A side-scan sonar method for seabed elevation detection combining deep learning and image processing

Abstract

This invention belongs to the field of marine surveying technology and discloses a side-scan sonar seabed elevation detection method combining deep learning and image processing. First, this invention utilizes the Unet deep learning network to classify seabed and non-seabed areas in side-scan sonar source strip images. Then, based on the segmentation, image processing methods such as filtering, gradient template matching, and symmetry checking are applied to utilize prior information such as gradient features and symmetry of seabed boundaries that are difficult to express in the Unet deep learning network in subsequent image processing methods. This combines the advantages of deep learning and image processing to achieve accurate and reliable identification of seabed lines, and thus measure seabed elevation.

Description

Technical Field

[0001] This invention belongs to the field of marine surveying and mapping technology, and more specifically, relates to a side-scan sonar seabed elevation detection method that combines deep learning and image processing. Background Technology

[0002] Side-scan sonar is one of the main instruments for underwater exploration and marine surveys. It possesses irreplaceable comprehensive advantages in high resolution, high mapping efficiency, and low cost, and is widely used. Conventional side-scan sonar only records the intensity of the backscattered sound waves emitted in chronological order. Then, based on the target slant range and the measured elevation, the horizontal distance is calculated to obtain an underwater topographic image. It is evident that the accuracy of the underwater elevation measurement at each measurement position of the side-scan sonar directly determines the positional accuracy of each point on each cross-section of the topographic image, thus determining the accuracy of the topographic image. The sound waves emitted by the side-scan sonar typically travel a distance through the water body, reaching the seabed and reflecting back. The echo intensity reflects the underwater characteristics; recording the echo intensity sequence yields the underwater features of the current cross-section. As the sonar platform moves, it continuously records the backscattered echo intensity, thus obtaining a strip image of the underwater topography from lines to surfaces. The image characteristics are: a relatively obvious gradient change in backscattered intensity from weak to strong at the seabed, starting from the center point and moving outwards.

[0003] Existing methods for measuring seabed elevation using side-scan sonar employ gradient changes or manual marking of the seabed line, or extract the seabed line from side-scan sonar images by setting echo intensity thresholds, and then calculate the seabed elevation. However, manual marking methods are time-consuming and labor-intensive, and their measurement efficiency and cost do not meet the application requirements of large-scale marine surveys. Threshold or gradient change detection methods are easily affected by sea surface echoes or echoes from cluttered targets in the water, making it difficult to achieve accurate and reliable automated detection results. Summary of the Invention

[0004] This invention provides a side-scan sonar seabed elevation detection method that combines deep learning and image processing, thereby solving the problems of low measurement efficiency and low accuracy of side-scan sonar seabed elevation in the prior art.

[0005] This invention provides a side-scan sonar method for seabed elevation detection combining deep learning and image processing, comprising the following steps:

[0006] Step 1: Obtain the side-scan sonar source strip image;

[0007] Step 2: Input the side-scan sonar source strip image into the trained Unet deep learning network. The trained Unet deep learning network outputs a segmentation image, which classifies and labels the pixels in the side-scan sonar source strip image as seabed areas and non-seabed areas.

[0008] Step 3: Filter the segmented image to remove interference clusters;

[0009] Step 4: Using the ping data from the seabed line tracing, the template is matched with the starboard and port regions of the filtered segmented image and then processed accordingly to obtain the starboard and port seabed positions.

[0010] Step 5: Perform symmetry verification and seabed repositioning on the starboard and port seabed positions to obtain the starboard and port seabed lines;

[0011] Step 6: Calculate the seabed elevation based on the starboard seabed line and the port seabed line.

[0012] Preferably, in step 1, the side-scan sonar source strip image is formed by stitching together the port side strip image and the starboard side strip image obtained using a two-dimensional side-scan sonar measurement platform, and the centerline of the side-scan sonar source strip image represents the emission starting point of the side-scan sonar detection sound source.

[0013] Preferably, step 2 includes constructing a Unet deep learning network and a dataset; the dataset is formed based on several measured side-scan sonar source strip images, combined with image processing and manual verification and correction, and the dataset includes a training set and a validation set; the constructed Unet deep learning network is trained and validated based on the training set and the validation set to obtain the trained Unet deep learning network.

[0014] Preferably, the dataset further includes a test set, and the detection results of the starboard sealine and the port sealine are evaluated based on the ground truth results of the sealine positions in the test set.

[0015] Preferably, the Unet deep learning network first performs two 3×3 convolution operations with the ReLU function as the activation function, and then performs a max pooling downsampling operation to halve the image size and increase the feature dimension; the above operation is repeated 4 times, the image size is halved 4 times, and the feature dimension is increased 4 times, forming a feature extraction network that encodes the high-resolution image into high-dimensional low-resolution image features.

[0016] The high-dimensional low-resolution image features are upsampled once and then subjected to two 3×3 convolution operations to double the image size and reduce the feature dimensionality. The above features are then concatenated with the corresponding coding layer output features to fuse the shallow and deep features. The above upsampling, convolution, and concatenation operations are repeated four times to obtain a classification prediction result with the same size as the input image.

[0017] Preferably, in step 3, the area of ​​closed clusters and clustered regions in the segmented image is calculated by using the number of pixels of the detected connected components, and interference clusters in small regions are removed.

[0018] Preferably, in step 4, the ping data matching template for the seabed line tracking includes a first matching template and a second matching template, wherein the first matching template is denoted as Tr = [0,0,1,1,1,1] and the second matching template is denoted as Tl = [1,1,1,1,0,0];

[0019] The first matching template is correlated with the starboard region of the filtered segmented image ping-by-ping to obtain the starboard correlation value sequence Rr. i , i is the ping sequence number; from the starboard related value sequence Rr i Find the maximum correlation value in the i-th ping. The location of this maximum value is the starboard seabed boundary point of the i-th ping. The pixel distance between the starboard seabed boundary point of the i-th ping and the line in the side-scan sonar source strip image is taken as the starboard seabed pixel distance of the i-th ping and denoted as Br. i The pixel distance sequence of all ping's starboard seabed positions is denoted as Br;

[0020] The second matching template is correlated with the port side region of the filtered segmented image ping-by-ping to obtain the port side correlation value sequence Rl. i i is the ping sequence number; from the port side related value sequence Rl i Find the maximum correlation value in the i-th ping. The location of this maximum value is the port side seabed boundary point of the i-th ping. The pixel distance between the port side seabed boundary point of the i-th ping and the center line of the side-scan sonar source strip image is taken as the port side seabed pixel distance of the i-th ping and denoted as Bl. i The pixel distance sequence of all ping's port side seabed positions is denoted as Bl.

[0021] Preferably, in step 5, the symmetry check is implemented as follows: A symmetry check is performed on all pings for both port and starboard sides. For the i-th ping, it is determined whether |Bl| is satisfied. i –Br i |>T0, where T0 is the symmetry threshold; ping that satisfies this formula is considered not to satisfy the symmetry assumption and requires seabed relocation.

[0022] Preferably, the specific implementation method of the seabed repositioning is as follows:

[0023] For each ping data that does not satisfy the symmetry assumption, the distance from the sound source's emission origin to the port side seabed pixel distance Bl is taken as the center. iBased on this, find its starboard symmetrical point P(i,j) such that its seabed pixel distance j = Bl i In the subregion [Ps(i,jL / 2):Pe(i,j+L / 2)] of length L+1 at point P(i,j), find the starboard-related value sequence Rr. i The largest point, whose seabed pixel distance is denoted as Br i ', with |Bl i –Br i '| represents the degree of symmetry deviation between the two when the port side is used as a reference;

[0024] Centered on the origin of the sound source, the distance of the seabed pixels on the starboard side of the ping from Br i Using this as a reference, find its port side symmetrical point, and within its sub-region of length L+1, find the port side correlation value sequence Rl. i The largest point, whose seabed pixel distance is denoted as Bl. i ', with |Bl i '–Br i |This characterizes the degree of symmetry deviation between the two when the starboard side is taken as the reference.

[0025] Comparing the two sets of symmetry deviations above, if |Bl i –Br i '|<|Bl i '–Br i |, then let Br i =Br i ', otherwise let Bl i =Bl i This enables the relocation function.

[0026] Preferably, in step 6, the seabed elevation of each ping is calculated using the following formula:

[0027]

[0028] Among them, H i Let B be the elevation of the i-th ping, and k be the actual distance between adjacent pixels; i Let B be the pixel distance to the seabed on one side of the i-th ping. If it is the port side, then B i The port side seabed pixel distance Bl corresponding to the i-th ping i If it is on the starboard side, then B i The starboard seabed pixel distance Br corresponding to the i-th ping i c is the speed of sound underwater, T s The sampling period for the side-scan sonar image data points.

[0029] One or more technical solutions provided in this invention have at least the following technical effects or advantages:

[0030] This invention first utilizes the Unet deep learning network to classify seabed and non-seabed areas in side-scan sonar source strip images. Then, based on the segmentation, it applies image processing methods such as filtering, gradient template matching, and symmetry verification to utilize prior information such as gradient features and symmetry of seabed boundaries that are difficult to express in the Unet deep learning network in subsequent image processing methods. This combines the advantages of deep learning and image processing to achieve accurate and reliable identification of seabed lines and measurement of seabed elevation, laying the foundation for subsequent conversion of slant range information to horizontal range information and synthesis of topographic images. Attached Figure Description

[0031] Figure 1 This is a side-scan sonar source strip image;

[0032] Figure 2 This is a map showing the results of seabed line extraction using side-scan sonar; among them, Figure 2 (a) is a side-scan sonar source strip image. Figure 2 (b) is the segmented image obtained after classifying the side-scan sonar source strip image using a trained Unet deep learning network. Figure 2 (c) is the final extracted seabed line;

[0033] Figure 3 A flowchart illustrating a side-scan sonar seabed elevation detection method combining deep learning and image processing, provided as an embodiment of the present invention;

[0034] Figure 4 This is a schematic diagram of the Unet deep learning network in a side-scan sonar seabed elevation detection method combining deep learning and image processing, provided in an embodiment of the present invention. Detailed Implementation

[0035] The side-scan sonar emits two beams of sound waves towards the seabed. The beams are narrow in the direction of the sonar platform's advance, but wide in the transverse direction perpendicular to the advance direction, covering a transverse section. After the emitted sound waves encounter a seabed target and reflect, the side-scan sonar platform records the intensity of the reflected echoes sequentially over time, thus obtaining a one-dimensional intensity sequence for each of the two sections. During navigation, the side-scan sonar platform continuously acquires one-dimensional intensity sequences of consecutive sections on both sides of the port and starboard sides, and stacks these sequences to form two two-dimensional side-scan sonar topographic strip images covering the port and starboard side sections—the side-scan sonar source strip images.

[0036] Figure 1This is a side-scan sonar source strip image, directly stitched together from the port and starboard side strip images. The left half is the port side strip image, and the right half is the starboard side strip image. The centerline of the image indicates the source emission location; the darker areas on both sides indicate strong echo intensity, usually seabed echoes; the darker areas between the source and the seabed indicate no target echo received, usually water bodies; irregular areas of strong echo intensity between the source and the seabed usually indicate surface echoes or reverberation / clutter interference. These factors greatly affect the accurate extraction of the seabed line.

[0037] To meet the need for automatic, accurate, and reliable extraction of seabed elevation in complex environments and improve the mapping accuracy and quality of side-scan sonar topographic images, this invention uses a combination of deep learning and image processing to identify and extract seabed lines from port and starboard strip images acquired by a two-dimensional side-scan sonar measurement platform. Based on the extracted seabed lines, the elevation information of the sonar platform relative to the seabed directly below it is calculated. Using this elevation information as a foundation, the slant distance information in the strip image data can be further converted into horizontal distance information, thereby improving the accuracy of the side-scan sonar topographic images.

[0038] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0039] See Figure 3 The present invention provides a side-scan sonar method for seabed elevation detection combining deep learning and image processing, comprising the following steps:

[0040] Step 1: Obtain the side-scan sonar source strip image.

[0041] Specifically, a two-dimensional side-scan sonar measurement platform is used to acquire port and starboard stripe images. These images are then directly stitched together to form a regional topographic image, resulting in a side-scan sonar source stripe image. Figure 1 As shown in the image, the center line represents the starting point of the sound wave emitted by the side-scan sonar. The echo data on both sides of the center line are the port and starboard echo data, respectively. The dark band in the middle represents the noise and weak scattered echo received by the sound wave during its propagation in the water, while the bright areas on both sides represent the stronger back-reflected echoes from the target after the sound wave reaches the seabed.

[0042] Step 2: Input the side-scan sonar source strip image into the trained Unet deep learning network. The trained Unet deep learning network outputs a segmentation image, which classifies and labels the pixels in the side-scan sonar source strip image as seabed areas and non-seabed areas.

[0043] This includes constructing the Unet deep learning network and a dataset. The dataset is based on several measured side-scan sonar source strip images, combined with image processing and manual verification and correction. The dataset includes a training set and a validation set. The constructed Unet deep learning network is trained and validated based on the training set and the validation set to obtain the trained Unet deep learning network.

[0044] The trained Unet deep learning network classifies and identifies the input side-scan sonar source strip images, ultimately obtaining a segmented image of the original image size. The segmented image labels the source image as seabed or non-seabed (including water bodies, sea surface reflections, and clutter interference, etc.), see [link to documentation]. Figure 2 .in, Figure 2 (a) is a side-scan sonar source strip image. Figure 2 (b) is the segmented image after the strip image is classified by the Unet network.

[0045] Step 3: Filter the segmented image to remove interference clusters.

[0046] The segmented image is filtered, and the area of ​​closed clusters and clustered regions in the image is calculated by detecting the number of pixels in the connected components, thereby removing interference clusters in small regions.

[0047] Step 4: Using the ping data from the seabed line tracing, match the template with the starboard and port regions of the filtered segmented image, and perform related processing on each ping to obtain the starboard and port seabed positions.

[0048] Specifically, the ping data matching template for the submarine line tracking includes a first matching template and a second matching template. The first matching template is denoted as Tr = [0,0,1,1,1,1], and the second matching template is denoted as Tl = [1,1,1,1,0,0].

[0049] The first matching template is correlated with the starboard region of the filtered segmented image ping-by-ping to obtain the starboard correlation value sequence Rr. i , i is the ping sequence number; from the starboard related value sequence Rr i Find the maximum correlation value in the i-th ping. The location of this maximum value is the starboard seabed boundary point of the i-th ping. The pixel distance between the starboard seabed boundary point of the i-th ping and the centerline (i.e., the sound source location) of the side-scan sonar source strip image is taken as the starboard seabed pixel distance of the i-th ping and denoted as Br. i The pixel distance sequence of all ping's starboard seabed positions is denoted as Br.

[0050] The second matching template is correlated with the port side region of the filtered segmented image ping-by-ping to obtain the port side correlation value sequence Rl. i i is the ping sequence number; from the port side related value sequence Rl i Find the maximum correlation value in the i-th ping. The location of this maximum value is the port side seabed boundary point of the i-th ping. The pixel distance between the port side seabed boundary point of the i-th ping and the center line of the side-scan sonar source strip image is taken as the port side seabed pixel distance of the i-th ping and denoted as Bl. i The pixel distance sequence of all ping's port side seabed positions is denoted as Bl.

[0051] That is, the first matching template is correlated with the starboard region of the filtered segmented image ping by ping, and the maximum correlation value in a ping is found. The location of the maximum correlation value is the starboard seabed position; the second matching template is correlated with the port side region of the filtered segmented image ping by ping, and the maximum correlation value in a ping is found. The location of the maximum correlation value is the port side seabed position.

[0052] like Figure 1 As shown, for a cross-sectional data point of the image, i.e., a ping data point of the side-scan sonar strip image, the neighborhood of the seabed echo line exhibits significant gradient characteristics: with the seabed line as the boundary, one side is a continuous non-seabed region, and the other side is a seabed region. Utilizing this characteristic, this invention establishes a ping data matching template Tr = [0,0,1,1,1,1] for seabed line tracking. Using the Tr template and the right half of the segmented image, i.e., the starboard region, correlation processing is performed on each ping to find the maximum correlation value in a ping. The matching position of this maximum correlation value is then considered the starboard seabed position. Similarly, a ping data matching template Tl = [1,1,1,1,0,0] is established for seabed line tracking. Using the Tl template and the left half of the segmented image, i.e., the port region, correlation processing is performed on each ping to find the maximum correlation value in a ping. The matching position of this maximum correlation value is then considered the port seabed position.

[0053] Step 5: Perform symmetry verification and seabed repositioning on the starboard and port seabed positions to obtain the starboard and port seabed lines.

[0054] Based on the imaging principle of side-scan sonar images and Figure 1 As shown, the seabed positions detected in the port and starboard side-scan sonar strip images are symmetrical. Therefore, the seabed line positions obtained in step 4 are checked for symmetry. Points with significant differences in the degree of deviation of the port and starboard seabed line positions from the symmetrical positions are detected by thresholding, and seabed repositioning calculations are performed on them.

[0055] Specifically, the symmetry check is implemented as follows: A symmetry check is performed on all pings for both port and starboard sides. For the i-th ping, it is determined whether |Bl| is satisfied. i –Br i |>T0, where T0 is the symmetry threshold; ping that satisfies this formula is considered not to satisfy the symmetry assumption and requires seabed relocation to correct erroneous seabed position detection results.

[0056] The specific implementation method of the seabed repositioning is as follows:

[0057] For each ping data that does not satisfy the symmetry assumption, the distance from the sound source's emission origin to the port side seabed pixel distance Bl is taken as the center. i Based on this, find its starboard symmetrical point P(i,j) such that its seabed pixel distance j = Bl i In the subregion [Ps(i,jL / 2):Pe(i,j+L / 2)] of length L+1 at point P(i,j), find the starboard-related value sequence Rr. i The largest point, whose seabed pixel distance is denoted as Br i ', with |Bl i –Br i '| represents the degree of symmetry deviation between the two when the port side is used as a reference.

[0058] Similarly, with the sound source's emission origin as the center, and the seabed pixel distance Br on the starboard side of the ping,... i Using this as a reference, find its port side symmetrical point, and within its sub-region of length L+1, find the port side correlation value sequence Rl. i The largest point, whose seabed pixel distance is denoted as Bl. i ', with |Bl i '–Br i |This characterizes the degree of symmetry deviation between the two when the starboard side is used as a reference.

[0059] Comparing the two sets of symmetry deviations above, if |Bl i –Br i '|<|Bl i '–Br i |, then let Br i =Br i ', otherwise let Bl i =Bl i This enables the relocation function.

[0060] After symmetry verification and seabed relocation, the final extracted seabed line is as follows: Figure 2 As shown in (c).

[0061] In addition, the dataset may also include a test set, which is used to evaluate the detection results of the starboard sealine and the port sealine.

[0062] Specifically, the seabed line detection results obtained by the combined deep learning and image processing algorithms are comprehensively evaluated based on the ground truth seabed line position results from the test set in the side-scan sonar seabed tracking dataset to obtain an objective evaluation of the processing results.

[0063] Step 6: Calculate the seabed elevation based on the starboard seabed line and the port seabed line.

[0064] The seabed elevation of each ping (i.e., each cross section) is calculated using the following formula:

[0065]

[0066] Among them, H i Let B be the elevation of the i-th ping, and k be the actual distance between adjacent pixels; i Let B be the pixel distance to the seabed on one side of the i-th ping. If it is the port side, then B i The port side seabed pixel distance Bl corresponding to the i-th ping i If it is on the starboard side, then B i The starboard seabed pixel distance Br corresponding to the i-th ping i c is the speed of sound underwater, T s The sampling period for the side-scan sonar image data points.

[0067] The Unet deep learning network in this invention will be further described below.

[0068] The dataset for the Unet deep learning network is used for creation, training, validation, and final testing, including the following aspects:

[0069] (1) Based on several measured side-scan sonar source strip image data, the training set, validation set, and test dataset required for the Unet deep learning network were established. Since the automatic seabed elevation detection task of side-scan sonar images differs fundamentally from conventional image classification tasks, and the mechanism and image quality of side-scan sonar images differ significantly from conventional optical images, existing public datasets cannot be used to train the model. Therefore, a dedicated dataset for seabed tracking based on measured side-scan sonar source strip images was established. Using the method described in step 4 above, the seabed lines were pre-extracted from the measured data. Due to the influence of sea surface echoes or water clutter target echoes, the extracted seabed lines contained errors. Through manual verification and labeling, the incorrectly detected seabed lines were corrected, thereby obtaining the true values ​​of the seabed line positions for all side-scan sonar source strip images.

[0070] (2) For the port side strip image of the side-scan sonar, the seabed line is used as the boundary. All pixels in the left side of the seabed are marked as seabed, and all pixels in the right side of the seabed are marked as non-seabed. The non-seabed classification includes water bodies, sea surface reflections, and clutter interference.

[0071] (3) The dataset was formed by the measured side-scan sonar source strip images and their corresponding seabed and non-seabed classification labels. The dataset was divided into training set, validation set and test set in a 6:2:2 ratio.

[0072] (4) Establish such Figure 4 The Unet deep learning network model is shown. After stitching together the measured striped images from the dataset, the network is input into the Unet deep learning network. First, two 3×3 convolution operations are performed with the ReLU activation function. Then, a max-pooling downsampling operation is performed to halve the image size and increase the feature dimension. This process is repeated four times, halving the image size four times and increasing the feature dimension four times, thus forming a feature extraction network, which can also be considered an encoder. This encodes the high-resolution image into high-dimensional low-resolution image features, which contain the multi-dimensional abstract semantic features of the source image. The obtained high-dimensional low-resolution semantic features are then upsampled once and subjected to two more 3×3 convolution operations to double the image size and reduce the feature dimension. These features are then stitched together with the corresponding encoding layer output features, fusing the shallow and deep features. Repeating the upsampling, convolution, and stitching process four times yields a classification prediction result with the same size as the input image.

[0073] (5) Input the measured strip images from the training set into the Unet deep learning network to obtain the predicted classification results of the strip images. Compare the predicted results with the label data from the training set and adjust the parameters of the Unet deep learning network. Evaluate the Unet deep learning network trained on the training set using test set data. Finally, obtain a Unet deep learning network with small generalization error for use in step 2 above.

[0074] The Unet deep learning network in this invention learns the representation of the seabed echo distribution from the dataset, overcoming the interference of surface reflection waves, underwater impurity echoes, and other interference.

[0075] In summary, based on the above-mentioned scheme combining deep learning and image processing, this invention can automatically extract accurate and reliable seabed lines directly from the port and starboard strip images obtained by the two-dimensional side-scan sonar measurement platform, and then measure the seabed elevation.

[0076] Finally, it should be noted that the above specific embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A side-scan sonar method for seabed elevation detection combining deep learning and image processing, characterized in that, Includes the following steps: Step 1: Obtain the side-scan sonar source strip image; Step 2: Input the side-scan sonar source strip image into the trained Unet deep learning network. The trained Unet deep learning network outputs a segmentation image, which classifies and labels the pixels in the side-scan sonar source strip image as seabed areas and non-seabed areas. Step 3: Filter the segmented image to remove interference clusters; Step 4: Using the ping data from the seabed line tracing, the template is matched with the starboard and port regions of the filtered segmented image and then processed accordingly to obtain the starboard and port seabed positions. Step 5: Perform symmetry verification and seabed repositioning on the starboard and port seabed positions to obtain the starboard and port seabed lines; Step 6: Calculate the seabed elevation based on the starboard seabed line and the port seabed line.

2. The side-scan sonar seabed elevation detection method combining deep learning and image processing according to claim 1, characterized in that, In step 1, the side-scan sonar source strip image is formed by stitching together the port side strip image and the starboard side strip image obtained using a two-dimensional side-scan sonar measurement platform. The centerline of the side-scan sonar source strip image represents the emission starting point of the side-scan sonar detection sound source.

3. The side-scan sonar seabed elevation detection method combining deep learning and image processing according to claim 1, characterized in that, Step 2 includes constructing the Unet deep learning network and a dataset; the dataset is formed based on several measured side-scan sonar source strip images, combined with image processing and manual verification and correction, and includes a training set and a validation set; the constructed Unet deep learning network is trained and validated based on the training set and the validation set to obtain the trained Unet deep learning network.

4. The side-scan sonar seabed elevation detection method combining deep learning and image processing according to claim 3, characterized in that, The dataset also includes a test set, based on which the detection results of the starboard and port seafloor lines are evaluated.

5. The side-scan sonar seabed elevation detection method combining deep learning and image processing according to claim 3, characterized in that, The Unet deep learning network first performs two 3×3 convolution operations with the ReLU activation function, and then performs a max pooling downsampling operation to halve the image size and increase the feature dimension. The above operation is repeated 4 times, halving the image size 4 times and increasing the feature dimension 4 times, to form a feature extraction network that encodes high-resolution images into high-dimensional low-resolution image features. The high-dimensional low-resolution image features are upsampled once and then subjected to two 3×3 convolution operations to double the image size and reduce the feature dimensionality. The above features are then concatenated with the corresponding coding layer output features to fuse the shallow and deep features. The above upsampling, convolution, and concatenation operations are repeated four times to obtain a classification prediction result with the same size as the input image.

6. The side-scan sonar seabed elevation detection method combining deep learning and image processing according to claim 1, characterized in that, In step 3, the area of ​​closed clusters and clustered regions in the segmented image is calculated by using the number of pixels in the detected connected regions, and interference clusters in small regions are removed.

7. The side-scan sonar seabed elevation detection method combining deep learning and image processing according to claim 1, characterized in that, In step 4, the ping data matching template for the submarine line tracking includes a first matching template and a second matching template. The first matching template is denoted as Tr = [0,0,1,1,1,1], and the second matching template is denoted as Tl = [1,1,1,1,0,0]. The first matching template is correlated with the starboard region of the filtered segmented image ping-by-ping to obtain the starboard correlation value sequence Rr. i , i is the ping sequence number; from the starboard related value sequence Rr i Find the maximum correlation value in the i-th ping. The location of this maximum value is the starboard seabed boundary point of the i-th ping. The pixel distance between the starboard seabed boundary point of the i-th ping and the line in the side-scan sonar source strip image is taken as the starboard seabed pixel distance of the i-th ping and denoted as Br. i The pixel distance sequence of all ping's starboard seabed positions is denoted as Br; The second matching template is correlated with the port side region of the filtered segmented image ping-by-ping to obtain the port side correlation value sequence Rl. i i is the ping sequence number; from the port side related value sequence Rl i Find the maximum correlation value in the i-th ping. The location of this maximum value is the port side seabed boundary point of the i-th ping. The pixel distance between the port side seabed boundary point of the i-th ping and the center line of the side-scan sonar source strip image is taken as the port side seabed pixel distance of the i-th ping and denoted as Bl. i The pixel distance sequence of all ping's port side seabed positions is denoted as Bl.

8. The side-scan sonar seabed elevation detection method combining deep learning and image processing according to claim 7, characterized in that, In step 5, the symmetry check is implemented as follows: A symmetry check is performed on the port and starboard sides for all pings. For the i-th ping, it is determined whether |Bl| is satisfied. i –Br i |>T0, where T0 is the symmetry threshold; ping that satisfies this formula is considered not to satisfy the symmetry assumption and requires seabed relocation.

9. The side-scan sonar seabed elevation detection method combining deep learning and image processing according to claim 8, characterized in that, The specific implementation method of the seabed repositioning is as follows: For each ping data that does not satisfy the symmetry assumption, the distance from the sound source's emission origin to the port side seabed pixel distance Bl is taken as the center. i Based on this, find its starboard symmetrical point P(i,j) such that its seabed pixel distance j = Bl i In the subregion [Ps(i,jL / 2):Pe(i,j+L / 2)] of length L+1 at point P(i,j), find the starboard-related value sequence Rr. i The largest point, whose seabed pixel distance is denoted as Br i ', with |Bl i –Br i '| represents the degree of symmetry deviation between the two when the port side is used as a reference; Centered on the origin of the sound source, the distance of the seabed pixels on the starboard side of the ping from Br i Using this as a reference, find its port side symmetrical point, and within its sub-region of length L+1, find the port side correlation value sequence Rl. i The largest point, whose seabed pixel distance is denoted as Bl. i ', with |Bl i '–Br i |This characterizes the degree of symmetry deviation between the two when the starboard side is taken as the reference. Comparing the two sets of symmetry deviations above, if |Bl i –Br i '|<|Bl i '–Br i |, then let Br i =Br i ', otherwise let Bl i =Bl i This enables the relocation function.

10. The side-scan sonar seabed elevation detection method combining deep learning and image processing according to claim 7, characterized in that, In step 6, the seabed elevation of each ping is calculated using the following formula: Among them, H i Let B be the elevation of the i-th ping, and k be the actual distance between adjacent pixels; i Let B be the pixel distance to the seabed on one side of the i-th ping. If it is the port side, then B i The port side seabed pixel distance Bl corresponding to the i-th ping i If it is on the starboard side, then B i The starboard seabed pixel distance Br corresponding to the i-th ping i c is the speed of sound underwater, T s The sampling period for the side-scan sonar image data points.

Images