A method for constructing an adaptive semantic segmentation model for detecting linear targets on the seabed

Abstract

The present application relates to the technical field of seabed linear target detection model construction, and particularly relates to a method for constructing an adaptive semantic segmentation seabed linear target detection model. The method comprises the following steps: obtaining seabed scene original image data, extracting first feature points and second feature points, and analyzing the structural response difference therebetween; based on the structural response difference, calculating the linear rotation angle and curvature change trend of the connecting path, constructing a spatial distribution change graph, extracting the cable continuous trend, and labeling the stable section and abnormal variation section; designing a deep neural network comprising a multi-level convolution structure and a recurrent memory structure, extracting the structure extension and repair features; fusing the deep features of the coherent region and the incoherent region, constructing a seabed target detection model based on DeepLabv3+; performing semantic segmentation on the waterfall graph, and evaluating the segmentation accuracy of the model to achieve optimization; the present application realizes more accurate and efficient target detection through the construction of the target detection model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of seabed linear target detection model construction, and particularly relates to a construction method of a self-adaptive semantic segmentation seabed linear target detection model. BACKGROUND

[0002] Traditional sonar image target detection mainly uses a method of manually extracting features and then classifying by a classifier. Sonar image target detection based on deep learning usually focuses on the differences between optical images and sonar images, small data sets, and high noise characteristics of sonar images. The demand for automatic recognition and monitoring of linear structures such as seabed cables and pipelines is growing. Synthetic aperture sonar (SAS) has been widely used in seabed target detection tasks as a high-resolution underwater imaging means. However, due to the complex and changeable seabed environment, seabed images are often affected by water column occlusion, terrain undulation, sound wave interference, and structure morphology changes, resulting in discontinuous, unclear, and sudden changes in the direction of linear targets in the image, which seriously affects the accuracy and stability of traditional image processing and target detection algorithms. Existing methods rely on fixed feature templates or simple edge detection algorithms, which are difficult to adapt to the multi-scale, weak texture, and incoherent characteristics of cable structures in different seabed scenes, and lack robustness to local occlusion and structure inflection, making it difficult to achieve continuous tracking and accurate segmentation of linear targets. SUMMARY

[0003] Therefore, it is necessary to provide a construction method of a self-adaptive semantic segmentation seabed linear target detection model to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a construction method of a self-adaptive semantic segmentation seabed linear target detection model includes the following steps:

[0005] Step S1: Obtain the original image data of the seabed scene and perform feature response processing, wherein the starting response point of the target structure is identified as the first feature point by analyzing the image; the second feature point representing the position of the structure change is extracted; and the structure response difference between the first feature point and the second feature point is analyzed.

[0006] Step S2: Calculate the linear rotation angle and curvature change trend of the connection path based on the structure response difference, and construct a spatial distribution change graph of the linear target. By tracking the offset rate and local connectivity of the structure extension direction, and combining seabed line tracking, the continuous direction of the seabed cable is extracted, and the stable section and abnormal variation section of the seabed cable in the image are labeled.

[0007] Step S3: a deep neural network containing a multi-level convolution structure and a recurrent memory structure is designed, the convolution network is used to capture multi-scale edge trends and linear trends in the stable section, and the recurrent network is used to extract structure repair features and context sequence features in the detected incoherent area;

[0008] Step S4: the comprehensive features of the detected coherent area and the detected incoherent area are fused, the seabed target sample set is identified, the model input is the seabed target sample set, and a seabed target detection model based on DeepLabv3+ is constructed;

[0009] Step S5: the obtained waterfall graph is subjected to cable target segmentation by using the seabed cable detection model, the segmentation accuracy of the seabed target detection model is evaluated based on semantic segmentation, and the model is optimized.

[0010] The present application has the following beneficial effects:

[0011] First aspect: by introducing the extraction mechanism of the first feature point and the second feature point in the seabed scene, and analyzing the structural response difference therebetween, the problems of weak texture, discontinuity and edge blur of the linear target in the seabed image are effectively solved, the limitations of the traditional method in accurately identifying the start and end positions of the cable in the low contrast and non-structured area are made up, and more physically meaningful initial feature basis is provided for subsequent path construction and connectivity analysis.

[0012] Second aspect: by guiding the extraction of the path linear rotation angle and the curvature change trend through the structural response difference, the spatial distribution change graph of the seabed linear target is constructed, the offset rate of the cable extension direction and the local connectivity are combined, the continuous trend features of the target structure in space can be effectively captured, and the automatic identification ability of the cable stable section and the abnormal section in the complex seabed environment is significantly improved, thereby providing a reliable basis for the state evaluation and abnormal warning of the seabed target.

[0013] Third aspect: the deep neural network model introduced in the present application fuses the multi-level convolution structure and the recurrent memory structure, the convolution module can fully capture the multi-scale edge and continuity features in the cable stable section, the recurrent network extracts structure repair and context association information for the areas with interruption, shielding and mutation in the image, thereby improving the structure recognition and recovery ability under the condition of incomplete image, and enhancing the robustness and adaptability of the model.

[0014] Fourth aspect: by fusing the comprehensive features of the coherent area and the incoherent area, and constructing the seabed semantic segmentation model based on DeepLabv3+, not only the target generalization ability of the model in different seabed scenes is improved, but also the cable segmentation accuracy is significantly improved, especially under the interference conditions such as water column shielding, texture aliasing and complex background, the high accuracy can still be maintained, and the complete reconstruction of the target structure in the image and the expression of the spatial coherence are ensured.

[0015] The fifth aspect: through the model, the submarine cable target in the waterfall chart is segmented with high precision, the model performance is evaluated and optimized in combination with the semantic segmentation index, an end-to-end automatic processing flow from image preprocessing, feature extraction, spatial modeling to model segmentation is realized, technical support is provided for underwater engineering inspection, automatic laying monitoring and submarine communication maintenance scenes, and good application and popularization value is obtained. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 It is a step flowchart of a method for constructing an adaptive semantic segmentation submarine linear target detection model.

[0017] Figure 2 It is a DeepLabv3+ network structure schematic diagram.

[0018] Figure 3 It is a submarine cable detection model construction flowchart.

[0019] Figure 4 It is a synthetic aperture sonar data submarine linear target detection flowchart.

[0020] Figure 5 It is a submarine cable target image set, wherein (a) is a pipeline image; (b) is a prediction result; (c) is a manual annotation result.

[0021] Figure 6 It is a pipeline image detection result, wherein (a) is an original sonar image; (b) is a sonar image superimposed with a prediction result.

[0022] The implementation of the object, functional characteristics and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0023] The technical method of the present application will be described in detail below in combination with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0024] In addition, the accompanying drawings are only schematic drawings of the present application, and are not necessarily drawn to scale. The same reference signs in the drawings represent the same or similar parts, and thus repeated descriptions thereof will be omitted. Some of the block diagrams shown in the drawings are functional entities, which do not necessarily have to correspond to physically or logically independent entities. The functional entities can be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0025] It should be understood that, although the terms "first", "second" or the like can be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, a first element could be termed a second element, and, similarly, a second element could be termed a first element without departing from the scope of the example embodiments. As used herein, the term "and / or" includes any and all combinations of one or more of the associated associated items.

[0026] To achieve the above object, please refer to Figures 1 to 6 A method for constructing an adaptive semantic segmentation seabed linear target detection model, comprising the following steps:

[0027] Step S1: Obtain the original image data of the seabed scene, and perform feature response processing thereon, wherein the starting response point of the target structure is identified by analyzing the image and is taken as a first feature point; a second feature point representing the position of the structure change is extracted; and the structure response difference between the first feature point and the second feature point is analyzed.

[0028] In an embodiment, a synthetic aperture sonar is used to collect seabed image data sets in a certain sea area, and the original image resolution is 2048x1024. The image is preliminarily processed by a preset feature response algorithm, and a candidate linear structure is extracted by combining edge enhancement and straight line detection. Among them, combined with Hessian matrix and directional gradient information, the significant response area in the image is detected, and the starting point thereof is taken as the first feature point. The inflection point with large local response gradient change on the linear extension path thereof is searched and marked as the second feature point. Then, the structure connection path between the first feature point and the second feature point is constructed, the local directional change amount on the path is counted, and the structure response difference distribution is obtained.

[0029] In another embodiment, the synthetic aperture sonar raw data is a binary stored file, and the file content mainly includes: file header, option data, data type and offset, imaging parameters, image data packet, etc. The file header is used to indicate the overall information of the file, and the useful information for image processing includes left and right side type, high and low frequency type, image width, real and imaginary part arrangement. The option data is a user-defined part, which is an optional item. If there is no option data, the data option length of the file header should be 0. The data type and offset are used to indicate the offset of each part in the image data packet. The type 512-1023 is used for imgpc file. The imaging parameters are the parameters required for synthetic aperture sonar imaging, mainly including center frequency, bandwidth, pulse width, sampling rate, pulse repetition period, sampling point number, channel number, transmitting array length, receiving array length, range resolution, azimuth resolution, reference speed, image width, etc. The image data packet includes image data and image additional data, and one image data includes one row of image. The image data is stored by row, and the data is real image or complex image, and the storage mode is determined by the real and imaginary part arrangement mode of the file header. The current data is stored in the form of real and imaginary part interleaving arrangement. The image additional data mainly includes the sensor parameters of each row of image, including attitude information such as heading, pitch, roll and heave, cable length, GAPS ultra-short baseline underwater positioning information and GNSS position information, etc. After reading the binary data, it is assigned to a self-defined data structure and saved to a txt format text file for subsequent processing and calling.

[0030] Step S2: based on the structural response difference, the linear rotation angle of the connection path and the curvature change trend are calculated, and the spatial distribution change graph of the linear target is constructed, the offset rate of the structure extension direction and the local connectivity degree are tracked, combined with the submarine line tracking, the continuous trend of the submarine cable is extracted, and the stable section and the abnormal variation section of the submarine cable in the image are marked;

[0031] In an embodiment, based on the structural response difference, the linear rotation angle of the local connection path and the curvature change trend are calculated by using the piecewise fitting method. The spatial extension skeleton line of the linear target in the image is obtained by calculating the direction gradient accumulation pixel by pixel. Further, the spatial distribution change graph reflecting the change trend of the linear structure in the image plane is constructed with the skeleton line as the center. Combined with the offset rate of the structure extension direction and the local connectivity degree evaluation, the submarine cable tracking module is called to track the potential cable trend path. For the identified continuous trend, it is marked as a stable section in the graph; and when there is a curvature mutation or a fracture sign, it is identified as an abnormal variation section.

[0032] Step S3: design a deep neural network containing multi-level convolution structure and recurrent memory structure, use convolution network to capture multi-scale edge trend and linear trend in stable section, use recurrent network to extract structure repair features and context sequence features in discontinuous area;

[0033] In an embodiment, a deep neural network structure is designed, the front end is composed of three convolutional coding modules, respectively using 3x3 convolution kernel to extract multi-scale edge features and local direction consistency features, and the convolution channels are 32, 64 and 128 in turn. In order to enhance the structure reconstruction ability of the image discontinuous area, two layers of bidirectional GRU network are connected in series in the back end to extract the context structure repair features in the time sequence. In the training stage, the stable section labeled in step S2 is taken as the sample of detecting continuous area, and the discontinuous section is taken as the sample of abnormal area, and the convolution module and the recurrent module are trained respectively, so as to improve the sensitivity and robustness of the model to the continuity identification of linear target.

[0034] In another embodiment, DeepLabv3+ adopts a typical Encoder-Decoder structure as a whole. The main process is: high-level semantic feature extraction is performed on the input image by the encoder, then context information at different scales is obtained by means of a multi-scale feature extraction and enhancement module, and finally spatial details and edge contours are gradually restored by the decoder, and a segmentation mask map consistent with the size of the input image is output. In the encoder part, DeepLabv3+ uses an improved Xception network as the backbone feature extraction structure, which combines depthwise separable convolution and atrous convolution technology. On the one hand, depthwise separable convolution significantly reduces the parameter quantity and computational complexity, and improves the operation efficiency of the network; on the other hand, atrous convolution enhances the perception ability of long-distance spatial patterns such as linear structures by introducing interval expansion receptive field in the convolution kernel, without significantly reducing the resolution of the feature map. For modeling of multi-scale environmental context information, DeepLabv3+ uses an atrous spatial pyramid pooling (ASPP) module. Unlike the traditional pyramid pooling (Pyramid Pooling) which obtains semantic context through multi-scale pooling, ASPP uses multiple atrous convolutions with different rates in parallel to extract features, thereby maintaining the spatial resolution of the feature map while improving the receptive field of the network and avoiding information loss caused by downsampling. In a typical configuration, ASPP contains three atrous convolution branches with rates of 6, 12, and 18, a standard 1x1 convolution branch, and a global average pooling branch to capture the overall background semantic information of the image. The outputs of the five branches are spliced and fused to form a feature vector containing rich scale information. The whole network process is: the input image extracts the basic feature map through the deep encoding network, these features are then input into the ASPP module, and after being fused after being modeled by atrous convolution with different rates and global context, multi-scale structure perception features are formed. Then, these fused features are input into the lightweight decoding module to recover the detailed information layer by layer, and output fine semantic segmentation results, especially with good edge retention ability for slender linear structures (such as submarine cables). This structure and the multi-level convolution and recurrent memory network module described in step S3 form a synergistic cooperation: the former focuses on extracting stable target spatial edges and direction trend features from the overall scale, and the latter focuses on modeling and repairing the time series of local non-continuous structures, thereby realizing high-robust detection of submarine linear targets in multiple scenes and multiple interference conditions.

[0035] Step S4: fuse the comprehensive features of the extracted detected continuous region and the detected discontinuous region, identify 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;

[0036] In an embodiment, the obtained cable target image is manually annotated. A polygon annotation mode is selected, and the edge contour of the target can be finely annotated. After annotation, the corresponding polygon mask can be saved, thereby constructing a seabed cable target sample set. The continuous region features and the discontinuous region features extracted in step S3 are fused to construct a complete seabed linear target sample set, which is used as the input of model training. Based on the DeepLabv3+ backbone network, the response ability to different scale targets is enhanced by introducing the atrous spatial pyramid structure (ASPP), and branch loss functions are used to supervise the learning of continuous segments and broken segments respectively, thereby improving the robustness of the model in recognizing seabed cable targets in complex scenes.

[0037] In another embodiment, the seabed cable target sample set is randomly divided into a training set and a test set. The model is trained on the training set. The input of the model is the seabed cable target image, which is subjected to feature extraction, enhancement, and edge information recovery by the DeepLabv3+ model, and outputs a mask of the same size as the original image. The loss value is calculated with the manually annotated mask. The random gradient algorithm is used to update the model parameters based on the error loss, and the iteration is performed until the loss value basically no longer changes, and the model is considered to have converged, and the training is stopped. The test set data is used for model testing. Based on the trained model, the samples in the test set are subjected to cable target segmentation, and the model is quantitatively evaluated based on the segmentation accuracy evaluation index, thereby constructing a seabed cable target detection model based on DeepLabv3+.

[0038] Step S5: using the seabed cable detection model, the obtained waterfall image is subjected to cable target segmentation, the segmentation accuracy of the seabed target detection model is evaluated based on semantic segmentation, and the model is optimized.

[0039] In an embodiment, the trained detection model is deployed on an underwater robot platform to perform real-time semantic segmentation on the seabed sonar waterfall image collected by the underwater robot. The length constraint and morphological correction are performed on the slender target region in the prediction result by the post-processing module, so as to ensure that the segmentation result is consistent with the physical cable shape. In the evaluation, the mIoU and FWIoU values and other indicators are used to evaluate the performance of the model, and different depths and different topography scenes are used as test groups to comprehensively verify the adaptability and segmentation accuracy of the model.

[0040] In another embodiment, the semantic segmentation is evaluated by the mean Intersection over Union (mIoU). The greater the mIoU value, the higher the segmentation accuracy of the model, and the expression is:

[0041]

[0042] wherein, and are label values of different categories, and the label value starts from 0; represents the number of pixels belonging to category but predicted as category ; represents the number of correctly classified pixels; is the maximum value of all category labels, i.e., the number of categories.

[0043] Considering that the underwater pipeline target pixels account for a small proportion in the image, if the Intersection over Union of each target is directly averaged, the prediction result of the pipeline target will have a great influence on the average Intersection over Union index. In order to better evaluate the performance of the model, the Frequency Wighted Intersection over Union (FWIoU) is used to evaluate the performance of the segmentation model in the experiment:

[0044]

[0045] wherein, and represent the labels of different categories; represents the number of pixels belonging to category target but predicted as category ; represents the number of pixels belonging to category j target but predicted as category ; represents the number of pixels correctly predicted as category target; is the maximum category label value; the category label value starts from 0, so, represents the number of all categories. For the pipeline target detection experiment, the value is 1. The FWIoU weights each target according to the proportion of the pixels of each target, and uses this weight value to weight the average of the Intersection over Union of each target, so it can more reasonably evaluate the performance of the model.

[0046] Preferably, step S1 comprises the following steps:

[0047] Step S11: Obtain original image data of the seabed scene by using synthetic aperture sonar original data;

[0048] Step S12: Decode the original image and generate a waterfall image of the seabed scene;

[0049] Step S13: Identify the starting response point of the target structure as a first feature point by analyzing the first texture feature and the first spatial gray feature map in the waterfall image;

[0050] Step S14: Extract a second feature point representing the position of the structure change by analyzing the second texture feature and the second spatial gray feature map in the waterfall image;

[0051] Step S15: Analyze the structure response difference between the first feature point and the second feature point.

[0052] In an embodiment, the system takes synthetic aperture sonar (SAS) raw data as input to obtain raw image data of the seabed scene to be analyzed. The raw image data includes information of the change of the intensity of the sound wave echo over time and space, and has rich texture and gray scale change characteristics. The system decodes the collected raw image data, including amplitude correction, coordinate projection conversion, and image amplitude mapping, and other preprocessing steps, to generate a standard format of the seabed waterfall image. The waterfall image can clearly present the linear extension state of the seabed target structure and the surrounding background environment in a two-dimensional manner. After generating the waterfall image, the system first optimizes a single strip image. The single strip image refers to a continuous imaging strip covering a certain width of the seabed area obtained by the synthetic aperture sonar in a single side-scan scanning process. The width is determined by the sonar array layout and working parameters, and the length changes with the extension of the track. For the single strip image, the system performs random noise and water column noise removal, amplitude equalization and radiation correction to eliminate brightness unevenness, geometric distortion correction and attitude compensation to correct the offset caused by platform roll, pitch or yaw, and unifies the spatial resolution and enhances the local contrast to improve the detail distinguishability. Subsequently, the system performs strip splicing processing on multiple strip images covering adjacent seabed areas, realizes spatial registration according to the navigation information and feature matching results, adjusts the brightness and texture consistency of the overlapping areas, and uses a smoothing fusion strategy to eliminate the splicing seams to generate a continuous and seamless large-area synthetic aperture sonar image, providing a complete background for the continuous trend analysis of the global linear target. The system extracts texture and spatial gray scale features from the waterfall image based on image analysis algorithms, constructs a first texture feature map and a first spatial gray scale feature map, and identifies the starting response position of the target structure in the image. The response point is usually manifested as a texture edge enhancement, a gray scale change mutation or a linear texture concentrated area, and the identification result is used as the first feature point of the target structure for further analysis of the change trend of the structure. The system continues to extract a second texture feature map and a second spatial gray scale feature map, and identifies the abnormal feature positions such as morphological deflection, trend change and brightness jump in the target structure by combining an edge enhancement operator and a direction gradient change detection technology, and extracts the second feature point. Finally, the system extracts the pixel connection path based on the image area between the first feature point and the second feature point, analyzes the difference features such as gray scale change, texture distribution and edge structure response, and obtains the structure response difference result to provide a basis for subsequent path structure analysis.

[0053] In another embodiment, after decoding the raw data, the real part and the imaginary part data in the raw image are squared and then the square root is taken. The result obtained by this processing is considered as the echo data corresponding to each pixel position of the image. The linear quantization of this set of echo data to the [0, 255] interval obtains the image data of the 8-bit gray scale image interval, and each scanning line containsN column, the scanning lines are stacked vertically according to the recording order to obtain a M row, obtaining a M x N image data matrix, and then adding a corresponding image header file to obtain a single-channel grayscale image.

[0054] Preferably, step S13 comprises the following steps:

[0055] Step S131: enhancing the local contrast of the waterfall image and performing histogram equalization processing to generate a single-channel grayscale image reflecting the brightness distribution characteristics as the first spatial grayscale feature map of the seabed image;

[0056] Step S132: constructing a local analysis window centered on an arbitrary pixel in the first spatial grayscale feature map, and evaluating the distribution trend of each pixel in the local analysis window in terms of texture gradient direction, brightness difference characteristics and regional contrast;

[0057] Step S133: determining the dominant direction of the most significant change in texture direction in the local window based on the distribution trend as the first texture feature of the pixel point;

[0058] Step S134: performing fusion analysis on the first texture feature of the pixel point and its corresponding grayscale response value to screen out a region with texture mutation and grayscale response extreme value characteristics in the image;

[0059] Step S135: extracting a response starting point with 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.

[0060] In an embodiment, to extract the starting response point of the seabed linear structure, the waterfall image is preprocessed, an adaptive histogram equalization algorithm (such as CLAHE) is used to enhance the local contrast and suppress the interference of background texture, and a first spatial grayscale feature map reflecting the brightness distribution characteristics is obtained. In the grayscale image, a sliding window of size MxM is constructed centered on each pixel, and the distribution trend of each pixel in the local region in terms of gradient direction, grayscale change rate and contrast index is analyzed. Specifically, the gradient direction is extracted by the Sobel operator, the grayscale change rate is obtained by grayscale difference absolute value normalization, and the contrast is calculated based on the local standard deviation. Subsequently, based on these statistical characteristics, the dominant direction of the most significant change in local texture direction is determined as the first texture feature of the pixel point. On this basis, the first texture feature of each pixel and its corresponding grayscale response value are fused and analyzed, and the edge starting region with good continuity and coherent response is extracted in the fusion score map, and the feature response point with stable local structure direction and complete morphology is screened out from the edge starting region as the first feature point of the seabed linear structure.

[0061] Preferably, step S14 comprises the following steps:

[0062] Step S141: adjusting the brightness dynamic range and edge enhancement processing on the waterfall image, enhancing the response intensity of the structural inflection point and the boundary transition zone in the image, to obtain a second spatial gray feature map;

[0063] Step S142: for any one pixel point in the image, a fixed size local analysis window is constructed with the pixel point as the center, and a plurality of direction and scale Gabor filter banks are used to calculate the response of the window respectively, and the maximum amplitude of all responses is extracted as the second texture feature of the pixel point;

[0064] Step S143: the second texture feature and the second spatial gray feature are fused by weighted superposition to generate a fusion response map highlighting the degree of local structure change;

[0065] Step S144: in the fusion response map, the gray scale change region, the texture direction mutation region and the structure trend jump position are analyzed, and the key pixel points with obvious inflection trend or spatial discontinuity are extracted;

[0066] Step S145: for the extracted key pixel points, response intensity screening and spatial distribution clustering analysis are performed, and the point set with clear structure change trend and continuous regional distribution is retained as the second feature point of the target structure.

[0067] In an embodiment, the system adopts a strategy based on the fusion of gray scale and texture features for the region where there is a structural change in the waterfall image, identifies the second feature points of the target structure to represent the local inflection, abrupt change in direction or connection interruption of the structure, and the like. The system performs brightness dynamic range adjustment and edge enhancement processing on the original waterfall image to improve the response intensity of the image in the structural boundary and transition area, and generates a second spatial gray scale feature map. This step effectively improves the image detail level by stretching the image histogram and introducing the Sobel edge enhancement operator. The system constructs a local analysis window of a fixed size (such as 11x11 or 15x15) centered on any pixel in the image. For each window, the system uses a multi-direction (such as 0°, 45°, 90°, 135°) and multi-scale parameter configured Gabor filter set to perform convolution response calculation, and extracts the result with the maximum response amplitude as the second texture feature of the pixel. The system performs feature fusion processing on the second texture feature and the second spatial gray scale feature obtained in step S141, adopts a weighted superposition method or a joint response scoring strategy, and constructs a fusion response map. The fusion image can reflect the local texture change and brightness structure response at the same time, and enhance the sensitivity to the inflection region. In the fusion response map, the system further analyzes the gray scale change region, the texture direction mutation region and the structure direction jump position, and extracts key pixel points through response threshold method. These points are usually concentrated in the positions where there is a significant direction change or morphological boundary mutation in the image. The system performs response intensity screening on the above extracted key pixel points, and removes weak response points; and uses a density-based spatial clustering analysis method (such as the DBSCAN algorithm) to remove isolated points and small clusters, and only retains the pixel point set with obvious structure change trend and continuous spatial distribution. The obtained point set is the second feature point of the target structure.

[0068] In another embodiment, the system processes a seabed synthetic aperture sonar image with a resolution of 768x512, generates a second spatial gray scale feature map of the image through histogram equalization and Laplacian enhancement processing, and the cable structure boundary and inflection region in the image are significantly enhanced. The image is processed by local sliding window, and a set of Gabor filters containing four directions and three scales is used for texture response extraction in each 11x11 window. The system retains the maximum amplitude response for each pixel point to form a second texture feature map. The map shows a higher texture response energy in the structural morphology change area. By weighted fusion (proportion 0.6:0.4) of the second gray scale feature map and the second texture feature map, a fusion response map is obtained. In the map, the system extracts the gray scale mutation points and texture direction jump points by threshold segmentation, and preliminarily identifies a batch of suspected structure change points. The system removes weak response points by response intensity sorting, and performs clustering analysis on the remaining point set in space to retain the region points with consistent structure trend and good continuity as the second feature points of the target structure.

[0069] Preferably, the step S15 comprises the following steps:

[0070] Step S151: Taking the first feature point as the starting point, acquiring the second feature point within its adjacent region in the image, and extracting the pixel connection path between the two points;

[0071] Step S152: Extracting the gray value sequence of the continuous pixels along the connection path; calculating the gray response gradient between the adjacent pixels in the gray value sequence;

[0072] Step S153: Based on the texture direction information of each pixel in the connection path, calculating the texture fluctuation trend of the texture direction;

[0073] Step S154: Combining the gray response gradient and the texture fluctuation trend, and counting the structure connectivity parameter of the path region;

[0074] Step S155: According to the gray response gradient and the texture fluctuation trend and the structure connectivity parameter, scoring the structure difference degree of the pixel connection path to generate the structure response difference.

[0075] In an embodiment, to effectively identify the local connection features of the subsea linear structure, the system analyzes the response difference of the path structure in the gray level and texture level based on the pixel connection path between the feature points, and generates a structure response difference result, which is used for subsequent path structure modeling and stability analysis. Specifically, as a first feature point has been identified in the image, the system takes the first feature point as the starting point, searches for a second feature point in its adjacent area, and extracts the pixel connection path between the first feature point and the second feature point through the spatial adjacency relationship between the image pixels. The path can be connected at the pixel level through the minimum gray gradient method or the dynamic programming method to ensure the continuity of the path and the stability of the response. Subsequently, the system extracts the gray value sequence of all continuous pixels along the extracted connection path, and calculates the gray response gradient between adjacent pixel pairs in the gray value sequence. The gradient reflects the gray change rate of the structure region, and is an important basis for judging the image structure boundary. Further, the system extracts and analyzes the change trend of the texture direction of each pixel in the connection path based on the texture direction information of the pixel, and constructs a fluctuation curve of the texture direction. This step usually uses a main direction filter to estimate the local main direction of the pixel window, form a texture direction sequence, and analyze the fluctuation amplitude of the sequence. Based on the gray response gradient and the texture fluctuation trend, the system statistically analyzes the structure connectivity of the connection path region, and extracts the structure connectivity parameters reflecting the path continuity and morphological consistency. The parameters can include gray consistency measure, texture direction continuity index, and path continuous curvature, etc. Finally, the system comprehensively scores the extracted connection path based on the gray response gradient, the texture direction fluctuation trend and the structure connectivity parameters, and generates a structure response difference value to represent the change degree between the structure regions connected by the path. The response difference value can be used as an important index for subsequent judgment of the stability of the cable direction and the structure mutation point.

[0076] In another embodiment, a synthetic aperture sonar waterfall image is processed, and a starting point of a known target cable segment in the image is taken as a first feature point. The system sets a neighborhood window with a radius of 20 pixels around the feature point, and locates a second feature point with a structural mutation in the region through texture jump detection and spatial gradient analysis. Then, 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 the path, the system extracts a gray value sequence, and calculates the gray response gradient between each two adjacent pixels through a difference method to form a complete gradient sequence. A local direction consistency estimation algorithm is further used to extract the texture main direction of each pixel on the path, and a first difference is taken on the direction sequence to obtain a texture direction change curve. The texture fluctuation degree is quantitatively analyzed by setting a threshold. The structural connectivity parameters such as the gray difference stability between each pixel on the path, the direction continuity, and the pixel continuity rate are calculated to evaluate the consistency of the overall shape of the path. A multi-dimensional scoring model is constructed according to the above three types of indexes to comprehensively determine the structural response difference degree of the connecting path. If the score value exceeds a set threshold, it is determined that the two regions connected by the path have significant structural changes, and the structural mutation path is marked, which is used for subsequent corner detection and direction division analysis.

[0077] Preferably, step S2 comprises the following steps:

[0078] Step S21: based on the structural response difference, calculating the linear rotation angle of the connecting path and the curvature change trend;

[0079] Step S22: using the linear rotation angle and the curvature change trend, dividing the target region into a continuous section with stable direction extension and a turning section with morphological inflection;

[0080] Step S23: taking the center skeleton line of the continuous section and the turning section as a reference, constructing a spatial distribution change map of the linear target;

[0081] Step S24: using the spatial distribution change map, combining with the submarine line tracking, extracting the continuous trend of the submarine cable;

[0082] Step S25: through the continuous trend of the submarine cable, tracking the offset rate of the structural extension direction and the local connectivity degree, marking the stable section and the abnormal variation section of the target structure in the image.

[0083] In an embodiment, based on the structural response difference obtained in step S1, the corresponding pixel connection path is extracted, and the linear rotation angle between the starting point and the ending point of the path is calculated, that is, the angle between the connection line direction of the path and the image reference axis (such as the X axis), to obtain a direction change index. At the same time, by the gradient change and direction offset of the pixel points on the path, the curvature change trend of the path as a whole is calculated, which is used to judge whether the path has a turning or sharp turning phenomenon. Based on the above two indexes, the target region is further divided into a continuous section (such as a straight line section) with stable direction extension and a turning section (such as a bending section) with shape turning. The center skeleton line of each section is extracted respectively and used as the main axis reference of the linear target to construct a spatial distribution change graph containing information such as direction change, curvature fluctuation, and connectivity strength. Subsequently, in combination with the cable tracking task in the seabed image, the spatial distribution change graph is used to extract the continuous running trajectory of the seabed cable. By analyzing the offset rate of the extension direction (that is, the variation amplitude of the direction per unit length) in the running trajectory and the structural connectivity between the path sections, the stable section (directionally coherent and gently curved) and the abnormal variation section (directionally jumping and sharply changing in curvature) of the target cable are automatically labeled.

[0084] In another embodiment, the structural response difference path in a 512x512 seabed image is analyzed, and the linear rotation angle range is calculated to be 5° to 42°, and the curvature change trend curve has 3 inflection points. According to the set curvature threshold, the path is divided into 2 continuous sections and 1 turning section. The center skeleton line of each section is used to interpolate and reconstruct the overall linear running direction, and a spatial distribution change graph is generated. Further analysis shows that the path main axis direction has a peak value of the offset rate at the 20th pixel, and the connectivity decreases, which is determined as an "abnormal variation section"; the rest is a "stable section". The segmentation information can be directly used for automatic inspection and path reconstruction tasks of seabed cables or communication pipelines. Preferably, step S21 comprises the following steps:

[0085] Step S211: based on the structural response difference, extracting a response path between the first feature point and the second feature point;

[0086] Step S212: along the response path, extracting a center line sequence of the path according to the spatial coordinates of each pixel point in the path;

[0087] Step S213: based on the center line sequence, calculating a direction vector between any two continuous pixel points in the path; and combining all the direction vectors into a path direction sequence;

[0088] Step S214: according to the path direction sequence, calculating the angle change between the direction vectors section by section to obtain the linear rotation angle of the path;

[0089] Step S215: Calculate the curvature change rate based on the linear rotation angle change of each segment in the path, and draw the curvature change trend curve of the path.

[0090] In an 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 pixel point sequence of the path, the center line sequence is extracted to represent the main trend of the overall response path. Then, the direction vectors between any two consecutive pixel points in the center line sequence are constructed, and the path direction sequence is obtained by sequential combination. According to the angle change between adjacent direction vectors, the linear rotation angle of the path is calculated segment by segment to describe the direction extension of the path. Further, based on the linear rotation angle change of each segment, the curvature change rate of the path is derived, and the curvature change trend curve is drawn to represent the bending degree and direction stability of the path in space.

[0091] In another embodiment, for a response path containing 40 pixel points, after extracting the center line sequence, 39 direction vectors are generated, and the direction angle change range is calculated to be 2.3° to 19.7°, and the average linear rotation angle is 7.6°. Further, the curvature change rate curve is obtained by difference calculation of the angle change of each segment, and obvious peaks appear at the 12th to 18th pixels, indicating that this segment has a morphological turning feature. This trend chart can be used to assist in identifying the bending position or risk segment of the submarine cable.

[0092] Preferably, step S22 comprises the following steps:

[0093] Step S221: Based on the linear rotation angle and the curvature change trend curve, the response path is divided into several adjacent sub-segments in pixel order;

[0094] Step S222: Calculate the average rotation angle and the average curvature change of each sub-segment, and compare them with the preset stable threshold value. If the linear rotation angle and the curvature value are both lower than the threshold value, mark the sub-segment as a continuous section with stable direction extension;

[0095] Step S223: If the linear rotation angle or the curvature change rate of a sub-segment is higher than the threshold value, mark the sub-segment as a turning section with morphological inflection;

[0096] Step S224: Perform continuity verification on the marking results of all sub-segments, eliminate short and isolated abnormal sections, and retain the main path section with continuous structural pattern and obvious change trend;

[0097] Step S225: Output the divided continuous section with stable direction extension and the turning section with morphological inflection.

[0098] In an embodiment, for the linear rotation angle and curvature variation trend curve of the aforementioned path, the response path is divided into several equal-length adjacent sub-segments (e.g., each containing 5 pixel points) in pixel order. For each sub-segment, the linear rotation angle mean and the curvature change rate mean are calculated respectively, and compared with the set stable threshold (e.g., angle threshold θ t =10°, curvature threshold κ t =0.1). If both indicators in a sub-segment are lower than the corresponding threshold, the sub-segment is marked as a "directionally stable continuous segment"; otherwise, if any indicator exceeds the threshold, it is marked as a "turning segment with shape inflection". To ensure the continuity and effectiveness of the path segmentation marking, the preliminary division result is post-processed to remove isolated segments with a length less than 3 pixels, and to merge continuous segments of the same type to enhance the structural coherence. The path structure division result is output, which includes continuous segments and turning segments.

[0099] In another embodiment, for a response path with a total length of 45 pixels, set the length of each sub-segment to 5 pixels, and divide it into 9 sub-segments. After calculation, the linear rotation angle means of the 1st-3rd, 7th-9th segments are 4.2°, 6.1°, 5.7°, etc., and the curvature change rate means are less than 0.08, meeting the stability judgment condition, so they are marked as directionally stable continuous segments; in the 4th-6th segments, the average angles are 12.9°, 14.6°, 11.3°, and the curvature fluctuates sharply, so they are marked as turning segments with shape inflection. Further, the 6th segment is removed as an irregular segment with only 3 pixels, and it is merged into the 5th segment. Three continuous segments and two main turning segments are identified, and this structure division provides significant morphological boundary basis for subsequent path skeleton fitting.

[0100] Referring to Figure 5 , the seabed cable target image set;

[0101] Preferably, step S24 includes the following steps:

[0102] Step S241: Based on the skeleton line structure extracted in the spatial distribution change map, the direction vector of the continuous segment and the turning segment in the image coordinate system is calculated, and the relative direction information of the skeleton line along the line is obtained;

[0103] Step S242: According to the relative direction information of the skeleton line in the spatial distribution map, the dynamic tracking processing based on the direction vector is performed on the extension path of the cable;

[0104] Step S243: In the tracking process, the interference area with water column shielding in the image is detected, and the cable trajectory nodes on both sides of the shielding area are identified respectively;

[0105] Step S244: Interpolation prediction and geometric fitting of the occluded interference region are performed by using a trajectory fitting algorithm to construct a structure extension trend completion result across the occluded region;

[0106] Step S245: An inter-frame structure connectivity graph is constructed based on the structure extension trend completion result and the extension path of the cable;

[0107] Step S246: The spatial cable path result extracted in the inter-frame structure connectivity graph is used to establish a global inter-frame connectivity graph, integrate the multi-segment submarine cable trend information, and extract the continuous trend of the submarine cable.

[0108] In an embodiment, the system obtains the center skeleton line coordinate sequence of each continuous segment and turning segment based on the skeleton line structure extracted from the aforementioned spatial distribution change graph. For each segment of the skeleton line in the image coordinate system, the trend direction vector thereof is calculated, and the extension direction information of the whole skeleton line is further integrated to form the direction sequence of the target structure. On this basis, the system performs dynamic tracking of the submarine cable target along the skeleton line direction by using the direction vector. In the process, when it is detected that there is an occluded segment caused by obvious water column interference or water body reflection in the image region, the cable trajectory nodes at the front and rear ends thereof are extracted, and a trajectory prediction model (such as spline interpolation, least square fitting, etc.) based on curve fitting is used to complete the path of the occluded region to construct a complete structure extension trend. Subsequently, the cable path in the multiple image frames is integrated by combining the completed trend data and the image frame sequence to construct an inter-frame structure connectivity graph, and further form a complete global inter-frame connectivity graph structure in the global frame set to realize the accurate extraction of the continuous trend of the submarine cable.

[0109] In another embodiment, for a cable detection task of a segment containing 10 underwater images, the system extracts the skeleton line structure and calculates the local direction vector thereof in each frame. In the 4th frame, an obvious water column occluded region is detected, the skeleton line end points before and after the occlusion are identified, and the path fitting is completed by using a cubic spline interpolation method to complete the missing cable segment. Subsequently, the system registers the skeleton line generated by prediction in the completion process with the actual skeleton line structure in the adjacent frame in the inter-frame connectivity graph to ensure the continuity of the trend and the spatial geometric consistency. All the path information extracted in the image frames is integrated to establish a global structure connectivity graph covering all the image frames to realize the continuous and complete identification of the trend of the submarine cable in the region.

[0110] Please refer to Figure 6 , the pipeline image detection result;

[0111] Preferably, step S25 comprises the following steps:

[0112] Step S251: Obtain the extension direction of the submarine cable based on the continuous trend of the submarine cable;

[0113] Step S252: In each frame of image, the geometric center point offset rate of the submarine cable between adjacent frames is calculated along the extension direction of the submarine cable, and the time sequence of the offset rate is smoothed to obtain a structural extension trend curve;

[0114] Step S253: Based on the skeleton connectivity in the spatial distribution change map of the submarine cable, the path continuity measure of the cable in the local window is counted, including the coherent length of the skeleton line segment, the fracture frequency and the turning angle change amplitude;

[0115] Step S254: The structural extension trend curve and the local path connectivity measure are integrated to set a joint threshold criterion of the offset rate and the connectivity;

[0116] Step S255: When the target cable region meets the smooth change of the offset rate and the connectivity index is higher than the set threshold, it is marked as a stable section, and when the target cable region has a sudden change in the offset rate or a significant decrease in the connectivity, it is marked as an abnormal variation section, and the spatial labeling results of the stable section and the abnormal variation section in the image are output.

[0117] In an embodiment, the system further models the extension direction of the submarine cable based on the continuous extension result of the submarine cable extracted as described above. Specifically, the system extracts the extension direction vector of the submarine cable based on the spatial coordinate sequence of the continuous skeleton line, and takes the direction as the reference direction. In the image sequence, the system traces the geometric center point of the submarine cable along the extension direction in each frame, and calculates the offset distance of the center point between adjacent frames, thereby obtaining the offset rate between frames. In order to remove the influence of jitter and small errors, the time domain smoothing filter method (such as moving average, Gaussian filter) is used to process the offset rate sequence, and the structural extension trend curve is generated. At the same time, based on the skeleton structure in the spatial distribution change map, the system counts the connectivity measure indicators (step S253) such as the coherent length of the skeleton line segment, the fracture frequency and the change amplitude of the turning angle in the local sliding window. Then, combined with the extracted extension trend curve, the joint determination criterion of the offset rate and the connectivity evaluation is set to evaluate the structural stability. If the cable offset rate in the target region changes smoothly and the local connectivity index is higher than the threshold, it is determined that the region is a structural extension stable section; if there is a sudden change in the offset rate, an increase in the skeleton fracture or a sharp change in the turning angle, it is marked as an abnormal variation section. The system outputs the spatial labeling results of the corresponding stable section and abnormal section in the image for subsequent analysis or processing.

[0118] In another embodiment, for a sequence of 30 frames of seabed images, the system identifies the continuous skeleton structure of the main cable and extracts its global extension direction. Then, the position of the center line of the cable in each frame is extracted, and the center point offset rate is calculated between frames to obtain a sequence of 29 values. To enhance robustness, the system uses a five-point sliding average algorithm to smooth the sequence and obtain the structural extension trend.

[0119] Next, the system selects a local 5x5 pixel window, evaluates the continuity of each position point corresponding to the skeleton line segment, and counts the number of breaks and the direction turning fluctuation amplitude in consecutive frames. By setting the combined threshold criteria of the offset rate change amplitude ≤ 2 pixels / frame and the skeleton continuity length ≥ 20 pixels, the system marks the structure in the range of frame 8 to frame 18 as a stable segment, and marks the structure from frame 24 to frame 27 as an abnormal variation segment due to frequent skeleton breaks and dramatic changes in the direction of the skeleton, and displays the spatial region marked with different colors in the image.

[0120] The above description is only a specific implementation of the present application, enabling those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope 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 method comprises the following steps: Step S1: obtaining original image data of a seabed scene, performing feature response processing thereon, wherein a starting response point of a target structure is identified as a first feature point by analyzing the image and recognizing the starting response point; a second feature point representing a structure change position is extracted; and a structure response difference between the first feature point and the second feature point is analyzed; Step S2: calculating a linear rotation angle of a connection path and a curvature change trend thereof based on the structure response difference, and constructing a spatial distribution change graph of a linear target, extracting a continuous trend of a seabed cable by tracking a displacement rate of a structure extension direction and a local connectivity degree, and combining seabed line tracking to label a stable section and an abnormal variation section of the seabed cable in the image; Step S3: designing a deep neural network comprising a multi-level convolution structure and a recurrent memory structure, using a convolution network to capture a multi-scale edge trend and a linear trend in the stable section, and using a recurrent network to extract structure repair features and context sequence features in an incoherent region, wherein the deep neural network structure comprises three convolution encoding modules at the front end, respectively using a 3*3 convolution kernel to extract multi-scale edge features and local direction consistency features, and the convolution channels are 32, 64 and 128 in sequence, and two layers of bidirectional GRU networks are connected at the back end to extract context structure repair features in a time sequence, and in the training stage, the labeled stable section is used as a sample of a detected coherent region, and the incoherent section is used as a sample of an abnormal region, and the convolution module and the recurrent module are trained respectively; Step S4: fusing comprehensive features of the detected coherent region and the detected incoherent region to identify a seabed target sample set, inputting the seabed target sample set into a model to construct a seabed linear target detection model based on DeepLabv3+; Step S5: using the seabed linear target detection model to perform cable target segmentation on the obtained waterfall image, evaluating the segmentation accuracy of the seabed linear target detection model based on semantic segmentation, and optimizing the model.

2. The method of claim 1, wherein the adaptive semantic segmentation of the seafloor linear target detection model is constructed by, Step S1 comprises the following steps: Step S11: obtaining original image data of a seabed scene from synthetic aperture sonar original data; Step S12: performing decoding processing on the original image to generate a waterfall image of the seabed scene; Step S13: identifying a starting response point of a target structure as a first feature point by analyzing a first texture feature and a first spatial gray feature map in the waterfall image; Step S14: extracting a second feature point representing a structure change position by analyzing a second texture feature and a second spatial gray feature map in the waterfall image; Step S15: analyzing a structure response difference between the first feature point and the second feature point.

3. The method of claim 2, wherein the adaptive semantic segmentation of the seafloor linear target detection model is constructed by, Step S13 comprises the following steps: Step S131: performing local contrast enhancement and gray histogram equalization processing on the waterfall image to generate a single-channel gray image reflecting brightness distribution characteristics, as a first spatial gray feature map of the seabed image; Step S132: constructing a local analysis window in the first spatial gray feature map with any pixel point as the center, and evaluating the distribution trend of each pixel in the local analysis window in terms of texture gradient direction, brightness difference characteristics and regional contrast; Step S133: based on the distribution trend, determine the dominant direction of the most significant texture direction change in the local window as the first texture feature of the pixel point; Step S134: fuse and analyze the first texture feature of the pixel point and its corresponding gray response value, and screen out the area with texture mutation and gray response extreme value characteristics in the image; Step S135: extract the response starting point of structure continuity and distribution continuity from the area with texture mutation and gray response extreme value characteristics as the first feature point of the target linear structure.

4. The method of claim 2, wherein the adaptive semantic segmentation of the seafloor linear target detection model is constructed by, Step S14 includes the following steps: Step S141: adjust the brightness dynamic range and edge enhancement of the waterfall image, enhance the response intensity of the structure inflection point and the boundary transition area in the image, and obtain the second spatial gray feature map; Step S142: for any one pixel point in the image, a fixed size local analysis window is constructed with the pixel point as the center, and a plurality of direction and scale Gabor filter banks are used to calculate the response of the window respectively, and the maximum amplitude of all responses is extracted as the second texture feature of the pixel point; Step S143: fuse the second texture feature and the second spatial gray feature, and generate a fusion response map that highlights the degree of local structure change by weighted superposition; Step S144: in the fusion response map, analyze the gray change area, texture direction mutation area and structure trend jump position, and extract the key pixel points with obvious inflection trend or spatial discontinuity; Step S145: select and cluster the extracted key pixel points according to the response intensity and spatial distribution, and retain the point set with clear structure change trend and continuous regional distribution as the second feature point of the target structure.

5. The method of claim 2, wherein the adaptive semantic segmentation of the seafloor linear target detection model is constructed by, Step S15 includes the following steps: Step S151: take the first feature point as the starting point, obtain the second feature point in its adjacent area in the image, and extract the pixel connection path between the two points; Step S152: extract the gray value sequence of the continuous pixels along the connection path; calculate the gray response gradient between adjacent pixels in the gray value sequence; Step S153: based on the texture direction information of each pixel in the connection path, calculate the texture fluctuation trend of the texture direction; Step S154: combine the gray response gradient and the texture fluctuation trend to calculate the structure connectivity parameter of the path area; Step S155: according to the gray response gradient and the texture fluctuation trend and the structure connectivity parameter, score the structure difference degree of the pixel connection path to generate the structure response difference.

6. The method of claim 1, wherein the adaptive semantic segmentation of the seafloor linear target detection model is constructed by, Step S2 includes the following steps: Step S21: based on the structure response difference, calculate the linear rotation angle and curvature change trend of the connection path; Step S22: divide the target area into continuous sections with stable direction extension and turning sections with shape inflection by using the linear rotation angle and curvature change trend; Step S23: take the center skeleton line of the continuous section and the turning section as a reference to construct the spatial distribution change map of the linear target; Step S24: use the spatial distribution change map to extract the continuous trend of the submarine cable in combination with the submarine line tracking. Step S25: Track the offset rate and local connectivity degree of the extension direction of the structure by the submarine cable, and mark the stable section and abnormal section of the target structure in the image.

7. The method of claim 6, wherein the adaptive semantic segmentation of the seafloor linear target detection model is constructed by, Step S21 includes the following steps: Step S211: Based on the structural response difference, extract the response path between the first feature point and the second feature point; Step S212: Along the response path, according to the spatial coordinates of each pixel point in the path, extract the center line sequence of the path; Step S213: Based on the center line sequence, calculate the direction vector between any two continuous pixel points in the path; all direction vectors are combined into a path direction sequence; Step S214: According to the path direction sequence, calculate the angle change between the direction vectors piece by piece to obtain the linear rotation angle of the path; Step S215: Calculate the curvature change rate based on the change of the linear rotation angle of each small section in the path, and form the curvature change trend curve of the path.

8. The method of claim 6, wherein the adaptive semantic segmentation of the seafloor linear target detection model is constructed by, Step S22 includes the following steps: Step S221: Based on the linear rotation angle and the curvature change trend curve, the response path is divided into several adjacent sub sections according to the pixel order; Step S222: Calculate the average rotation angle and the average curvature change of each sub section, and compare them with the preset stable threshold value. If the linear rotation angle and the curvature value are lower than the threshold value, mark the sub section as a continuous section with stable direction extension; Step S223: If the linear rotation angle or the curvature change rate of a sub section is higher than the threshold value, mark the sub section as a turning section with morphological inflection; Step S224: Perform continuity verification on the marking results of all sub sections, remove the abnormal sections with too short length and isolation, and keep the main path section with continuous structure mode and obvious change trend; Step S225: Output the divided continuous section with stable direction extension and the turning section with morphological inflection.

9. The method of claim 6, wherein the adaptive semantic segmentation of the seafloor linear target detection model is constructed by, Step S24 includes the following steps: Step S241: Based on the skeleton line structure extracted from the spatial distribution change graph, calculate the direction vector of the continuous section and the turning section in the image coordinate system, and obtain the relative direction information of the skeleton line along the line extension; Step S242: According to the relative direction information of the skeleton line in the spatial distribution graph, perform dynamic tracking processing on the extension path of the cable based on the direction vector; Step S243: In the tracking process, detect the interference area with water column shielding in the image, and identify the cable trajectory nodes on both sides of the shielding area respectively; Step S244: Use the trajectory fitting algorithm to perform interpolation prediction and geometric fitting on the shielded interference area, and construct the structure extension direction completion result across the shielding area; Step S245: Based on the structure extension direction completion result and the extension path of the cable, construct the inter frame structure connectivity graph; Step S246: Extract the spatial cable path result in the inter frame structure connectivity graph to establish a global inter frame connectivity graph, integrate the submarine cable extension information, and extract the continuous extension direction of the submarine cable.

10. The method of claim 6, wherein the adaptive semantic segmentation of the seafloor linear target detection model is constructed by, Step S25 includes the following steps: Step S251: Based on the continuous extension direction of the submarine cable, obtain the extension direction of the submarine cable; Step S252: In each frame of image, the geometric center point offset rate of the submarine cable between adjacent frames is calculated along the extension direction of the submarine cable, and the time sequence 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 map, the path continuity measure of the cable in the local window is counted, including the coherent length of the skeleton line segment, the fracture frequency and the turning angle change amplitude; Step S254: The structural extension trend curve and the local path connectivity measure are integrated to set the joint threshold standard of the offset rate and the connectivity; Step S255: When the target cable region meets the smooth change of the offset rate and the connectivity index is higher than the set threshold, it is marked as a stable section; When there is a sudden change in the offset rate or a significant decrease in the connectivity of the target cable region, it is marked as an abnormal variation section, and the spatial labeling results of the stable section and the abnormal variation section are output in the image.

Images