matrix and compared with the original low-dimensional echo measurement matrix to obtain its residual Y in the measurement domain. The residual is then corrected using the transpose of the observation matrix Φ, that is, the residual is back-projected back into the signal domain through the conjugate transpose of the observation matrix, multiplied by the learnable step size parameter, and then weighted and fused into the current reconstructed signal to output the data consistency calibrated signal Z(k). This module forces the network output to be consistent with the actual observation data by projecting the signal residual back into the signal domain, thereby improving the physical reliability of the reconstruction.

[0054] The data consistency correction formula is:

[0055]

[0056] in, Let be the learnable step size parameter for the k-th iteration;

[0057] This is the transpose of the observation matrix Φ;

[0058] Y is the low-dimensional measurement matrix of the original signal.

[0059] Complex gain correction submodule: The data consistency enhancement signal Z(k) is input into the complex gain correction module. Based on the geometry of the multi-beam sonar receiver array, the theoretical phase difference relationship between adjacent channels is pre-constructed, and a set of learnable complex gain correction factors is initialized. Each learnable complex gain correction factor contains trainable amplitude and phase components. Complex gain correction factors are used to enhance signal consistency. Each channel performs combined amplitude and phase complex gain correction: This makes the inter-channel phase structure of the calibrated signal approximate the ideal array manifold, thus obtaining the calibrated signal. The output is the calibrated signal P(k). A learnable complex gain factor is introduced. It can adaptively compensate for amplitude and phase errors in the m-th channel caused by installation deviations or environmental disturbances, thereby correcting array manifold mismatch and avoiding beam pointing offset.

[0060] in This represents the amplitude and phase correction parameters of the m-th channel in the k-th iteration.

[0061] Soft threshold denoising submodule: The calibration signal P(k) after complex gain correction is input into the soft threshold denoising module. Since this module needs to perform denoising in the sparse domain, and in this example the sparse domain is the angle domain, while the domain where P(k) is located is the array element domain. First, based on the array geometry and the preset azimuth angle set... Construct the corresponding guide vector , guide vector The angle-domain beamforming spectrum is obtained by performing matrix operations on the calibration signal P(k) after complex gain correction. Subsequently, a learnable threshold parameter is introduced, and an element-wise soft thresholding shrinkage operation is performed on s(k) in the angle domain to obtain S(k). Finally, the denoised angle domain beamforming spectrum S(k) is transformed back into the element domain by performing a matrix operation with the steering vector. The sparsely denoised signal X(k) is obtained, and the k-th iteration processing is completed, thus obtaining the reconstructed signal X(k) for the k-th iteration. This module effectively suppresses noise and preserves the energy of the true target direction by performing soft thresholding denoising in the sparse domain.

[0062] The method for constructing the corresponding steering vector based on the array geometry and a preset azimuth set is as follows:

[0063] In this example, an M-element uniform linear array is selected, with an element spacing of d and a signal wavelength of λ. The formula for calculating the steering vector is:

[0064]

[0065] The formula for element-wise soft threshold shrinkage is:

[0066]

[0067] Wherein, s(k) is the beamforming spectrum of the calibration signal P(k) after complex gain correction; It is a soft threshold learnable parameter.

[0068] Use the reconstructed signal output from the k-th iteration as the input to the (k+1)-th layer, and repeat steps S32 to S33 until the K iterations are completed; output the reconstructed signal X(K) from the K-th iteration.

[0069] S34, Differentiable Beamforming Module: The reconstructed signal X(K) output from the Kth iteration is input into the beamforming module, and the output beamforming spectrum S(K) is used as the final output of the network. Optionally, this beamforming module uses MVDR beamforming (but is not limited to the beamforming method), and calculates its regularized covariance matrix based on the network output reconstructed signal X(K). εI is used to ensure that R is positive definite and invertible, and I is an identity matrix with the same dimension as R. The MVDR weight vector is calculated. The MVDR weights are then used to perform beamforming on the reconstructed signal to obtain the beamforming spectrum of the signal. S(θ) is used as the final output of the network.

[0070] The formula for calculating the MVDR weight vector is:

[0071]

[0072] Where R is the covariance of the reconstructed signal X(K);

[0073] This is the array guide vector.

[0074] S35 utilizes the constructed deep unfolded beamforming network for end-to-end training, as follows:

[0075] Multiple sets of training samples were acquired, each set consisting of a low-dimensional measurement matrix obtained by compressed sampling of the original multibeam sonar echo signal. And the beamforming spectrum generated by applying a high-resolution beamforming algorithm to the original signal. As a truth label.

[0076] Measurement matrix Input a depth unrolling network, and output the reconstructed beamforming spectrum after feedforward computation. Based on the reconstructed spectrum With truth spectrum Construct a loss function; in this example, the loss function uses the reconstructed spectrum. With truth spectrum The mean squared error (MSE) between the two is calculated. The gradient of the loss function with respect to all learnable parameters in the network is calculated using the backpropagation algorithm, and the parameters are updated using the Adam optimizer. The above feedforward-backpropagation process is repeated until the loss function converges, resulting in a trained deep unfolded beamforming network.

[0077] During the real-time imaging phase:

[0078] During S1 implementation, an underwater multibeam bathymetry sonar system actively and continuously transmits linear frequency modulated pulse signals. These sound waves propagate in the water and are reflected off the seabed, then synchronously received by a uniform linear array of sonar transducers. The array contains M equally spaced receiving elements, with an element spacing of d. Each receiving element is equipped with an independent analog receiving link, forming M parallel receiving channels. The analog echo signals from each channel are amplified with low noise and bandpass filtered before being synchronously sampled and digitized by a high-speed analog-to-digital converter at the same sampling rate. T time sampling points are acquired per frame, resulting in an M-channel, T-sample-point digital time-series echo signal matrix X. This signal fully preserves the time amplitude information of the echoes from each channel, serving as the raw data basis for subsequent compressed sampling and sonar bathymetry images.

[0079] During S2 implementation, the acquired multi-channel digital time-series echo signals are input into the compressed sampling module to construct a random observation matrix of dimension N×M. Where M is the number of receiving channels and N is the preset compressed channel dimension, satisfying N < M. According to compressed sensing theory, this random matrix satisfies the restricted isometry property with high probability, making it suitable for effective compressed sampling of multi-channel sonar signals with joint sparsity characteristics. In this example, the elements of the observation matrix Φ are independently and identically distributed, following a Gaussian distribution with a mean of 0 and a variance of 1 / N. Matrix operations are performed between the acquired M-channel digital time-series echo signals X and the observation matrix Φ to obtain the compressed measurement matrix. The low-dimensional measurement matrix Y is used as the input to the subsequent depth-unfolding beamforming network.

[0080] During S4 implementation, the low-dimensional measurement matrix Y to be processed is input into the trained depth-unfolded beamforming network to obtain a high-resolution beamforming spectrum. The depth-unfolded beamforming network has been trained end-to-end offline, and its parameters are optimized based on multiple sets of labeled compressed measurements and corresponding high-resolution beamforming spectrum samples. For the beamforming spectrum S(θ), the time sampling point T is first converted into the corresponding slant range R using the two-way propagation distance formula, and the (θ, T) plane of the beamforming spectrum S(θ) is mapped to an image grid (θ, R). The converted slant range R is combined with the corresponding beam pointing angle θ to form a polar coordinate sonar depth image, where each pixel represents the echo intensity at a certain direction and slant range, and the spectral intensity value of each pixel ||S(θ, T)|| is... 2 The image is mapped to grayscale or pseudocolor, ultimately generating a high-resolution sonar depth image in the angle-range domain.

[0081] The formula for two-way propagation distance is:

[0082]

[0083] Where T is the time sampling point; Fs is the time sampling frequency; and C is the speed of sound in the medium.

[0084] Secondly, embodiments of the present invention also provide a multi-beam sonar compression imaging system based on depth unfolding, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the above embodiments.

[0085] In summary, the depth-unfold-based multibeam sonar compressed imaging method and system of this invention, through compressed sampling and depth-unfold reconstruction architecture, achieves high-resolution underwater imaging while significantly reducing data acquisition volume and transmission bandwidth. This feature is particularly suitable for scenarios such as autonomous underwater vehicles performing long-term seabed topographic mapping tasks, effectively alleviating the problem of operation interruption caused by limited storage capacity and communication bandwidth, and supporting large-scale, high-efficiency continuous depth sounding operations; it can be widely applied to practical scenarios with high real-time, high-precision, and low-resource-consumption requirements for underwater environmental perception.

[0086] This invention introduces a complex gain correction submodule and angle-domain sparse modeling based on steering vectors, enabling adaptive compensation for inter-channel amplitude and phase errors and suppression of reverberation and sidelobe interference. In weak target detection scenarios such as underwater search and rescue, shipwreck detection, or port security, this mechanism significantly improves the detectability of low signal-to-noise ratio targets, avoiding missed detections or false alarms caused by array manifold mismatch or noise contamination. By jointly optimizing the learnable module and the physical model of sonar beamforming, the invention adaptively compensates for array manifold mismatch, overcoming the dependence on a preset sparse dictionary and insufficient robustness of traditional compressed sensing, effectively suppressing beam sidelobes and improving compressed imaging accuracy.

[0087] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0088] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0089] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first and second features are in direct contact, or that they are in indirect contact through an intermediate medium. Furthermore, "above," "over," or "on top" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," or "beneath" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0090] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0091] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A multi-beam sonar compressed imaging method based on depth unfolding, characterized in that, Includes the following steps: S1, using a multibeam sonar system to transmit sound waves to the seabed and receiving the reflected echo signals through a receiving device to obtain multi-channel digital time-series echo signals; S2, construct the observation matrix, compress and sample the multi-channel digital time-series echo signal to obtain a low-dimensional measurement matrix; S4. Input the low-dimensional measurement matrix into the trained depth unwrap beamforming network, output the reconstructed beamforming spectrum, and generate a sonar depth sounding image based on the beamforming spectrum.

2. The multibeam sonar compression imaging method based on depth unfolding according to claim 1, characterized in that, The deep unfolded beamforming network includes an initial reconstruction module, K iterative reconstruction modules, and a differentiable beamforming module, where K ≥ 1.

3. The multi-beam sonar compression imaging method based on depth unfolding according to claim 1 or 2, characterized in that, The depth-unfolding beamforming network is trained through the following steps: S31, Construct a training dataset: Obtain multiple sets of original multibeam sonar echo signals as original signals, compress and sample each set of original signals to obtain a low-dimensional measurement matrix; perform beamforming processing on the original signals to generate a target beamforming spectrum, pair the low-dimensional measurement matrix corresponding to each set of original signals with the target beamforming spectrum to form training samples, multiple sets of training samples form a dataset, and divide the dataset into a training set and a test set; S32, Initial Reconstruction Module: Converts the low-dimensional measurement matrix using the transpose Φ of the observation matrix. T A preliminary reconstruction is performed to obtain the initial signal X(0); S33, Iterative Reconstruction Module: For k=1, 2, ..., K iterations, the data consistency sub-step, complex gain correction sub-step, and soft threshold denoising sub-step are executed sequentially, and the reconstructed signal X(K) of the Kth iteration is output; S34, Differentiable beamforming module: Input the reconstructed signal X(K) of the Kth iteration into the differentiable beamforming module and output the reconstructed beamforming spectrum; S35, Network Training: The deep unfolded beamforming network is trained end-to-end using training set samples. The reconstruction error between the reconstructed beamforming spectrum output by the deep unfolded beamforming network and the target beamforming spectrum is used as the loss function. The minimization of the loss function is used as the optimization objective. Other learnable parameters in the deep unfolded beamforming network are jointly optimized for training until the model converges.

4. The multibeam sonar compression imaging method based on depth unfolding according to claim 3, characterized in that, The data consistency sub-step includes: inputting the initial signal X(0) or the reconstructed signal X(k-1) output by the (k-1)th iteration into the data consistency sub-module; calculating the forward projection of the current reconstructed signal in the measurement domain using the observation matrix, and comparing it with the original low-dimensional echo measurement matrix to obtain the measurement domain residual; backprojecting the residual back to the signal domain through the conjugate transpose of the observation matrix, multiplying it by the learnable step size parameter, and then weighting and fusing it into the current reconstructed signal to output the consistency-calibrated signal Z(k).

5. The multibeam sonar compression imaging method based on depth unfolding according to claim 3, characterized in that, The specific steps of the complex gain correction sub-step are as follows: inputting the uniformly calibrated signal Z(k) into the complex gain correction sub-module; constructing an ideal steering vector set as an array manifold prior based on the geometry and operating frequency of the sonar receiving array; introducing a set of learnable complex gain correction factors corresponding one-to-one with the receiving channels, each learnable complex gain correction factor containing trainable amplitude and phase components; using the complex gain correction factors to perform amplitude and phase joint correction on each channel of the uniformly calibrated signal Z(k), so that the phase structure between channels of the corrected signal approximates the ideal array manifold, and outputting the complex gain corrected signal P(k).

6. The multibeam sonar compression imaging method based on depth unfolding according to claim 3, characterized in that, The soft thresholding denoising sub-step is specifically as follows: inputting the complex gain-corrected signal P(k) into the soft thresholding denoising sub-module; constructing a guide vector based on the array geometry and a preset azimuth angle set, and performing matrix operations on the guide vector and signal P(k) to obtain the angle domain beamforming spectrum s(k); introducing a learnable threshold parameter, performing an element-wise soft thresholding operation on the angle domain beamforming spectrum s(k) to obtain the denoised angle domain beamforming spectrum S(k); and performing matrix operations on the denoised angle domain beamforming spectrum S(k) with the transpose of the guide vector to return it to the array element domain, thereby obtaining the reconstructed signal X(k) for the kth iteration.

7. The multibeam sonar compression imaging method based on depth unfolding according to claim 3, characterized in that, The differentiable beamforming module is specifically used for: calculating the regularized covariance matrix based on the reconstructed signal X(K) output by the network; calculating the MVDR weight vector based on the array steering vector and the regularized covariance matrix; and performing beamforming operations on the MVDR weights and the reconstructed signal X(K) to obtain the beamforming spectrum S(θ) as the final output of the network.

8. The multibeam sonar compression imaging method based on depth unfolding according to claim 3, characterized in that, The loss function is the mean square error (MSE) between the reconstructed beamforming spectrum and the ground truth beamforming spectrum; the backpropagation algorithm is used to calculate the gradient of the loss with respect to all learnable parameters in the network, including the data consistency step size parameter, complex gain correction factor, and soft threshold parameter in each iteration stage; the Adam optimizer is used to update the parameters until the loss function converges.

9. The multibeam sonar compression imaging method based on depth unfolding according to claim 1, characterized in that, The generation of the sonar depth sounding image includes the following steps: The low-dimensional measurement matrix to be processed is input into a trained depth unwrap beamforming network, which outputs a reconstructed beamforming spectrum. The time sampling points in the beamforming spectrum are converted into slant ranges using the two-way propagation distance formula, and the data is remapped to an angular slant range image grid. The spectral intensity at each grid point is mapped to grayscale or pseudo-color to generate a high-resolution sonar depth sounding image in the angular range domain. The two-way propagation distance formula is: R = (C × T) / (2 × Fs), where R is the slant range, C is the speed of sound in the medium, T is the time sampling point, and Fs is the time sampling frequency.

10. A multibeam sonar compressed imaging system based on depth unfolding, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.

Images