Construction method of adaptive semantic segmentation seabed linear target detection model

Abstract

The invention relates to the technical field of construction of seabed linear target detection models, in particular to a construction method of a self-adaptive semantic segmentation seabed linear target detection model. The method comprises the following steps: obtaining original image data of a seabed scene, extracting a first feature point and a second feature point, and analyzing a structural response difference between the first feature point and the second feature point; calculating a linear rotation angle and a curvature change trend of a connection path based on the structure response difference, constructing a spatial distribution change diagram, extracting a cable continuous trend, and marking a stable section and an abnormal variation section; designing a deep neural network comprising a multi-level convolution structure and a cyclic memory structure, and extracting structure extension and repair features; the depth features of the coherent region and the incoherent region are fused, and a seabed target detection model based on DeepLabv3 + is constructed; semantic segmentation is carried out on the waterfall plot, and the segmentation precision of the model is evaluated to realize optimization; according to the invention, through the construction of the target detection model, the target detection is more accurate and efficient.

Description

Technical Field

[0001] The present invention relates to the technical field of constructing a submarine linear target detection model, and in particular to a method for constructing an adaptive semantic segmentation submarine linear target detection model. Background Art

[0002] Traditional sonar image target detection relies primarily on manually extracted features followed by classification using a classifier. Deep learning-based sonar image target detection typically focuses on issues such as differences between optical and acoustic image domains, small datasets, and high noise characteristics of acoustic images. Demand for automated identification and monitoring of linear structures such as submarine cables and pipelines is growing, and synthetic aperture sonar (SAS), as a high-resolution underwater imaging method, has been widely used for submarine target detection. However, due to the complex and changing submarine environment, submarine images are often affected by factors such as water column obstruction, topographical fluctuations, acoustic interference, and structural morphology variations. This results in linear targets appearing discontinuous, unclear, and with sudden changes in their orientation, severely impacting the accuracy and stability of traditional image processing and target detection algorithms. Existing methods often rely on fixed feature templates or simple edge detection algorithms, which struggle to adapt to the multi-scale, weakly textured, and discontinuous characteristics of cable structures in diverse submarine scenarios. They also lack robustness to anomalies such as partial occlusion and structural inflections, making it difficult to achieve continuous tracking and accurate segmentation of linear targets. Summary of the Invention

[0003] Based on this, it is necessary to provide a method for constructing an adaptive semantic segmentation seabed linear target detection model to solve at least one of the above technical problems.

[0004] To achieve the above objectives, a method for constructing an adaptive semantic segmentation submarine linear target detection model includes the following steps: Step S1: acquiring original image data of a seabed scene and performing feature response processing on the data, wherein the image is analyzed and the starting response point of the target structure is identified as a first feature point; a second feature point representing the location of the structural change is extracted; and the difference in structural response between the first feature point and the second feature point is analyzed; Step S2: Based on the structural response differences, the linear rotation angle and curvature change trend of the connection path are calculated, and a spatial distribution change map of the linear target is constructed. By tracking the offset rate and local connectivity of the structural extension direction, combined with seabed line tracking, the continuous direction of the submarine cable is extracted, and the stable and abnormal variation segments of the submarine cable in the image are marked; Step S3: Design a deep neural network with a multi-level convolutional structure and a recurrent memory structure. Use the convolutional network to capture multi-scale edge trends and linear trends in stable segments, and use the recurrent network to extract and detect structural repair features and contextual sequence features in incoherent regions. Step S4: Fusing the comprehensive features of the extracted detection coherent areas and detection incoherent areas, identifying the seabed target sample set, the model input is the seabed target sample set, and a seabed target detection model based on DeepLabv3+ is constructed; Step S5: Use the submarine cable detection model to segment the cable targets on the waterfall diagram, evaluate the segmentation accuracy of the submarine target detection model based on semantic segmentation, and optimize the model.

[0005] The present invention has the following beneficial effects: First: By introducing the extraction mechanism of the first feature point and the second feature point in the seabed scene and analyzing the structural response differences between them, the problems of weak texture, discontinuity, and blurred edges of linear targets in seabed images are effectively solved. It makes up for the limitation of traditional methods that cannot accurately identify the starting and ending positions of cables in low-contrast and unstructured areas, and provides a more physically meaningful initial feature basis for subsequent path construction and connectivity analysis.

[0006] The second aspect: By extracting the linear rotation angle and curvature change trend of the guided path based on the structural response difference, a spatial distribution change map of the seabed linear target is constructed. Combined with the offset rate and local connectivity in the cable extension direction, the continuous direction characteristics of the target structure in space can be effectively captured, thereby significantly improving the ability to automatically identify stable and abnormal sections of the cable in complex seabed environments, providing a reliable basis for the status assessment and abnormal warning of seabed targets.

[0007] Third aspect: The deep neural network model introduced in the present invention integrates a multi-level convolutional structure and a recurrent memory structure. The convolution module can fully capture the multi-scale edge and continuity features in the stable segment of the cable, and the recurrent network extracts structural repair and contextual association information for areas with interruptions, occlusions, mutations, etc. in the image, thereby improving the structural recognition and recovery capabilities under incomplete image conditions and enhancing the robustness and adaptability of the model.

[0008] Fourth aspect: By fusing the comprehensive features of coherent and incoherent areas and building a seabed semantic segmentation model based on DeepLabv3+, not only the target generalization ability of the model in different seabed scenarios is improved, but also the cable segmentation accuracy is significantly improved, especially under interference conditions such as water column occlusion, texture aliasing, and complex background, it can still maintain high accuracy, ensuring the complete reconstruction and spatial coherence expression of the target structure in the image.

[0009] Fifth aspect: Through this model, the submarine cable targets in the waterfall diagram are segmented with high precision, and the model performance is evaluated and optimized in combination with semantic segmentation indicators, realizing an end-to-end automated processing flow from image preprocessing, feature extraction, spatial modeling to model segmentation, providing technical support for scenarios such as underwater engineering inspection, automatic deployment monitoring and submarine communication maintenance, and has good application and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A schematic diagram of the steps of a method for constructing an adaptive semantic segmentation submarine linear target detection model; Figure 2 This is a schematic diagram of the DeepLabv3+ network structure; Figure 3 Construct a flow chart for the submarine cable detection model; Figure 4 Flowchart for seabed linear target detection using synthetic aperture sonar data; Figure 5 This is a set of submarine cable target images, including (a) pipeline image; (b) prediction result; (c) manual annotation result; Figure 6 The pipeline image detection results, where (a) the original sonar image; (b) the sonar image with the superimposed prediction results; The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0011] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0012] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0013] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0014] To achieve this, please refer to Figures 1 to 6 , a method for constructing an adaptive semantic segmentation seabed linear target detection model, comprising the following steps: Step S1: acquiring original image data of a seabed scene and performing feature response processing on the data, wherein the image is analyzed and the starting response point of the target structure is identified as a first feature point; a second feature point representing the location of the structural change is extracted; and the difference in structural response between the first feature point and the second feature point is analyzed; In one embodiment, a seafloor image dataset of a specific sea area was collected using synthetic aperture sonar. The original image resolution was 2048×1024. The images were initially processed using a preset feature response algorithm, and candidate linear structures were extracted using a combination of edge enhancement and line detection. This algorithm combines the Hessian matrix with directional gradient information to detect significant response regions in the image. Its starting point is designated as the first feature point. Inflection points with significant local response gradient changes are searched for along its linear extension path and marked as second feature points. A structural connection path is then constructed between the first and second feature points. Local directional changes along the path are statistically analyzed to obtain a distribution of structural response differences.

[0015] In another embodiment, synthetic aperture sonar raw data is stored in a binary file. The file contents primarily include: a file header, option data, data type and offset, imaging parameters, image data packet, etc. The file header indicates the overall information of the file. Useful information used for image processing includes port and starboard types, high and low frequency types, image width, and real and imaginary part arrangement. Option data is a user-defined portion, which is optional. If no option data is present, the data option length of the file header should be 0. Data type and offset. Each image data packet contains a line of image data, sensor data, and other data. The data type and offset indicate the offset of each part in the image data packet. Types 512 to 1023 are used for imgpc files. Imaging parameters are parameters required for synthetic aperture sonar imaging, primarily including center frequency, bandwidth, pulse width, sampling rate, pulse repetition period, number of sampling points, number of channels, transmit array length, receive array length, range resolution, azimuth resolution, reference velocity, image width, and other information. Image data packet consists of two parts: image data and image supplementary data. One image data packet contains a line of image data. Image data. Data is stored row-by-row, either real or complex, depending on the arrangement of the real and imaginary components in the file header. Currently, data is stored in an interleaved arrangement of real and imaginary components. Image data is primarily supplemented with sensor parameters for each row of the image, including attitude information such as heading, pitch, roll, and height, cable length, GAPS ultra-short baseline underwater positioning information, and GNSS position information. After the binary data is read, it is assigned to a custom data structure and saved to a txt file for subsequent processing.

[0016] Step S2: Based on the structural response differences, the linear rotation angle and curvature change trend of the connection path are calculated, and a spatial distribution change map of the linear target is constructed. By tracking the offset rate and local connectivity of the structural extension direction, combined with seabed line tracking, the continuous direction of the submarine cable is extracted, and the stable and abnormal variation segments of the submarine cable in the image are marked; In one embodiment, based on the differences in structural responses, a piecewise fitting method is used to calculate the linear rotation angles and curvature change trends of local connection paths. By accumulating gradients pixel by pixel, the spatial extension skeleton of the linear target in the image is obtained. Then, with this skeleton as the center, a spatial distribution change map reflecting the change trend of the linear structure in the image plane is constructed. Combining the offset rate of the structural extension direction with the assessment of the degree of local connectivity, a self-developed submarine cable tracking module is used to track potential cable paths. Identified continuous directions are marked as stable segments in the diagram; when there are signs of sudden curvature changes or breaks, they are identified as abnormal variation segments.

[0017] Step S3: Design a deep neural network with a multi-level convolutional structure and a recurrent memory structure. Use the convolutional network to capture multi-scale edge trends and linear trends in stable segments, and use the recurrent network to extract and detect structural repair features and contextual sequence features in incoherent regions. In one embodiment, a deep neural network architecture is designed. The front-end consists of three convolutional encoding modules, each employing a 3×3 convolution kernel to extract multi-scale edge features and local directional consistency features. The convolution channels are 32, 64, and 128, respectively. To enhance the ability to reconstruct the structure of discontinuous image regions, a two-layer bidirectional GRU network is connected in series on the back-end to extract contextual structural restoration features from time series. During the training phase, the stable segments marked in step S2 are used as samples for detecting coherent regions, and the incoherent segments are used as samples for detecting abnormal regions. The convolutional module and the recurrent module are trained separately to improve the model's sensitivity and robustness in recognizing the coherence of linear objects.

[0018] In another embodiment, DeepLabv3+ employs a typical encoder-decoder architecture. Its main process is as follows: the encoder extracts high-level semantic features from the input image, then uses a multi-scale feature extraction and enhancement module to obtain contextual information at different scales. Finally, the decoder gradually restores spatial details and edge contours, outputting a segmentation mask with the same size as the input image. In the encoder, DeepLabv3+ uses a modified Xception network as its backbone feature extraction architecture, integrating depthwise separable convolution and atrous convolution techniques. Depthwise separable convolution significantly reduces the number of parameters and computational complexity, improving network efficiency. Atrous convolution, by introducing gaps within the convolution kernel, expands the receptive field and enhances the perception of long-range spatial patterns, such as linear structures, without significantly reducing feature map resolution. To model multi-scale environmental context, DeepLabv3+ employs the atrous spatial pyramid pooling (ASPP) module. Unlike traditional pyramid pooling, which uses multi-scale pooling to capture semantic context, ASPP utilizes multiple dilated convolutions with different dilation rates to extract features in parallel. This improves the network's receptive field while maintaining the spatial resolution of the feature map, avoiding information loss caused by downsampling. In a typical configuration, ASPP includes three dilated convolution branches with dilation rates of 6, 12, and 18, a standard 1×1 convolution branch, and a global average pooling branch to capture the overall image context. The outputs of these five branches are concatenated and fused to form a feature vector rich in scale information. The overall network process is as follows: the input image passes through a deep encoder network to extract basic feature maps. These features are then fed into the ASPP module, where they undergo dilated convolutions with different dilation rates and global context modeling before being fused to form multi-scale structure-aware features. These fused features are then fed into a lightweight decoder module, which recovers detailed information layer by layer, outputting refined semantic segmentation results. The results are particularly effective at preserving the edges of slender linear structures, such as submarine cables. This structure works synergistically with the multi-level convolution and recurrent memory network modules described in step S3: the former focuses on extracting the spatial edge and directional trend features of stable targets from the overall scale, while the latter focuses on time series modeling and repairing and completing local incoherent structures, thereby achieving highly robust detection of seabed linear targets in multiple scenarios and interference conditions.

[0019] Step S4: Fusing the comprehensive features of the extracted detection coherent areas and detection incoherent areas, identifying the seabed target sample set, the model input is the seabed target sample set, and a seabed target detection model based on DeepLabv3+ is constructed; In one embodiment, the obtained cable target image is manually annotated. Polygonal annotation mode is selected to finely annotate the target's edge contours. After annotation, the corresponding polygon mask can be saved to construct a submarine cable target sample set. The coherent region features extracted in step S3 are fused with the discontinuous region features to construct a complete submarine linear target sample set, which serves as input for model training. Based on the DeepLabv3+ backbone network, the Atrous Spatial Pyramid (ASPP) structure is introduced to enhance the responsiveness to targets of different scales. A branch loss function is used for supervised learning of coherent segments and discontinuous segments, thereby improving the model's robust recognition capabilities for submarine cable targets in complex scenarios.

[0020] In another embodiment, a submarine cable target sample set is randomly divided into a training set and a test set. The model is trained on the training set. The model input is a submarine cable target image. The DeepLabv3+ model performs feature extraction, enhancement, and edge information restoration, outputting a mask of the same size as the original image. The loss value is calculated compared with the manually annotated mask. A stochastic gradient algorithm is used to update the model parameters based on the error loss. This algorithm is iterated until the loss value remains essentially unchanged, indicating that the model has converged and training is terminated. The test set data is used to test the model. Cable target segmentation is performed on the test set samples based on the trained model. The model is quantitatively evaluated based on segmentation accuracy evaluation metrics, thereby constructing a submarine cable target detection model based on DeepLabv3+.

[0021] Step S5: Use the submarine cable detection model to segment the cable targets on the waterfall diagram, evaluate the segmentation accuracy of the submarine target detection model based on semantic segmentation, and optimize the model.

[0022] In one embodiment, the trained detection model is deployed on an underwater robotic platform to perform real-time semantic segmentation on the collected seafloor sonar waterfall images. A post-processing module applies length constraints and morphological correction to the slender target regions in the prediction results, ensuring that the segmentation results are consistent with the physical cable shape. Model performance is evaluated using metrics such as mIoU and FWIoU. The model's adaptability and segmentation accuracy are fully verified using a test set of scenes at different depths and terrains.

[0023] In another embodiment, semantic segmentation is often evaluated using the mean intersection over union (mIoU) as an evaluation criterion. A larger mIoU value indicates a higher segmentation accuracy of the model, and its expression is: in, and are label values ​​of different categories, and the label value starts from 0; Indicates belonging to a category But it is predicted to be a category The number of pixels; Indicates the number of correctly classified pixels; is the maximum value of all category labels, That is the number of categories.

[0024] Considering that underwater pipeline target pixels account for a relatively small proportion of the image, if the IoU of each type of target is directly averaged, the prediction results of the pipeline target will have a significant impact on the average IoU indicator. To better evaluate the performance of the model, the experiment uses the Frequency Wighted Intersection over Union (FWIoU) to evaluate the performance of the segmentation model: in, and Labels representing different categories; Indicates belonging class target but is predicted to be The number of pixels of the class; Indicates belonging j class target but is predicted to be The number of pixels of the class; The correct prediction is The number of pixels of the class object; is the maximum category label value; category label values ​​start from 0, so, Indicates the number of all categories. For pipeline target detection experiments, The value is 1. FWIoU weights each type of target according to the proportion of each type of target pixels, and uses this weight to perform a weighted average of the intersection-over-union ratio of each type of target, thus more reasonably evaluating the performance of the model.

[0025] Preferably, step S1 includes the following steps: Step S11: obtaining original image data of the seabed scene using synthetic aperture sonar raw data; Step S12: decoding the original image and generating a waterfall image of the seabed scene; Step S13: identifying the starting response point of the target structure as the first feature point by analyzing the first texture feature and the first spatial grayscale feature map in the waterfall image; Step S14: extracting a second feature point representing a location of structural change by analyzing the second texture feature and the second spatial grayscale feature map in the waterfall image; Step S15: Analyze the structural response difference between the first characteristic point and the second characteristic point.

[0026] In one embodiment, the system collects raw synthetic aperture sonar (SAS) data as input to obtain raw image data of the seafloor scene to be analyzed. The raw image data includes information on how the intensity of acoustic echoes changes over time and space, and exhibits rich texture and grayscale variations. The system decodes the collected raw image data, including preprocessing steps such as amplitude correction, coordinate projection conversion, and image amplitude mapping, to generate a standard-format seafloor waterfall image. This waterfall image clearly presents the linear extension of seafloor target structures and their surroundings in two dimensions. After generating the waterfall image, the system first optimizes the single-strip image. A single-strip image is a continuous imaging strip covering a certain width of seafloor area acquired by a synthetic aperture sonar during a single side-scan scan. Its width is determined by the sonar array layout and operating parameters, and its length varies with the track extension. For this single-strip image, the system removes random noise and water column noise, performs amplitude equalization and radiation correction to eliminate brightness unevenness, and performs geometric distortion correction and attitude compensation to correct the offset caused by platform roll, pitch or yaw, while unifying the spatial resolution and enhancing local contrast to improve detail recognizability. Subsequently, the system performs strip splicing processing on multiple strip images covering adjacent seabed areas, realizes spatial registration based on navigation information and feature matching results, adjusts the brightness and texture consistency of overlapping areas, and adopts a smooth fusion strategy to eliminate splicing seams, generating continuous and seamless large-area synthetic aperture sonar images, providing a complete background for the continuous direction analysis of global linear targets. Based on the image analysis algorithm, the system extracts texture and spatial grayscale features from the waterfall image, constructs the first texture feature map and the first spatial grayscale feature map, and identifies the starting response position of the target structure in the image. This response point usually manifests as texture edge enhancement, grayscale change mutation or linear texture concentration area. The recognition result is used as the first feature point of the target structure. In order to further analyze the change trend of the structure, the system continues to extract the second texture feature map and the second spatial grayscale feature map, and combines the edge enhancement operator with the directional gradient change detection technology to identify abnormal feature positions such as morphological deflection, direction change, and brightness jump in the target structure, and extract them as the second feature point. Finally, based on the image area between the first feature point and the second feature point, the system extracts the pixel connection path, analyzes the difference characteristics in its grayscale change, texture distribution, edge structure response, etc., and obtains the structural response difference result, which provides a basis for subsequent path structure analysis.

[0027] In another embodiment, after the original data is decoded, the real and imaginary data in the original image are squared and then squared. The result obtained by this processing is considered to be the echo data corresponding to each pixel position in the image. This set of echo data is linearly quantized to the interval [0, 255] to obtain image data with pixel values ​​quantized to the 8-bit grayscale image interval. Each scan line containsN Column, stack each scan line vertically in the order of recording M OK, get one M×N The image data matrix is ​​then added to the corresponding image header file to obtain a single-channel grayscale image.

[0028] Preferably, step S13 includes the following steps: Step S131: performing local contrast enhancement and grayscale histogram equalization processing on the waterfall image to generate a single-channel grayscale image reflecting the brightness distribution characteristics as the first spatial grayscale feature map of the seabed image; Step S132: In the first spatial grayscale feature map, a local analysis window is constructed with an arbitrary pixel point as the center, and the distribution trend of each pixel in the local analysis window in terms of texture gradient direction, brightness difference characteristics and regional contrast is evaluated; Step S133: Based on the distribution trend, determine the dominant direction in which the texture direction changes most significantly in the local window as the first texture feature of the pixel point; Step S134: performing a fusion analysis on the first texture feature of the pixel point and its corresponding grayscale response value, and screening out regions with texture mutation and grayscale response extreme value characteristics in the image; Step S135: extracting the response starting point of structural continuity and distribution coherence from the region with texture mutation and grayscale response extreme value characteristics as the first feature point of the target linear structure.

[0029] In one embodiment, to extract the starting response points of submarine linear structures, the waterfall image is preprocessed using an adaptive histogram equalization algorithm (such as CLAHE) to enhance local contrast and suppress background texture interference, resulting in a first spatial grayscale feature map reflecting the brightness distribution characteristics. Within this grayscale map, an M×M sliding window is constructed centered around each pixel to analyze the distribution trends of each pixel within the local region in terms of gradient direction, grayscale change rate, and contrast. Specifically, the gradient direction is extracted using the Sobel operator, the grayscale change rate is obtained by normalizing the grayscale difference by absolute value, and the contrast is calculated based on the local standard deviation. Subsequently, based on these statistical characteristics, the dominant direction with the most significant local texture direction change is determined as the first texture feature of that pixel. Based on this, the first texture feature of each pixel is fused with its corresponding grayscale response value for analysis. A weighted scoring function is then used to highlight regions with strong texture abruptness and high grayscale response. Edge starting regions with good continuity and coherent response are extracted from the fused score map. Feature response points with stable local structural direction and complete morphology are then selected as the first feature points of submarine linear structures.

[0030] Preferably, step S14 includes the following steps: Step S141: adjusting the brightness dynamic range and edge enhancement processing of the waterfall image to enhance the response intensity of the structural inflection points and boundary transition areas in the image, thereby obtaining a second spatial grayscale feature map; Step S142: For any pixel in the image, a local analysis window of fixed size is constructed with the pixel as the center. Responses of the window are calculated using a preset Gabor filter bank in multiple directions and at multiple scales. The maximum value of all responses is extracted as the second texture feature of the pixel. Step S143: performing feature fusion on the second texture feature and the second spatial grayscale feature, using weighted superposition to generate a fusion response map that highlights the degree of local structural change; Step S144: In the fusion response image, analyze the areas with drastic grayscale changes, areas with sudden texture changes, and locations where the structure changes, and extract key pixels with obvious inflection trends or spatial discontinuities; Step S145: Perform response intensity screening and spatial distribution cluster analysis on the extracted key pixel points, and retain the point set with clear structural change trend and continuous regional distribution as the second feature point of the target structure.

[0031] In one embodiment, the system uses an analysis strategy based on grayscale and texture fusion features to identify second feature points of the target structure in regions with structural changes within the waterfall image. These features are used to characterize spatial variations such as local structural inflections, sudden changes in orientation, or disconnected connections. The system then performs brightness dynamic range adjustment and edge enhancement on the original waterfall image to enhance the image response at structural boundaries and transition regions, generating a second spatial grayscale feature map. This step effectively enhances the level of image detail by stretching the image histogram and introducing the Sobel edge enhancement operator. The system constructs a fixed-size local analysis window (e.g., 11×11 or 15×15) centered on any pixel in the image. For each window, convolution responses are calculated using a Gabor filter bank configured with multiple orientations (e.g., 0°, 45°, 90°, and 135°) and multiple scale parameters. The result with the maximum response amplitude is extracted as the second texture feature for that pixel. The system then fuses the second texture feature with the second spatial grayscale feature obtained in step S141, constructing a fused response map using a weighted superposition approach or a joint response scoring strategy. This fused image can simultaneously reflect local texture changes and brightness structural responses, enhancing sensitivity to inflection areas. Within the fused response image, the system further analyzes areas of dramatic grayscale change, areas of sudden texture direction changes, and locations of structural orientation jumps, extracting key pixels using a response threshold method. These points are typically concentrated in locations within the image where there are significant directional shifts or sudden morphological boundary changes. The system then screens the extracted key pixels for response strength, removing weakly responsive points. Density-based spatial clustering analysis methods, such as the DBSCAN algorithm, are then used to remove isolated points and small clusters, retaining only pixel sets with clear structural change trends and continuous spatial distribution. The resulting set of points serves as the second feature point of the target structure.

[0032] In another embodiment, the system processes a 768×512-resolution submarine synthetic aperture sonar image through histogram equalization and Laplacian enhancement to generate a second spatial grayscale feature map. Cable structure boundaries and inflection regions within the image are significantly enhanced. The image is then subjected to local sliding window processing, with each 11×11 window using a set of Gabor filters with four orientations and three scales to extract texture responses. The system retains the maximum magnitude response for each pixel to form a second texture feature map. This map exhibits high texture response energy in areas of structural morphological variation. A weighted fusion of the second grayscale feature map and the second texture feature map (in a ratio of 0.6:0.4) is performed to generate a fused response map. Within this map, the system uses threshold segmentation to extract grayscale and texture direction transition points, preliminarily identifying a number of suspected structural change points. The system uses response strength sorting to eliminate weak response points and performs spatial cluster analysis on the remaining point set, retaining points with consistent structural trends and good continuity as the second feature points of the target structure.

[0033] Preferably, step S15 includes the following steps: Step S151: Taking the first feature point as the starting point, obtaining the second feature point in the vicinity of the first feature point in the image, and extracting the pixel connection path between the two points; Step S152: extracting a grayscale value sequence of consecutive pixels along the connection path; calculating a grayscale response gradient between adjacent pixels in the grayscale value sequence; Step S153: Calculating the texture fluctuation trend in the texture direction based on the texture direction information of each pixel in the connection path; Step S154: combining the grayscale response gradient and texture fluctuation trend to calculate the structural connectivity parameters of the path area; Step S155: Score the degree of structural difference of the pixel connection path according to the grayscale response gradient, texture fluctuation trend and structural connectivity parameter to generate structural response difference.

[0034] In one embodiment, to effectively identify local connectivity features of submarine linear structures, the system analyzes the grayscale and texture-level response differences of the path structure based on the pixel connection paths between feature points. This generates structural response difference results for subsequent path structure modeling and stability analysis. Specifically, if a first feature point has been identified in an image, the system uses this first feature point as a starting point and searches for a second feature point in its vicinity. Based on the spatial adjacency between image pixels, the system extracts the pixel connection path connecting the first and second feature points. This path can be connected pixel-wise using the minimum grayscale gradient method or dynamic programming to ensure path continuity and stable response. The system then sequentially extracts the grayscale value sequence of all consecutive pixels along the extracted connection path and calculates the grayscale response gradient between adjacent pixel pairs in this grayscale value sequence. This gradient reflects the rate of grayscale change in the structural region and is an important basis for determining image structural boundaries. Furthermore, based on the texture direction information of each pixel in the connection path, the system extracts and analyzes its changing trend along the path direction, constructing a texture direction fluctuation curve. This step usually uses a main direction filter to estimate the local main direction of the pixel window, forming a texture direction sequence, and performing a variation amplitude analysis on the sequence. Based on the grayscale response gradient and texture fluctuation trend, the system performs a statistical analysis on the structural connectivity of the connection path area and extracts structural connectivity parameters that reflect the path continuity and morphological consistency. These parameters may include grayscale consistency metrics, texture direction coherence indicators, and path continuity curvature. Finally, the system comprehensively considers the grayscale response gradient, texture direction fluctuation trend, and structural connectivity parameters to score the proposed connection path and generate a structural response difference value to represent the degree of change between the structural areas connected by the path. This response difference value can be used as an important indicator for subsequent judgment of cable direction stability and structural mutation points.

[0035] In another embodiment, a synthetic aperture sonar waterfall image is processed, with the starting point of a known target cable segment in the image as the first feature point. The system sets a neighborhood window with a radius of 20 pixels around this feature point. Within this area, texture jump detection and spatial gradient analysis are used to locate a second feature point with a structural abrupt change. Next, the system uses an image edge gradient minimum energy path algorithm to obtain a pixel path connecting the two feature points. For each pixel on this path, the system extracts a grayscale value sequence and calculates the grayscale response gradient between every two adjacent pixels using a difference method to form a complete gradient sequence. Furthermore, a local directional consistency estimation algorithm is used to extract the main texture direction of each pixel on the path. The directional sequence is then differentiated to obtain a texture direction variation curve. The degree of texture fluctuation is quantitatively analyzed by setting a threshold. Structural connectivity parameters such as grayscale difference stability, directional continuity, and pixel coherence between pixels in the path are calculated to assess the consistency of the overall path morphology. A multidimensional scoring model is constructed based on the three aforementioned indicators to comprehensively assess the degree of structural response variation in the connecting path. If the score exceeds the set threshold, it is determined that the two regions connected by the path have significant structural changes, and the path is marked as a structural mutation path, which will be used for subsequent inflection detection and direction division analysis.

[0036] Preferably, step S2 includes the following steps: Step S21: Calculating the linear rotation angle and curvature change trend of the connection path based on the structural response difference; Step S22: using the linear rotation angle and curvature change trend, the target area is divided into a continuous segment with stable directional extension and a turning segment with morphological inflection; Step S23: constructing a spatial distribution change map of the linear target with reference to the central skeleton lines of the continuous segment and the turning segment; Step S24: using the spatial distribution change map in combination with submarine line tracking, extracting the continuous direction of the submarine cables; Step S25: Track the displacement rate and local connectivity of the structure in the extension direction through the continuous direction of the submarine cable, and mark the stable segments and abnormal variation segments of the target structure in the image.

[0037] In one embodiment, based on the structural response differences obtained in step S1, corresponding pixel connection paths are extracted. For each path, the linear rotation angle (i.e., the angle between the direction of the line connecting the path's starting and ending points and the image reference axis (e.g., the X-axis)) is calculated to obtain a directional change index. Simultaneously, the overall curvature trend of the path is calculated based on the gradient changes and directional offsets of the pixels along the path to determine whether the path exhibits inflections or sharp turns. Based on these two indices, the target area is further divided into continuous segments with stable directional extension (e.g., straight segments) and transitional segments with morphological inflections (e.g., bends). The central skeleton line of each segment is extracted and used as the principal axis reference for the linear target. A spatial distribution change map is constructed, containing information such as directional change, curvature fluctuations, and connectivity strength. Subsequently, combined with the cable tracking task in seabed imagery, the spatial distribution change map is used to extract the continuous trajectory of the submarine cable. By analyzing the deviation rate of the extension direction in the trajectory (i.e., the amplitude of the direction change within a unit length) and the structural connectivity between the path segments, the stable segments (direction consistency, gentle curvature) and abnormal variation segments (direction jump, drastic curvature change) of the target cable are automatically marked.

[0038] In another embodiment, the structural response difference path in a 512×512 seabed image is analyzed, and it is calculated that the linear rotation angle range is 5° to 42°, and there are 3 inflection points in the curvature change trend curve. According to the set curvature threshold, the path is divided into 2 continuous segments and 1 turning segment. The overall linear direction is reconstructed by interpolation of the central skeleton line of each segment, and a spatial distribution change map is generated. Further analysis shows that the main axis direction of the path has an offset rate peak at the 20th pixel, the connectivity is reduced, and it is determined to be an "abnormal variation segment"; the rest is a "stable segment". This segmentation information can be directly used for automatic inspection and path reconstruction tasks of submarine cables or communication pipelines. Preferably, step S21 includes the following steps: Step S211: extracting a response path between the first feature point and the second feature point based on the structural response difference; Step S212: extracting a centerline sequence of the path according to the spatial coordinates of each pixel point along the response path; Step S213: Based on the centerline sequence, calculate the direction vector between any two consecutive pixel points in the path; combine all direction vectors into a path direction sequence; Step S214: Calculate the angle change between the direction vectors segment by segment based on the path direction sequence to obtain the linear rotation angle of the path; Step S215: Calculate the curvature change rate based on the change in the linear rotation angle of each small segment in the path to form a curvature change trend curve of the path.

[0039] In one embodiment, based on the structural response difference, the connection path between the first feature point and the second feature point is extracted, and the spatial coordinates of all pixel points in the path are obtained. After smoothing the path pixel point sequence, its center line sequence is extracted to represent the main direction of the overall response path. Subsequently, a direction vector is constructed between any two consecutive pixel points in the center line sequence, and the path direction sequence is obtained by combining them in sequence. According to the change in the angle between adjacent direction vectors, the linear rotation angle of the path is calculated segment by segment to characterize the directional extensibility of the path. Further, based on the change in the linear rotation angle of each segment, the curvature change rate of the path is derived, and a curvature change trend curve is drawn to characterize the bending degree and directional stability of the path in space.

[0040] In another example, after extracting the centerline sequence of a response path consisting of 40 pixels, 39 directional vectors were generated. The calculated directional angle ranged from 2.3° to 19.7°, with an average linear rotation angle of 7.6°. Further differential calculation of the angle variation within each segment yielded a curvature rate curve, which showed a clear peak between the 12th and 18th pixels, indicating a morphological transition characteristic within this segment. This trend chart can be used to assist in identifying bends or sections at risk of breakage in submarine cables.

[0041] Preferably, step S22 includes the following steps: Step S221: Based on the linear rotation angle and curvature change trend curve, the response path is divided into a number of adjacent sub-segments in pixel order; Step S222: Calculate the average rotation angle and curvature change of each sub-segment and compare them with a preset stability threshold. If both the linear rotation angle and curvature value are lower than the threshold, mark the sub-segment as a continuous segment with stable directional extension. Step S223: If the linear rotation angle or curvature change rate in a sub-segment is higher than a threshold, the sub-segment is marked as a turning segment with a morphological inflection; Step S224: Perform continuity check on all sub-segment marking results, remove abnormal segments that are too short and isolated, and retain main path segments with continuous structural patterns and obvious change trends; Step S225: Output the divided continuous segments with stable directional extension and the turning segments with morphological inflection.

[0042] In one embodiment, for the linear rotation angle and curvature change trend curve of the aforementioned path, the response path is divided into a number of adjacent sub-segments of equal length (e.g., each segment contains 5 pixels) in pixel order. For each sub-segment, the average linear rotation angle and the average curvature change rate are calculated and compared with a set stability threshold (e.g., the angle threshold θ t =10°, curvature threshold κ t= 0.1) for comparison. If both of the above indicators in a subsegment are below the corresponding threshold, the subsegment is marked as a "continuous segment with stable directional extension"; conversely, if either indicator exceeds the threshold, it is marked as a "turning segment with morphological inflection." To ensure the continuity and effectiveness of the path segmentation marking, the preliminary segmentation results are post-processed to remove isolated segments less than 3 pixels in length and merge consecutive similar segments to enhance structural coherence. The resulting path structure segmentation results include both continuous segments and turning segments.

[0043] In another embodiment, a response path with a total length of 45 pixels was divided into nine subsegments, each subsegment being 5 pixels long. Calculations revealed that the average linear rotation angles of segments 1–3 and 7–9 were 4.2°, 6.1°, and 5.7°, respectively, with an average curvature change rate less than 0.08, meeting the stability criteria and therefore marking them as continuous segments with stable directional extension. In segments 4–6, the average angles were 12.9°, 14.6°, and 11.3°, indicating significant curvature fluctuations, marking them as turning segments with morphological inflection. Further, segment 6, consisting of only an isolated 3-pixel irregular segment, was removed and merged into segment 5, identifying three continuous segments and two major turning segments. This structural division provides a clear morphological demarcation basis for subsequent path skeleton fitting.

[0044] See also Figure 5 , is the submarine cable target image set; Preferably, step S24 includes the following steps: Step S241: Based on the skeleton line structure extracted from the spatial distribution change map, the strike direction vectors of the continuous segment and the turning segment in the image coordinate system are calculated, and the relative direction information along the skeleton line is obtained; Step S242: performing dynamic tracking processing on the extension path of the cable based on the direction vector according to the relative direction information of the skeleton line in the spatial distribution map; Step S243: During the tracking process, detecting the interference area blocked by the water column in the image, and identifying the cable track nodes on both sides of the blocked area respectively; Step S244: using a trajectory fitting algorithm to perform interpolation prediction and geometric fitting on the blocked interference area, and constructing a structural extension direction completion result across the blocked area; Step S245: constructing an inter-frame structural connectivity graph based on the structure extension direction completion result and the extension path of the cable; Step S246: The spatial cable path results extracted from the inter-frame structural connectivity graph are used to establish a global inter-frame connectivity graph, integrate the direction information of multiple sections of submarine cables, and extract the continuous direction of the submarine cables.

[0045] In one embodiment, based on the skeleton line structure extracted from the aforementioned spatial distribution change map, the system obtains a sequence of central skeleton line coordinates for each continuous segment and turning point. For each skeleton line segment in the image coordinate system, the system calculates its direction vector and further integrates the overall skeleton line extension direction information to form a directional sequence for the target structure. Based on this, the system uses the direction vectors to dynamically track the submarine cable target along the skeleton line direction. During this process, when significant occlusion caused by water column interference or water reflection is detected in the image area, the system extracts the cable trajectory nodes at the front and rear ends of the occluded segment. Using a curve-fitting trajectory prediction model (such as spline interpolation or least squares fitting), the path of the occluded segment is completed to construct a complete structural extension direction. Subsequently, the cable path across multiple frames is integrated, combining the completed direction data with the image frame sequence to construct an inter-frame structural connectivity graph. This global inter-frame connectivity graph structure is then formed within the global frame set, enabling accurate extraction of the continuous submarine cable path.

[0046] In another example, for a cable inspection task consisting of 10 frames of underwater imagery, the system extracts the skeleton line structure in each frame and calculates its local direction vector. In the fourth frame, a significant area of ​​water column occlusion is detected. The system identifies the corresponding skeleton line endpoints before and after the occlusion and uses cubic spline interpolation to complete path fitting and complete the missing cable segment. The system then aligns the skeleton line predicted during the completion process with the actual skeleton line structure in adjacent frames in an inter-frame connectivity graph, ensuring direction continuity and spatial geometric consistency. By integrating the path information extracted from all image frames, a global structural connectivity graph covering all image frames is constructed, enabling continuous and complete identification of the submarine cable direction in the area.

[0047] See also Figure 6 , is the pipeline image detection result; Preferably, step S25 includes the following steps: Step S251: obtaining the extension direction of the submarine cable based on the continuous direction of the submarine cable; Step S252: In each frame of the image, the geometric center point offset rate between adjacent frames is calculated along the extension direction of the submarine cable, and the time series of the offset rate is smoothed to obtain a structural extension trend curve; Step S253: Based on the skeleton connectivity of the submarine cable in the spatial distribution change diagram, the path continuity metric of the cable in the local window is calculated, including the coherent length of the skeleton line segment, the break frequency and the turning angle change amplitude; Step S254: integrating the structural extension trend curve and the local path connectivity metric to set a combined threshold standard for the offset rate and connectivity; Step S255: When the target cable area satisfies the requirement of a stable change in the offset rate and a connectivity index higher than the set threshold, it is marked as a stable segment. When the target cable area has a sudden change in the offset rate or a significant decrease in connectivity, it is marked as an abnormal variation segment, and the spatial annotation results of the stable segment and the abnormal variation segment are output in the image.

[0048] In one embodiment, based on the extracted continuous trajectory of the submarine cable, the system further spatially models its extension direction. Specifically, the system extracts the submarine cable's extension direction vector based on the spatial coordinate sequence of the continuous skeleton line and uses this direction as the reference direction. Within the image sequence, the system tracks the submarine cable's geometric center point in each frame along this extension direction and calculates the offset distance between its center points between adjacent frames to obtain the inter-frame offset rate. To remove the effects of jitter and minor errors, the offset rate sequence is processed using a time-domain smoothing filter (e.g., sliding average or Gaussian filtering) to generate a structural extension trend curve. Simultaneously, based on the skeleton structure in the spatial distribution change map, the system calculates connectivity metrics such as the coherent length of skeleton line segments, break frequency, and the magnitude of change in turning angle within a local sliding window (step S253). Combined with the extracted extension trend curve, a joint judgment criterion for excursion rate and connectivity evaluation is established to assess structural stability. If the cable excursion rate within the target area changes steadily and the local connectivity index is above the threshold, the area is considered a stable structural extension segment. If there is a sudden change in excursion rate, an increase in skeleton fractures, or a sharp change in turning angle, it is marked as an abnormal variation segment. The system outputs the spatial annotation results of the corresponding stable and abnormal segments in the image for subsequent analysis or processing.

[0049] In another example, the system identified the continuous skeleton structure of a main cable from a 30-frame seafloor image sequence and extracted its global extension direction. The system then extracted the cable centerline position in each frame and calculated the centerpoint offset rate between frames, generating a 29-value offset rate sequence. To enhance robustness, the system smoothed this sequence using a five-point sliding average algorithm to capture the structural extension trend.

[0050] Next, the system selects a local 5×5 pixel window and evaluates the continuity of the skeleton segments corresponding to each location point, counting the number of breaks and the amplitude of directional fluctuations in consecutive frames. By setting the offset rate change amplitude ≤ 2 pixels / frame and the skeleton continuity length ≥ 20 pixels as the joint threshold standard, the system marks the structure between frames 8 and 18 as a stable segment. Frames 24 to 27 are marked as abnormal variation segments due to frequent skeleton breaks and drastic changes in direction. These spatial regions are annotated and displayed in different colors in the image.

[0051] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing an adaptive semantic segmentation submarine linear target detection model, characterized in that: The following steps are involved: Step S1: acquiring original image data of a seabed scene and performing feature response processing on the data, wherein the image is analyzed and the starting response point of the target structure is identified as a first feature point; a second feature point representing the location of the structural change is extracted; and the difference in structural response between the first feature point and the second feature point is analyzed; Step S2: Based on the structural response differences, the linear rotation angle and curvature change trend of the connection path are calculated, and a spatial distribution change map of the linear target is constructed. By tracking the offset rate and local connectivity of the structural extension direction, combined with seabed line tracking, the continuous direction of the submarine cable is extracted, and the stable and abnormal variation segments of the submarine cable in the image are marked; Step S3: Design a deep neural network with a multi-level convolutional structure and a recurrent memory structure. Use the convolutional network to capture multi-scale edge trends and linear trends in stable segments, and use the recurrent network to extract and detect structural repair features and contextual sequence features in incoherent regions. Step S4: Fusing the comprehensive features of the extracted detection coherent areas and detection incoherent areas, identifying the seabed target sample set, the model input is the seabed target sample set, and a seabed target detection model based on DeepLabv3+ is constructed; Step S5: Use the submarine cable detection model to segment the cable targets on the waterfall diagram, evaluate the segmentation accuracy of the submarine target detection model based on semantic segmentation, and optimize the model.

2. The method for constructing an adaptive semantic segmentation submarine linear target detection model according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: obtaining original image data of the seabed scene using synthetic aperture sonar raw data; Step S12: decoding the original image and generating a waterfall image of the seabed scene; Step S13: identifying the starting response point of the target structure as the first feature point by analyzing the first texture feature and the first spatial grayscale feature map in the waterfall image; Step S14: extracting a second feature point representing a location of structural change by analyzing the second texture feature and the second spatial grayscale feature map in the waterfall image; Step S15: Analyze the structural response difference between the first characteristic point and the second characteristic point.

3. The method for constructing an adaptive semantic segmentation submarine linear target detection model according to claim 2, wherein: Step S13 includes the following steps: Step S131: performing local contrast enhancement and grayscale histogram equalization processing on the waterfall image to generate a single-channel grayscale image reflecting the brightness distribution characteristics as the first spatial grayscale feature map of the seabed image; Step S132: In the first spatial grayscale feature map, a local analysis window is constructed with an arbitrary pixel point as the center, and the distribution trend of each pixel in the local analysis window in terms of texture gradient direction, brightness difference characteristics and regional contrast is evaluated; Step S133: Based on the distribution trend, determine the dominant direction in which the texture direction changes most significantly in the local window as the first texture feature of the pixel point; Step S134: performing a fusion analysis on the first texture feature of the pixel point and its corresponding grayscale response value, and screening out regions with texture mutation and grayscale response extreme value characteristics in the image; Step S135: extracting the response starting point of structural continuity and distribution coherence from the region with texture mutation and grayscale response extreme value characteristics as the first feature point of the target linear structure.

4. The method for constructing an adaptive semantic segmentation submarine linear target detection model according to claim 2, wherein: Step S14 includes the following steps: Step S141: adjusting the brightness dynamic range and edge enhancement processing of the waterfall image to enhance the response intensity of the structural inflection points and boundary transition areas in the image, thereby obtaining a second spatial grayscale feature map; Step S142: For any pixel in the image, a local analysis window of fixed size is constructed with the pixel as the center. Responses of the window are calculated using a preset Gabor filter bank in multiple directions and at multiple scales. The maximum value of all responses is extracted as the second texture feature of the pixel. Step S143: performing feature fusion on the second texture feature and the second spatial grayscale feature, using weighted superposition to generate a fusion response map that highlights the degree of local structural change; Step S144: In the fusion response image, analyze the areas with drastic grayscale changes, areas with sudden texture changes, and locations where the structure changes, and extract key pixels with obvious inflection trends or spatial discontinuities; Step S145: Perform response intensity screening and spatial distribution cluster analysis on the extracted key pixel points, and retain the point set with clear structural change trend and continuous regional distribution as the second feature point of the target structure.

5. The method for constructing an adaptive semantic segmentation submarine linear target detection model according to claim 2, wherein: Step S15 includes the following steps: Step S151: Taking the first feature point as the starting point, obtaining the second feature point in the vicinity of the first feature point in the image, and extracting the pixel connection path between the two points; Step S152: extracting a grayscale value sequence of consecutive pixels along the connection path; calculating a grayscale response gradient between adjacent pixels in the grayscale value sequence; Step S153: Calculating the texture fluctuation trend in the texture direction based on the texture direction information of each pixel in the connection path; Step S154: combining the grayscale response gradient and texture fluctuation trend to calculate the structural connectivity parameters of the path area; Step S155: Score the degree of structural difference of the pixel connection path according to the grayscale response gradient, texture fluctuation trend and structural connectivity parameter to generate structural response difference.

6. The method for constructing an adaptive semantic segmentation submarine linear target detection model according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: Calculating the linear rotation angle and curvature change trend of the connection path based on the structural response difference; Step S22: using the linear rotation angle and curvature change trend, the target area is divided into a continuous segment with stable directional extension and a turning segment with morphological inflection; Step S23: constructing a spatial distribution change map of the linear target with reference to the central skeleton lines of the continuous segment and the turning segment; Step S24: using the spatial distribution change map in combination with submarine line tracking, extracting the continuous direction of the submarine cables; Step S25: Track the displacement rate and local connectivity of the structure in the extension direction through the continuous direction of the submarine cable, and mark the stable segments and abnormal variation segments of the target structure in the image.

7. The method for constructing an adaptive semantic segmentation submarine linear target detection model according to claim 6, characterized in that: Step S21 includes the following steps: Step S211: extracting a response path between the first feature point and the second feature point based on the structural response difference; Step S212: extracting a centerline sequence of the path according to the spatial coordinates of each pixel point along the response path; Step S213: Based on the centerline sequence, calculate the direction vector between any two consecutive pixel points in the path; combine all direction vectors into a path direction sequence; Step S214: Calculate the angle change between the direction vectors segment by segment based on the path direction sequence to obtain the linear rotation angle of the path; Step S215: Calculate the curvature change rate based on the change in the linear rotation angle of each small segment in the path to form a curvature change trend curve of the path.

8. The method for constructing an adaptive semantic segmentation submarine linear target detection model according to claim 6, wherein: Step S22 includes the following steps: Step S221: Based on the linear rotation angle and curvature change trend curve, the response path is divided into a number of adjacent sub-segments in pixel order; Step S222: Calculate the average rotation angle and curvature change of each sub-segment and compare them with a preset stability threshold. If both the linear rotation angle and curvature value are lower than the threshold, mark the sub-segment as a continuous segment with stable directional extension. Step S223: If the linear rotation angle or curvature change rate in a sub-segment is higher than a threshold, the sub-segment is marked as a turning segment with a morphological inflection; Step S224: Perform continuity check on all sub-segment marking results, remove abnormal segments that are too short and isolated, and retain main path segments with continuous structural patterns and obvious change trends; Step S225: Output the divided continuous segments with stable directional extension and the turning segments with morphological inflection.

9. The method for constructing an adaptive semantic segmentation submarine linear target detection model according to claim 6, characterized in that: Step S24 includes the following steps: Step S241: Based on the skeleton line structure extracted from the spatial distribution change map, the strike direction vectors of the continuous segment and the turning segment in the image coordinate system are calculated, and the relative direction information along the skeleton line is obtained; Step S242: performing dynamic tracking processing on the extension path of the cable based on the direction vector according to the relative direction information of the skeleton line in the spatial distribution map; Step S243: During the tracking process, detecting the interference area blocked by the water column in the image, and identifying the cable track nodes on both sides of the blocked area respectively; Step S244: using a trajectory fitting algorithm to perform interpolation prediction and geometric fitting on the blocked interference area, and constructing a structural extension direction completion result across the blocked area; Step S245: constructing an inter-frame structural connectivity graph based on the structure extension direction completion result and the extension path of the cable; Step S246: The spatial cable path results extracted from the inter-frame structural connectivity graph are used to establish a global inter-frame connectivity graph, integrate the direction information of multiple sections of submarine cables, and extract the continuous direction of the submarine cables.

10. The method for constructing an adaptive semantic segmentation submarine linear target detection model according to claim 6, characterized in that: Step S25 includes the following steps: Step S251: obtaining the extension direction of the submarine cable based on the continuous direction of the submarine cable; Step S252: In each frame of the image, the geometric center point offset rate between adjacent frames is calculated along the extension direction of the submarine cable, and the time series of the offset rate is smoothed to obtain a structural extension trend curve; Step S253: Based on the skeleton connectivity of the submarine cable in the spatial distribution change diagram, the path continuity metric of the cable in the local window is calculated, including the coherent length of the skeleton line segment, the break frequency and the turning angle change amplitude; Step S254: integrating the structural extension trend curve and the local path connectivity metric to set a combined threshold standard for the offset rate and connectivity; Step S255: When the target cable area satisfies the requirement of a stable change in the offset rate and a connectivity index higher than the set threshold, it is marked as a stable segment. When the target cable area has a sudden change in the offset rate or a significant decrease in connectivity, it is marked as an abnormal variation segment, and the spatial annotation results of the stable segment and the abnormal variation segment are output in the image.

Images