A seabed line detection method based on deep neural network and multi-beam sonar data

Abstract

The application belongs to the technical field of multi-beam sonar data analysis, and discloses a seabed line detection method based on a deep neural network and multi-beam sonar data, which comprises the following steps: collecting simulated and measured seabed line point clouds and sonar original images, reversely converting the point clouds into pixel tensors to construct a pair of data sets; training by using a MIMOUNet-MDTA multi-scale U-shaped network, optimizing the model by combining spatial domain pixel loss and frequency domain Fourier transform loss; converting the model into an inference model, adopting an overlapping sliding window inference strategy after pre-processing the measured sonar image, splicing and restoring into a global high-resolution prediction matrix; filtering and extracting effective pixel coordinates through a threshold, inversely mapping the pixel coordinates into physical space point cloud coordinates based on a bistatic acoustic model; and finally generating and drawing a seabed topographic profile line. The application realizes end-to-end automatic seabed line detection, does not require manual parameter adjustment, improves detection accuracy, and guarantees the spatial continuity and smoothness of the profile line.

Description

Technical Field

[0001] This invention belongs to the field of multibeam sonar data analysis technology, and particularly relates to a method for detecting seabed lines based on deep neural networks and multibeam sonar data. Background Technology

[0002] Multibeam bathymetry sonar is a key piece of equipment in modern marine surveying, underwater target detection, seabed topography and geomorphology surveying, and marine engineering, such as subsea pipeline laying and channel dredging. It can obtain large-scale three-dimensional seabed topographic data at high resolution. In the multibeam sonar data processing workflow, the accurate extraction of seabed lines is the most basic and critical step in constructing three-dimensional seabed topography. However, because sound waves are easily affected by the complex underwater acoustic environment, the original multibeam sonar images are often accompanied by severe speckle noise and sidelobe interference, resulting in blurred seabed edges in the images, which poses a great challenge to high-precision seabed line detection. At present, classic seabed line detection algorithms, such as energy thresholding, phase difference, or point spread model-based methods, essentially rely on the shallow features of local echo signals. The high computational complexity of these algorithms leads to poor real-time performance and weak edge detection and anti-interference capabilities in complex terrain. Secondly, classic algorithms lack robustness, requiring experienced technicians to manually adjust parameters in actual operations, which cannot meet the current needs of real-time on-site processing and fully automated high-precision detection of massive amounts of multibeam sonar data.

[0003] In recent years, deep learning technology has been gradually introduced into the field of sonar image processing, but it still has significant limitations in practical engineering applications of multibeam seabed line detection. First, existing methods deviate from the physical bathymetry mechanism and use a monostatic model during coordinate inversion, ignoring the spatial translational offset error caused by multibeam transmission and reception separation, resulting in systematic positional deviations in the physical point cloud. Second, there are blind spots in the network structure and receptive field. For example, 1D-UNet lacks cross-beam two-dimensional spatial constraints and is highly susceptible to strong sidelobe leakage interference, leading to false tracking jumps. The single-scale structure of basic two-dimensional U-Net is prone to losing high-frequency acoustic textures of the seabed during downsampling. In addition, existing fully connected, fully convolutional, and basic UNet networks have single feature constraints, relying only on spatial domain pixel loss and lacking extraction and constraints of frequency domain features, easily treating real high-frequency details as noise and over-smoothing them. Finally, limited by hardware memory, large-size data often adopts direct slicing and hard stitching strategies, severely fragmenting global contextual information, resulting in tortuosity or truncation artifacts at the boundaries of image blocks, destroying the spatial continuity of the profile lines. Summary of the Invention

[0004] The purpose of this invention is to provide a method for detecting seabed lines based on deep neural networks and multibeam sonar data, so as to solve the above-mentioned technical problems.

[0005] To address the aforementioned technical problems, the present invention provides a specific technical solution for a seabed line detection method based on deep neural networks and multibeam sonar data, as follows:

[0006] A method for detecting seabed lines based on deep neural networks and multibeam sonar data includes the following steps:

[0007] S1: Collect simulation and measured seabed line point clouds and original sonar images to form a dataset; convert the physical space point cloud of the seabed line into a real seabed line pixel tensor with the same shape as the original sonar image to construct a paired dataset.

[0008] S2: Model training is performed using the MIMOUNet-MDTA deep neural network;

[0009] S3: Convert the trained model into an inference model. After preprocessing the original multibeam sonar image matrix to be detected, the model is input into the inference model using an overlapping sliding window inference strategy for forward calculation, and then stitched together to restore the global high-resolution prediction matrix.

[0010] S4: Threshold filtering is performed on the high-resolution prediction matrix to extract the coordinates of effective pixels, and then the coordinates of the pixels are inversely mapped and converted into point cloud coordinates in physical space.

[0011] S5: Generate and draw the seabed topographic profile line based on the point cloud coordinates.

[0012] Furthermore, step S1 includes the following specific steps:

[0013] For a given physical point cloud coordinate of the seabed line, calculate its physical beam deflection angle relative to the vertical axis; obtain the column index by linear mapping based on the total number of beams and the angular coverage range; calculate the transmission slant range and the receiving slant range by combining the translation offset of the receiving array; obtain the round-trip propagation time based on the speed of sound in water, and convert it into a row index by combining the effective sampling frequency of the sonar system; assign high reflectivity amplitude to the pixel coordinates within the effective range, and assign zero to the remaining positions to generate the real seabed line pixel tensor.

[0014] Furthermore, step S2 includes the following specific steps:

[0015] The absolute values ​​of the elements of the raw sonar image matrix after beamforming are taken to eliminate complex phase information, and global normalization to a preset numerical range is achieved through a maximum-minimum linear mapping. The MIMOUNet-MDTA deep neural network is a U-shaped network structure that includes multi-scale feature extraction and multi-scale feature output. It uses the preprocessed raw sonar image tensor and its downsampled tensor as network input, and the corresponding real seabed line pixel tensor and its downsampled tensor as multi-scale supervision labels. A multi-head depthwise convolution transpose attention mechanism is added to the bottom bottleneck layer, and the spatial and frequency domains are calculated. A joint loss function is used to backpropagate and update the network weights until the model converges. During training, multi-scale seabed line prediction tensors are output from each decoding layer of MIMOUNet-MDTA, and the spatial pixel mean absolute error loss between the prediction tensors and the corresponding size real seabed line tensors is calculated. Two-dimensional real number fast Fourier transforms are performed on the prediction tensors and the real tensors, and the frequency domain mean absolute error loss of the real and imaginary parts of the frequency domain matrix is ​​calculated. Finally, a total loss function is constructed for backpropagation to update the model parameters, where the total loss function is a weighted sum of the spatial domain loss and the frequency domain loss.

[0016] Furthermore, the specific configuration of the Multi-Head Deep Convolutional Transposed Attention (MDTA) mechanism is as follows: its input is the semantic feature map tensor extracted from the deepest layer of the U-shaped network encoder; the MDTA consists of multiple cascaded attention blocks, which sequentially perform group normalization on the input feature map tensor, calculate the cross-covariance in the channel dimension, perform multi-head transposed attention feature recalibration, and then perform element-wise addition through residual connections. Further, step S3 includes the following specific steps: solidifying the trained deep neural network weights and exporting them as an inference model file; taking the absolute value of the original sonar image matrix to be detected and performing global normalization; setting the inference window size and an inference step size smaller than this size, so that there is an overlapping area between adjacent sliding windows; sliding and truncating sub-matrices on the sonar image matrix according to the inference step size and inputting them into the inference model; restoring and stitching the output prediction sub-matrices according to spatial coordinates; adopting a dynamic coverage strategy or a weighted fusion strategy for the overlapping area; outputting a global prediction tensor and truncating it to a preset numerical range to obtain a high-resolution prediction matrix.

[0017] Furthermore, S4 includes the following specific steps: applying a preset intensity threshold to truncate and filter the high-resolution prediction matrix output by S3, and extracting the set of valid pixel coordinates in the matrix whose response intensity is greater than the threshold; calculating the actual beam deflection angle based on the column index inverse mapping, and calculating the round-trip propagation time of the sound wave based on the row index; calculating the one-way transmission slant range using the bistatic equivalent slant range analytical formula by combining the spatial translation offset of the transmitting array and the receiving array; and obtaining the YZ point cloud coordinates of the physical space through the orthogonal decomposition of the beam deflection angle and the one-way transmission slant range.

[0018] Furthermore, in step S5, a seabed topographic profile line is generated and drawn based on the calculated YZ point cloud coordinates, thereby completing the high-precision extraction of the seabed line from the multibeam sonar.

[0019] Furthermore, the tensor shape of the original sonar image is the number of beam channels multiplied by the number of sampling points in each channel, and the element values ​​in the tensor are the signal amplitudes; the total number of effective sampling points for each beam channel is determined by the maximum detection distance and sampling frequency set by the sonar system.

[0020] Furthermore, the MIMOUNet-MDTA deep neural network integrates a feature extraction network, a spatial context module, a feature attention module, an encoding residual block, an asymmetric feature fusion module, a decoding residual block, and a multi-scale output convolutional layer. Its macroscopic flow logic does not depend on external multi-scale image input. Instead, it generates downsampled features of different scales from the single-scale sonar raw image tensor through adaptive bilinear interpolation downsampling within the network, thereby constructing a multi-scale parallel processing pipeline.

[0021] Furthermore, during model training, a chamfered distance index function is used to evaluate the spatial topological deviation between the seabed point cloud predicted by the deep neural network and the real seabed point cloud. The physical dimension of this chamfered distance is meters, which is used to directly characterize the absolute spatial topological deviation between the predicted seabed topographic profile and the real seabed topographic profile.

[0022] The seabed line detection method based on deep neural networks and multibeam sonar data of the present invention has the following advantages:

[0023] (1) It realizes the end-to-end process from the original multibeam sonar image to the extracted seabed line, meets the real-time processing requirements on site, and does not require manual parameter adjustment and manual processing.

[0024] (2) During the training phase, pixel-level loss in the spatial domain and frequency domain loss function based on two-dimensional real number fast Fourier transform are used to enable deep neural networks to achieve high-precision extraction of seabed lines.

[0025] (3) The overlapping sliding window inference strategy can process the original ultra-high resolution sonar image as a whole, maintain the contextual continuity of the sliding window edge area, eliminate the truncation artifacts, faults or jumps caused by traditional direct block hard stitching, and ensure the high smoothness and topological continuity of the finally extracted seabed profile line in space.

[0026] (4) Introducing a multi-head deep convolution transpose attention mechanism at the bottom bottleneck layer of the network reduces the computational complexity from quadratic to linear. This not only effectively breaks through the computational bottleneck of high-dimensional feature space and ensures the real-time inference efficiency of underwater edge devices, but also adaptively aggregates the key channel features of seabed echoes in the deep semantic space that filters out high-frequency speckle noise, thereby improving the accuracy of seabed line extraction in complex seabed reverberation environments. Attached Figure Description

[0027] Figure 1 This is a flowchart of the seabed line detection method based on deep neural networks and multibeam sonar data proposed in this invention;

[0028] Figure 2(a) is a schematic diagram of the MIMOUNet-MDTA neural network structure;

[0029] Figure 2(b) is a schematic diagram of the neural network structure of the multi-head deep convolution transposed attention mechanism;

[0030] Figure 3(a) is a schematic diagram of the raw data from the multibeam echo sounder;

[0031] Figure 3(b) is a schematic diagram of the seabed line inference of a multibeam bathymetry sonar;

[0032] Figure 3(c) shows the set of seabed line points extracted from the original sonar image; Detailed Implementation

[0033] To better understand the purpose, structure, and function of this invention, the following detailed description of a seabed line detection method based on deep neural networks and multibeam sonar data is provided in conjunction with the accompanying drawings.

[0034] like Figure 1 As shown, the present invention provides a method for detecting seabed lines based on deep neural networks and multibeam sonar data, which includes the following steps:

[0035] S1: The simulated and measured seabed line point cloud and the original sonar image are used to form a dataset. For the seabed line point cloud, based on the spatial geometric relationship of the multibeam sonar transmission and reception separation, it is converted from the YZ point cloud coordinates in physical space to the real seabed line pixel coordinates that are consistent with the shape of the original sonar image, so as to obtain a paired dataset.

[0036] Given the physical space point cloud coordinates of the seabed line Calculate its physical beam deflection angle relative to the vertical axis. Based on the total number of beams and angular coverage Linear mapping yields column indexes Combined with the translation offset of the receiving array Calculate the launch slant range and receiving slant range :

[0037]

[0038]

[0039] Based on the speed of sound in water Obtain round-trip propagation time And combined with the effective sampling frequency of the sonar system Convert to row index A high reflectivity magnitude is assigned to the pixel coordinates within the effective range, and zero is assigned to the remaining positions to generate a real seabed line pixel tensor. Here, round() is the rounding function.

[0040] S2: Model training is performed using the MIMOUNet-MDTA deep neural network. The MIMOUNet-MDTA is a U-shaped network structure that includes multi-scale feature extraction and multi-scale feature output. The preprocessed sonar original image tensor and its downsampled tensor are used as network inputs, and the corresponding real seabed line pixel tensor and its downsampled tensor are used as multi-scale supervision labels. A multi-head deep convolutional transpose attention mechanism (MDTA) is added to the bottom bottleneck layer. The network weights are updated by backpropagation through the calculation of the joint loss function in the spatial domain and frequency domain until the model converges.

[0041] The absolute values ​​of the elements of the raw sonar image matrix after beamforming are taken to eliminate complex phase information, and then globally normalized to the minimum-maximum linear mapping. The range, the original sonar image size is The total number of effective sampling points for each beam channel The calculation formula is as follows:

[0042]

[0043] number of beam channels Set parameters for the sonar system. The maximum detection range set for the sonar system.

[0044] The multi-head deep convolutional transposed attention mechanism is characterized by the following: its input is a semantic feature map tensor with low spatial resolution and high dimensionality extracted from the deepest layer of the U-shaped network encoder; the module consists of multiple cascaded attention blocks, which sequentially perform group normalization on the input feature map tensor and calculate the cross-covariance in the channel dimension for multi-head transposed attention feature recalibration, and then perform element-wise addition through residual connections; the output tensor after aggregation by the MDTA module retains the original spatial dimension and is input to the decoder upsampling module and asymmetric feature fusion module (AFF) respectively.

[0045] For the number of input channels ,length ,width Feature tensor ( ), MDTA first passed Pointwise convolution aggregates cross-channel information, and then... The spatial context of depthwise convolutional encoding. This can be obtained. , and matrix:

[0046]

[0047] in and These represent the weights of pointwise convolution and depthwise convolution, respectively. , and Given the query matrix, key matrix, and value matrix, respectively, the channel-level transpose attention map is calculated as follows:

[0048]

[0049] in A learnable temperature parameter is used to adjust the sharpness of the attention distribution. The acoustic features of the real seabed boundary are concentrated in specific implicit channels; MDTA calculates the cross-channel covariance. It adaptively assigns higher weights to channels containing valid tracking cues. This mechanism effectively suppresses ambient noise in water bodies and replaces deep stacked convolutional layers, thereby reducing computational cost.

[0050] During training, multi-scale seabed line prediction tensors are output from each decoding layer of MIMOUNet-MDTA, and the spatial pixel mean absolute error loss between each tensor and the corresponding size of the real seabed line tensor is calculated. Perform two-dimensional real-number fast Fourier transforms on the predicted tensor and the true tensor respectively, and calculate the frequency domain mean absolute error loss for the real and imaginary parts of the frequency domain matrix. Finally, construct the total loss function. Perform backpropagation to update the model parameters, where This represents the frequency domain loss weighting coefficient.

[0051] S3: The trained MIMOUNet deep neural network is converted into an inference model. The original image matrix of the measured multibeam sonar to be detected is preprocessed and then input into the inference model using an overlapping sliding window inference strategy for forward calculation. The matrix is ​​then stitched together to restore the global high-resolution prediction matrix.

[0052] The specific process of employing the overlapping sliding window inference strategy includes: solidifying the trained deep neural network weights and exporting them as an inference model file; taking the absolute value of the original sonar image matrix to be detected and performing max-min linear global normalization; setting the inference window size and an inference step size smaller than that size, so that there is an overlapping area between adjacent sliding windows; sliding and truncating sub-matrices on the sonar image matrix according to the inference step size and inputting them into the inference model; restoring and stitching the output prediction sub-matrices according to spatial coordinates; adopting a dynamic coverage strategy for the overlapping area; outputting the global prediction tensor and truncating it to... A high-resolution prediction matrix is ​​obtained within the range.

[0053] S4: Obtain the high-resolution prediction matrix output by the inference model, apply a preset intensity threshold to the global high-resolution prediction matrix for truncation and filtering, extract the set of effective pixel coordinates in the matrix whose response intensity is greater than the threshold, and inversely map the set of effective pixel coordinates into YZ point cloud coordinates in physical space according to the bistatic equivalent slant distance analytical formula and polar coordinate transformation relationship.

[0054] The set of pixel coordinates is obtained by filtering out valid pixels in the prediction matrix whose data is greater than a preset intensity threshold. The actual beam deflection angle is calculated by inverse mapping:

[0055]

[0056] and the round-trip time of sound waves Using the bistatic equivalent slant range analytical formula, a high-precision single-pass launch slant range is obtained:

[0057]

[0058] The YZ point cloud coordinates were obtained through orthogonal decomposition. and .

[0059] S5: Generate and draw seabed topographic profile lines based on the calculated YZ point cloud coordinates, completing the high-precision extraction of seabed lines from multibeam sonar.

[0060] Example:

[0061] This embodiment of a method for detecting seabed lines based on deep neural networks and multibeam sonar data includes the following steps:

[0062] S1: Acquire simulated and real-world seabed line and sonar raw images. The seabed line and sonar raw images together form a dataset. Transform the seabed line from YZ point cloud coordinates to pixel coordinates with the same shape as the sonar raw image. The tensor shape of the sonar raw image is... , Number of beam channels For each channel, the number of sampling points is given, and the element values ​​in the tensor are the signal amplitudes. The specific mapping process is as follows:

[0063] Calculate the beam deflection angle of the target point in the YZ plane:

[0064]

[0065] Based on the set number of beam channels and beam coverage angle range Map the angle to the column index of the pixel coordinate matrix:

[0066]

[0067] Based on the bistatic acoustic equivalent model with separate transmit and receive, the one-way transmit slant range is calculated. With receiving slant range :

[0068]

[0069]

[0070] Based on the speed of sound in water With sonar system sampling frequency Calculate the round-trip time of the sound wave , obtain the row indices of the pixel coordinate matrix:

[0071]

[0072] The pixel coordinate elements are assigned a preset high reflectivity amplitude, and the rest of the area is set to zero, thus obtaining a real seabed line tensor that is completely aligned with the original sonar tensor.

[0073] S2: In this embodiment, the raw sonar image is preprocessed by first taking the absolute value of the matrix elements to eliminate complex phase, and then using a max-min linear mapping to globally normalize the element amplitudes. The interval is converted to single-precision floating-point format. As shown in Figures 2(a) and 2(b), in terms of network training, the MIMOUNet-MDTA deep neural network adopts a multi-scale input and output structure. In order to eliminate sidelobe interference and retain high-precision seabed line information, the loss function in this embodiment includes the spatial domain mean absolute error loss between the prediction results at each scale and the real label. And the frequency domain loss after performing a two-dimensional real-number fast Fourier transform on the feature map. The final loss was ,Pick The network parameters are updated via backpropagation using a loss function. The MIMOUNet-MDTA deep neural network integrates a feature extraction network, a spatial context module (SCM), a feature attention module (FAM), an coded residual block (EBlock), an asymmetric feature fusion module (AFF), a multi-head deep convolutional transpose attention mechanism (MDTA), a decoded residual block (DBlock), and a multi-scale output convolutional layer (ConvsOut). Its macroscopic processing logic does not rely on external multi-scale image input. Instead, it generates 1 / 2 downsampled features and 1 / 4 downsampled features from the single-scale sonar raw image tensor within the network through adaptive bilinear interpolation downsampling, thereby constructing a multi-scale parallel processing pipeline.

[0074] S3: Convert the trained MIMOUNet-MDTA deep neural network into an inference model file in engine format. Input the measured original multibeam sonar image matrix to be detected into the inference model for calculation using an overlapping sliding window inference strategy. The overlapping sliding window inference strategy addresses the problem that high-resolution measured original sonar images are prone to memory overflow during inference, and that traditional direct block stitching is prone to truncation artifacts at the boundaries. Specifically, a fixed inference window size is set. It also includes a sliding step size of 1000. By controlling the sliding window compensation, it preserves the overlapping area between adjacent inference sub-blocks, uses the global context information of the network in the overlapping area for smooth transition, and performs weighted fusion of the overlapping area when finally splicing it into the result matrix, thereby outputting a seamless and continuous high-resolution seabed line prediction matrix.

[0075] S4: Obtain the high-resolution prediction matrix output by the inference model, and the set of pixel coordinates of the seabed line in the matrix. Converting to YZ point cloud coordinates in physical space, this embodiment sets an intensity threshold to filter out minimum noise in the prediction matrix and extracts the set of pixel coordinates greater than the threshold. The actual beam deflection angle is obtained by inverse mapping of the column index:

[0076]

[0077] Calculate the round-trip time of the sound wave based on the row index. Based on transmit and receive offsets Calculate the one-way launch slant range:

[0078]

[0079] Obtain the YZ coordinates through orthogonal decomposition:

[0080]

[0081]

[0082] S5: Based on the calculated discrete YZ point cloud coordinates, the seabed topographic profile is directly drawn to obtain a true-scale seabed topographic profile, thus completing the high-precision extraction of the seabed line by multibeam sonar.

[0083] In this embodiment, 270 pairs of original sonar images and corresponding seabed line pixel coordinate images are used for training. It is 512. The resolution is 9231, meaning the original sonar image resolution is a single channel, 512 pixels wide and 9231 pixels high. The chamfered distance index function is used during training to evaluate the YZ coordinate point cloud of the seabed line predicted by the deep neural network. Point cloud with YZ coordinates of the actual seabed line The distance is as follows:

[0084]

[0085] in, and They are sets and Total number of points included, chamfer distance A smaller chamfer distance indicates better model performance. The physical dimension of this chamfer distance is meters (m), directly representing the absolute spatial topological deviation between the predicted seabed topographic profile and the actual seabed topographic profile. A smaller average chamfer distance indicates higher accuracy in point cloud inversion in 3D geographic space and lower risk of overfitting. This example uses the optimal model after training and performs inference tests on a test set of 30 pairs of raw sonar images and corresponding seabed line pixel coordinate images. The average chamfer distance between the predicted seabed line YZ point cloud and the actual seabed line YZ point cloud is 6.81 meters. The average inference time is 0.0036 seconds / frame on an RTX 5070Ti graphics card and 0.486 seconds / frame on the Orin NX edge computing platform, demonstrating that deep neural networks can extract seabed lines in real-time with high accuracy.

[0086] Figures 3(a) and 3(b) show a comparison between the raw data and the inferred seabed lines from the multibeam echo sounder in a real, complex underwater acoustic environment according to an embodiment of the present invention. To facilitate a direct evaluation of the detection accuracy and terrain matching degree, all data in Figures 3(a) and 3(b) are converted to the YZ spatial coordinate system. Specifically, Figure 3(a) shows the raw data from the multibeam echo sounder. It can be clearly observed from the figure that, due to the complexity of the real underwater acoustic environment, there are multiple obvious sidelobe leakage interferences caused by strong reflections, which appear as arc-shaped bright lines running through multiple beam channels. The red lines in Figure 3(b) are the seabed topographic profile lines extracted using the method proposed in this invention. The red lines in Figure 3(c) are the point sets of the seabed topographic profile lines extracted using the method proposed in this invention.

[0087] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for detecting seabed lines based on deep neural networks and multibeam sonar data, characterized in that, Includes the following steps: S1: Collect simulation and measured seabed line point clouds and original sonar images to form a dataset; convert the physical space point cloud of the seabed line into a real seabed line pixel tensor with the same shape as the original sonar image to construct a pair of datasets; S2: Model training is performed using the MIMOUNet-MDTA deep neural network; The absolute value of the elements of the original sonar image matrix after beamforming is taken to eliminate complex phase information, and the matrix is ​​globally normalized to a preset numerical range through a linear mapping of maximum and minimum values. The MIMOUNet-MDTA deep neural network is a U-shaped network structure that includes multi-scale feature extraction and multi-scale feature output. It uses the preprocessed sonar original image tensor and its downsampled tensor as network input, and the corresponding real seabed line pixel tensor and its downsampled tensor as multi-scale supervision labels. A multi-head depthwise convolutional transpose attention mechanism (MDTA) is added to the bottom bottleneck layer. The network weights are updated by backpropagation through the calculation of the joint loss function in the spatial and frequency domains until the model converges. During training, multi-scale seabed line prediction tensors are output from each decoding layer of MIMOUNet-MDTA, and the spatial pixel mean absolute error loss between these prediction tensors and the corresponding size real seabed line tensors is calculated. Two-dimensional real-number fast Fourier transforms are performed on both the prediction and real tensors, and the frequency domain mean absolute error loss of the real and imaginary parts of the frequency domain matrix is ​​calculated. Finally, a total loss function is constructed for backpropagation to update the model parameters, where the total loss function is a weighted sum of the spatial domain loss and the frequency domain loss. S3: Convert the trained model into an inference model. After preprocessing the original multibeam sonar image matrix to be detected, the model is input into the inference model using an overlapping sliding window inference strategy for forward calculation, and then stitched together to restore the global high-resolution prediction matrix. S4: Threshold filtering is performed on the high-resolution prediction matrix to extract the coordinates of effective pixels, and then the coordinates of the pixels are inversely mapped and converted into point cloud coordinates in physical space. S5: Generate and draw the seabed topographic profile line based on the point cloud coordinates.

2. The method for detecting seabed lines according to claim 1, characterized in that, S1 includes the following specific steps: For a given physical point cloud coordinate of the seabed line, calculate its physical beam deflection angle relative to the vertical axis; obtain the column index by linear mapping based on the total number of beams and the angular coverage range; calculate the transmission slant range and the receiving slant range by combining the translation offset of the receiving array; obtain the round-trip propagation time based on the speed of sound in water, and convert it into a row index by combining the effective sampling frequency of the sonar system; assign high reflectivity amplitude to the pixel coordinates within the effective range, and assign zero to the remaining positions to generate the real seabed line pixel tensor.

3. The method for detecting submarine lines according to claim 1, characterized in that, The specific configuration of the Multi-Head Deep Convolutional Transposed Attention (MDTA) mechanism is as follows: its input is the semantic feature map tensor extracted by the deepest layer of the U-shaped network encoder; the MDTA is composed of multiple cascaded attention blocks, which sequentially perform group normalization on the input feature map tensor, calculate the cross-covariance in the channel dimension, perform multi-head transposed attention feature recalibration, and then perform element-wise addition through residual connections.

4. The method for detecting submarine lines according to claim 1, characterized in that, S3 includes the following specific steps: solidifying the trained deep neural network weights and exporting them as an inference model file; taking the absolute value of the original sonar image matrix to be detected and performing global normalization; Set the inference window size and an inference step size smaller than that size to create an overlapping area between adjacent sliding windows; slide and truncate a sub-matrix on the sonar image matrix according to the inference step size and input it into the inference model; restore and stitch the output prediction sub-matrix according to spatial coordinates; adopt a dynamic coverage strategy or a weighted fusion strategy for the overlapping area; output the global prediction tensor and truncate it to a preset value range to obtain a high-resolution prediction matrix.

5. The method for detecting submarine lines according to claim 1, characterized in that, S4 includes the following specific steps: applying a preset intensity threshold to truncate and filter the high-resolution prediction matrix output by S3, and extracting the set of valid pixel coordinates in the matrix whose response intensity is greater than the threshold; calculating the actual beam deflection angle based on the column index inverse mapping, and calculating the round-trip propagation time of the sound wave based on the row index; calculating the one-way transmission slant range using the bistatic equivalent slant range analytical formula by combining the spatial translation offset of the transmitting array and the receiving array; and obtaining the YZ point cloud coordinates of the physical space through the orthogonal decomposition of the beam deflection angle and the one-way transmission slant range.

6. The method for detecting submarine lines according to claim 5, characterized in that, In step S5, the seabed topographic profile line is generated and drawn based on the calculated YZ point cloud coordinates, thus completing the high-precision extraction of the seabed line by multibeam sonar.

7. The method for detecting submarine lines according to claim 1, characterized in that, The tensor shape of the sonar raw image is the number of beam channels multiplied by the number of sampling points in each channel, and the element values ​​in the tensor are the signal amplitudes; the total number of effective sampling points for each beam channel is determined by the maximum detection distance and sampling frequency set by the sonar system.

8. The method for detecting seabed lines according to claim 1, characterized in that, The MIMOUNet-MDTA deep neural network integrates a feature extraction network, a spatial context module, a feature attention module, an encoding residual block, an asymmetric feature fusion module, a decoding residual block, and a multi-scale output convolutional layer. Its macroscopic flow logic does not depend on external multi-scale image input. Instead, it generates downsampled features of different scales from the single-scale sonar raw image tensor within the network through adaptive bilinear interpolation downsampling, thereby constructing a multi-scale parallel processing pipeline.

9. The method for detecting submarine lines according to claim 1, characterized in that, During model training, the chamfer distance index function is used to evaluate the spatial topological deviation between the seabed line point cloud predicted by the deep neural network and the real seabed line point cloud. The physical dimension of this chamfer distance is meters, which is used to directly characterize the absolute spatial topological deviation between the predicted seabed topographic profile and the real seabed topographic profile.

Images