A method and system for assessing structural damage based on high-resolution remote sensing imagery

Abstract

This invention provides a method and system for assessing structural disaster damage based on high-resolution remote sensing imagery, comprising: Step 1, acquiring a high-resolution remote sensing imagery dataset for structural extraction, selecting a fixed number of samples as a training set to construct a sample library of structures in dense scenes, and selecting a fixed number of samples as a test set; Step 2, constructing a boundary buffer parsing network to extract dense structures and obtain the final structure raster result; Step 3, constructing a regularized vector fusion model, iteratively training it using the training set to obtain a structure vector generation model; Step 4, using the boundary buffer parsing network and regularized vector generation technology to predict the test set, obtaining a test set structure vector map, combining it with a disaster vector map extracted from multi-source disaster data, and outputting the structural disaster damage assessment result by fusing the structure vector map and the disaster vector map. This invention is used for the task of accurately extracting the vectorization of dense structures in disaster scenes.

Description

Technical Field

[0001] This invention belongs to the field of high-resolution remote sensing image interpretation, and specifically relates to a method and system for assessing structural disaster damage based on high-resolution remote sensing images. Background Technology

[0002] High-resolution remote sensing technology provides a clear and easily identifiable data foundation for the observation of structures. Compared with medium- and low-resolution remote sensing images, high-resolution remote sensing images provide a more intuitive overall layout of structures, more detailed geometric shapes, and richer spectral textures, making it possible to accurately extract structures and assess disaster damage, thereby providing a wealth of basic geographic information related to disasters.

[0003] Structure damage assessment is a comprehensive target interpretation task, encompassing structure extraction and damage assessment analysis. While much research exists on structure damage assessment, structure extraction in densely distributed environments remains a challenge. Influenced by geographical or human factors such as concentrated resource allocation, structures are often densely distributed. However, existing structure extraction methods struggle to accurately locate the ambiguous boundaries of targets in complex spatial distributions, leading to boundary adhesion problems and significantly reducing the accuracy of structure damage assessment. Existing boundary optimization-based structure extraction methods primarily focus on learning, optimizing, and accurately locating target boundaries. Classified by technical approach, these methods can be divided into two main categories: target component optimization algorithms and multi-stage optimization algorithms. Multi-stage optimization algorithms construct end-to-end model optimization strategies, achieving more robust extraction results in complex remote sensing scenarios. Cheng proposes a cascaded optimization network that progressively optimizes target boundary details while preserving the complete target structure, effectively extracting dense structures.

[0004] Although the aforementioned methods have significantly improved the performance of structure disaster damage assessment, they still struggle to effectively extract data in situations with densely distributed structures. Therefore, to address the problem of structure disaster damage assessment in densely distributed environments, a novel structure disaster damage assessment framework is proposed. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a method for assessing structural disaster damage based on high-resolution remote sensing imagery. First, the imagery is input into a boundary buffer parsing network to obtain structural raster results. Second, the structural raster results are input into a regularized vector fusion architecture to extract regularized structural vectors, which are then fused with a disaster vector map for disaster damage assessment. Finally, the fused results are output to achieve the assessment of structural disaster damage.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A method for assessing structural damage based on high-resolution remote sensing imagery includes the following steps:

[0008] Step 1: Obtain a high-resolution remote sensing image structure extraction dataset, select a fixed number of samples as the training set, construct a sample library of structures in dense scenes, and select a fixed number of samples as the test set.

[0009] Step 2: Construct a boundary buffer parsing network to extract dense structures from two perspectives: dense target boundary separation and fuzzy target boundary optimization, and obtain the final structure raster result;

[0010] Step 3: Construct a regularized vector fusion model and perform iterative training using the training set to finally obtain the structure vector generation model;

[0011] Step 4: Using the converged boundary buffer parsing network and regularized vector generation technology, predict the test set to obtain the test set structure vector map. Combine the disaster vector map extracted from multi-source disaster data, and perform structure disaster damage assessment and analysis by fusing the structure vector map and the disaster vector map, and output the structure disaster damage assessment results.

[0012] Furthermore, step 1 includes the following sub-steps:

[0013] Step 1.1: Use drones or high-resolution satellites to capture high spatial resolution images;

[0014] Step 1.2: Collect densely distributed structures in high spatial resolution images as samples, and create a high-resolution remote sensing image structure extraction dataset by combining the structure outline labels with the corresponding regional images.

[0015] Step 1.3: Divide the high-resolution remote sensing image structure extraction dataset into two parts: training set A, used to fit the parameters of the high-resolution remote sensing image structure extraction model, and test set B, used to evaluate the accuracy of the high-resolution remote sensing image structure extraction model.

[0016] Furthermore, step 2 includes the following sub-steps:

[0017] Step 2.1: Construct a boundary buffer parsing network. Use two stacked convolutional layers of a dual-task interactive learning architecture to capture the internal mask and boundary buffer of the target structure, respectively. The dual-task interactive learning architecture outputs the confidence map of the internal mask and the confidence map of the boundary buffer. Use a high-confidence voting mechanism to vote on the confidence maps of the internal mask and the boundary buffer, and output the raster result of the target structure with separated boundaries.

[0018] Step 2.2: Construct a high-frequency signal optimization module. The high spatial resolution remote sensing image processed by wavelet transform and the boundary-separated structure target raster results generated in Step 2.1 are used as input data and sent to the frequency-semantic fusion unit. The high-frequency signal at the blurred boundary is enhanced by channel-level data fusion. The enhanced input data is sent to the cascaded buffer parsing optimization structure. After the first stage, the two channels of the structure raster results in the input data are replaced by the raster results obtained from the previous loop refinement. The updated input data is sent to the second stage, where higher resolution features are used to repair the blurred boundary details. The process is repeated until the termination condition is met to obtain the final structure raster results.

[0019] Furthermore, the cascaded buffer parsing optimization structure in step 2.2 is a multi-resolution cascaded optimization architecture.

[0020] Furthermore, step 3 includes the following sub-steps:

[0021] Step 3.1: Construct a regularized vector fusion model using regularized vector generation technology;

[0022] Step 3.2: Train the structure boundary buffer parsing network based on the training set.

[0023] Furthermore, the regularized vector generation technique in step 3.1 includes target edge detection, redundant point removal, and polygon regularization.

[0024] Furthermore, step 3.2, which generates a disaster vector map using the threshold segmentation method, includes data preparation and preprocessing, threshold segmentation, and post-processing.

[0025] Data preparation and preprocessing include: collecting remote sensing image data from multiple remote sensing sensors as multi-source remote sensing images, and preprocessing the multi-source remote sensing images, including denoising, radiometric correction, and atmospheric correction.

[0026] Thresholding segmentation and post-processing include: traversing each pixel of the preprocessed multi-source remote sensing image, segmenting the pixels into different categories according to the relationship between their gray values ​​and thresholds, performing post-processing on the segmentation results to obtain a post-processed segmented image, and merging pixels in continuous regions into vector polygons on the post-processed segmented image to generate a disaster vector map.

[0027] Furthermore, in step 4, the generated disaster vector map is input into the geographic information system, overlaid and fused with the structure vector data, and the fused result is output.

[0028] On the other hand, the present invention provides a structure disaster damage assessment system based on high-resolution remote sensing imagery, comprising:

[0029] The structure extraction dataset acquisition module is used to acquire a high-resolution remote sensing image structure extraction dataset, select a fixed number of samples as the training set, construct a structure sample library in dense scenes, and select a fixed number of samples as the test set.

[0030] The boundary buffer parsing network construction module is used to build a boundary buffer parsing network to extract dense structures from two perspectives: dense target boundary separation and fuzzy target boundary optimization, and obtain the final structure raster result.

[0031] The regularized vector fusion architecture building module is used to build a regularized vector fusion architecture. It is trained and optimized through a training set and can be used to convert the final structure raster result into structure vectors.

[0032] The multi-source vector fusion module is used to predict the test set by using the boundary buffer parsing network after training convergence and the regularized vector generation technology to obtain the test set structure vector map. It combines the disaster vector map extracted from the multi-source disaster data, and performs structure disaster damage assessment and analysis by fusing the structure vector map and the disaster vector map, and outputs the structure disaster damage assessment results.

[0033] Furthermore, the boundary buffer resolution network includes:

[0034] The dual-task interactive learning architecture utilizes dual-task interactive learning to achieve joint optimization inference of the target's internal mask and boundary buffer.

[0035] The buffer parsing module parses the boundary buffer into an internal mask and a boundary mask of the structure target, and outputs a raster result of the structure target with separated boundaries;

[0036] The high-frequency signal optimization module enables fuzzy structure boundary optimization.

[0037] Compared with the prior art, the present invention has the following beneficial effects: the present invention can be used for the task of accurately extracting vectorized dense structures in disaster scenarios, and at the same time, by fusing with multi-source disaster information, the damage assessment and analysis results of the structures can be obtained. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0039] Figure 1 This is a schematic diagram of the dual-task interactive learning in step 2.1 of Embodiment 1 of the present invention.

[0040] Figure 2 This is a flowchart of the boundary buffer parsing network in step 2.2 of Embodiment 1 of the present invention.

[0041] Figure 3 This is the vector result diagram of the structure in step 4 of embodiment 1 of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0043] Example 1

[0044] This embodiment provides a method for assessing structural damage based on high-resolution remote sensing imagery, comprising the following steps:

[0045] Step 1: Obtain a high-resolution remote sensing image structure extraction dataset, select a fixed number of samples as the training set, construct a sample library of structures in dense scenes, and select a fixed number of samples as the test set.

[0046] Step 2: Construct a boundary buffer parsing network to extract dense structures from two perspectives: dense target boundary separation and fuzzy target boundary optimization, and obtain the final structure raster result;

[0047] Step 3: Construct a regularized vector fusion architecture and train a construct boundary buffer parsing network based on the training set;

[0048] Step 4: Using the converged boundary buffer parsing network and regularized vector generation technology, predict the test set to obtain the structure vector. Then, perform a structure disaster damage assessment and analysis by fusing the disaster vector map and the structure vector map, and output the structure disaster damage assessment results.

[0049] (I) A self-constructed high spatial resolution remote sensing imagery dataset for greenhouse structure extraction was used. The training dataset contains 1290 Google Earth images, each 512×512 pixels in size, with 44,176 greenhouse structures labeled for extraction. The test set consists of 430 images containing 14,237 greenhouse structures. This dataset includes images and annotations from six provinces and regions across China: Hubei, Liaoning, Shandong, Xinjiang Uygur Autonomous Region, Shaanxi, and Jiangsu.

[0050] 1.1. Select the training set and test set TrainA and TestB from the dataset to be used as sample data in this implementation.

[0051] (ii) Construct a boundary buffer parsing network to extract dense structures from two perspectives: dense target boundary separation and fuzzy target boundary optimization, and obtain the structure raster results.

[0052] Step 2.1 involves constructing a buffer parsing module for dense target boundary separation. Its core component is a dual-task interactive learning architecture. This architecture consists of two stacked convolutional layers, one for capturing general features and the other for generating task-specific features. On one hand, the two sub-tasks (target interior mask extraction and boundary buffer extraction) share features, which helps improve the model's generalization ability; on the other hand, the interactive learning of mutual information between the target interior mask and the boundary buffer strengthens the representation ability of task-specific features, simultaneously improving the accuracy of both tasks. Finally, as... Figure 1 As shown, the dual-task interactive learning architecture outputs the target's internal mask confidence map and the boundary buffer confidence map. A high-confidence voting mechanism is used to vote on the two confidence levels, and the target raster result with boundary separation is output.

[0053] Step 2.2: Construct a high-frequency signal optimization module for optimizing blurred target boundaries. First, the input image processed by wavelet transform and the raster results of structures generated by the buffer parsing module are used as input data. Combined with a frequency-semantic fusion unit, high-frequency enhancement is achieved through channel-level data fusion. A cascaded buffer parsing optimization structure is designed to progressively reconstruct the overall target structure and repair target boundary details, improving the extraction accuracy of dense structures. By combining the buffer parsing module and the high-frequency signal optimization module, dense target boundary separation and blurred target boundary optimization are jointly achieved. The overall framework of the boundary buffer parsing network is as follows: Figure 2 As shown.

[0054] (III) Construct a regularized vector fusion architecture and train a structure boundary buffer parsing network based on the training set.

[0055] Step 3.1: Convert the structure raster result into a structure vector using regularized vector generation technology. The basic process of regularized vector generation technology includes: target edge detection, redundant point removal, and polygon regularization.

[0056] The target edge detection algorithm aims to extract target edges from a raster image. It calculates the gradient magnitude and direction on the input structure raster result and uses non-maximum suppression to retain points with the largest magnitude along the gradient direction, thus refining the edges. Based on this, two thresholds are set: high and low. Pixels with gradient magnitudes above the high threshold are considered strong edge pixels, and pixels below the low threshold are considered weak edge pixels. Pixels in between are considered intermediate states. By connecting strong edge pixels and marking adjacent intermediate state pixels as edges, an edge map is output.

[0057] The redundant point removal algorithm aims to simplify complex edges into approximate polygons composed of fewer keypoints. First, the first and last points on the curve are selected as keypoints; points between them are called midpoints. Between the start and end points, the midpoint with the largest distance to the straight line segment connecting the start and end points is found. This largest distance is compared to a pre-set threshold. If the largest distance is less than the threshold, this curve segment can be approximated by this straight line segment, and the midpoint between the two points is removed. If the largest distance is greater than the threshold, this furthest point is selected as the new keypoint, and the curve segment is divided into two parts. All retained keypoints are combined to output the simplified polygon vector result.

[0058] The purpose of polygon regularization algorithms is to correct the distortion of polygon vectors and generate regularized vector results. Specifically, a dynamic programming method is constructed to correct the endpoints of the polygons, mainly ensuring the following constraints: the iterated line segments are within a specified tolerance range from the source line segments; the endpoints of the iterated line segments are within a specified tolerance range from the source endpoints; and the direction of the iterated line segments is consistent with the direction of the source. The output is the structure vector result, such as... Figure 3 As shown.

[0059] (iv) Using the boundary buffer parsing network and regularized vector generation technology after training convergence, the test set is predicted to obtain the structure vector. The disaster vector and the structure vector are fused to conduct the disaster damage assessment and analysis of the structure, and the disaster damage assessment results of the structure are output.

[0060] Input the constructed multi-source remote sensing image data to generate a disaster vector map. In this example, we will use the assessment and analysis of greenhouse damage caused by a blizzard disaster in Liaoning Province as a specific illustration. The multi-source remote sensing image data includes collected Sentinel 2 Top of Atmosphere (TOA) data, totaling 12 bands, with a resolution of 10 meters after panchromatic sharpening. To analyze the blizzard disaster situation, the Normalized Differential Snow Index (NDSI) was used, and threshold segmentation was employed to extract the snow cover range. The NDSI threshold was manually set to 0.62. To avoid water bodies being misidentified as snow, the Normalized Differential Water Index (NDWI) was used to remove incorrectly identified water bodies. Finally, a blizzard disaster vector map of Liaoning Province is generated.

[0061] Based on the structure vectors predicted by the model and the snowstorm disaster vector map of Liaoning Province extracted from multi-source remote sensing images, disaster damage assessment and analysis are conducted through vector fusion. The fused results are output.

[0062] Example 2

[0063] This embodiment provides a structure disaster damage assessment system based on high-resolution remote sensing imagery, including:

[0064] The structure extraction dataset acquisition module is used to acquire a high-resolution remote sensing image structure extraction dataset, select a fixed number of samples as the training set, construct a structure sample library in dense scenes, and select a fixed number of samples as the test set.

[0065] The boundary buffer parsing network construction module is used to build a boundary buffer parsing network to extract dense structures from two perspectives: dense target boundary separation and fuzzy target boundary optimization, and obtain the final structure raster result.

[0066] The regularized vector fusion architecture building module is used to build a regularized vector fusion architecture. It converts the final structure raster result into structure vector through regularized vector generation technology, and combines it with multi-source remote sensing imagery to generate disaster vector map through threshold segmentation.

[0067] The vector fusion module is used to predict the test set using a converged boundary buffer parsing network and regularized vector generation technology to obtain structure vectors. Combined with disaster vector maps extracted from multi-source remote sensing images, the module performs disaster damage assessment and analysis through vector fusion, and outputs the disaster damage assessment results.

[0068] Furthermore, the boundary buffer resolution network includes:

[0069] The dual-task interactive learning architecture utilizes dual-task interactive learning to achieve joint optimization inference of the target's internal mask and boundary buffer.

[0070] The buffer parsing module parses the boundary buffer into an internal mask and a boundary mask of the structure target, and outputs a raster result of the structure target with separated boundaries;

[0071] The high-frequency signal optimization module enables fuzzy structure boundary optimization.

[0072] It should be understood that any parts not described in detail in this specification belong to the prior art.

[0073] It should be understood that the above description of the preferred embodiments is quite detailed, but this should not be construed as limiting the scope of protection of this invention. It is neither necessary nor possible to exhaustively describe all possible implementations. Those skilled in the art, guided by this invention, can make substitutions or modifications without departing from the scope of the claims, all of which fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for evaluating damage to a structure based on high-resolution remote sensing images, characterized by, The method comprises the following steps: Step 1, obtaining a high-resolution remote sensing image construction extraction dataset, selecting a fixed number of samples as a training set, constructing a dense scene construction sample library, and selecting a fixed number of samples as a test set; Step 2, constructing a boundary buffer zone analysis network, separating dense target boundaries and optimizing fuzzy target boundaries for dense construction extraction, and obtaining a final construction raster result; Step 3, constructing a regularized vector fusion model and iteratively training the model using the training set to finally obtain a construction vector generation model; comprising the following sub-steps, Step 3.1, constructing a regularized vector fusion model through a regularized vector generation technique; the regularized vector generation technique comprises target edge detection, redundant point elimination, and polygon regularization; The target edge detection comprises: calculating the gradient amplitude and direction on the input construction raster result, and retaining the points with the maximum amplitude along the gradient direction through non-maximum suppression to refine the edges; on this basis, a high threshold and a low threshold are set, the pixels with a gradient amplitude higher than the high threshold are considered as strong edge pixels, the pixels with a gradient amplitude lower than the low threshold are considered as weak edge pixels; the pixels between the two are considered as intermediate states; the strong edge pixels are connected, and the intermediate state pixels adjacent to the strong edge are also marked as edges, and an edge map is output; The redundant point elimination comprises: first, selecting the first point and the last point on the curve as key points, and the points between the two points are called intermediate points; between the start point and the end point, find the point with the maximum distance from the start-end straight line among all the intermediate points, and compare the maximum distance with a pre-set threshold; if the maximum distance is less than the threshold, the curve segment can be approximated by the straight line segment, and the intermediate points between the two points are eliminated; if the maximum distance is greater than the threshold, the farthest point is taken as a new key point, and the curve segment is divided into two parts; the remaining key points are combined to output the simplified polygon vector result; The polygon regularization comprises: constructing a dynamic programming method to correct the endpoints of the polygon, and guaranteeing the following constraint conditions: the iterative line segment is within a specified tolerance range of the source line segment, the endpoint of the iterative line segment is within a specified tolerance range of the source endpoint, and the direction of the iterative line segment is consistent with the source direction, and outputting the construction vector result; Step 3.2, training the boundary buffer zone analysis network based on the training set; Step 4, using the trained boundary buffer zone analysis network and the regularized vector fusion model to predict the test set, obtaining a test set construction vector map, combining a disaster vector map extracted from multi-source disaster data, and performing construction disaster damage evaluation and analysis through construction vector map and disaster vector map fusion, and outputting a construction disaster damage evaluation result.

2. The method of claim 1, wherein the method comprises: The step 1 comprises the following sub-steps, Step 1.1, using a UAV or a high-resolution satellite to shoot and obtain high spatial resolution images; Step 1.2, collecting densely distributed constructions in the high spatial resolution images as samples, and labeling the construction outlines and corresponding area images to form a high-resolution remote sensing image construction extraction dataset; Step 1.3, the high-resolution remote sensing image construction extraction dataset is divided into two parts, a training set A for fitting the high-resolution remote sensing image construction extraction model parameters and a test set B for evaluating the precision of the high-resolution remote sensing image construction extraction model.

3. The method of claim 1, wherein the method comprises: The step 2 includes the following sub-steps: Step 2.1, a boundary buffer zone parsing network is constructed, two stacked convolutional layers of a dual-task interactive learning architecture are used to capture the mask inside the construction target and the boundary buffer zone respectively, the dual-task interactive learning architecture outputs the target inside mask confidence map and the boundary buffer zone confidence map; a high confidence voting mechanism is used to vote on the target inside mask confidence map and the boundary buffer zone confidence map, and a boundary-separated construction target raster result is outputted; Step 2.2, a high-frequency signal optimization module is constructed, the high spatial resolution remote sensing image processed by wavelet transform and the boundary-separated construction target raster result generated in step 2.1 are taken as input data and sent into a frequency-semantic fusion unit, the high-frequency signal at the fuzzy boundary is enhanced through channel-level data fusion, the enhanced input data is sent into a cascaded buffer zone parsing optimization structure, after the first stage, the two channels of the construction raster result in the input data are replaced by the raster result refined by the last cycle; the updated input data is sent into the second stage, the fuzzy boundary details are repaired using higher resolution features, and the cycle is repeated until the termination condition is met, and the final construction raster result is obtained.

4. The building damage assessment method based on high-resolution remote sensing images according to claim 3, wherein, The cascaded buffer zone parsing optimization structure in step 2.2 is a multi-resolution cascaded optimization architecture.

5. The method of claim 1, wherein the method further comprises: determining a damage level of the structure based on the damage area and the damage area ratio. The step 4 includes data preparation and preprocessing, threshold segmentation and post-processing through the threshold segmentation method to generate the disaster vector map; The data preparation and preprocessing includes: collecting remote sensing image data from multiple remote sensing sensors as multi-source remote sensing images, and preprocessing the multi-source remote sensing images, including denoising, radiation correction, and atmospheric correction; The threshold segmentation and post-processing includes: traversing each pixel of the preprocessed multi-source remote sensing image, segmenting the pixel into different categories according to its gray value and threshold, and post-processing the segmentation result to obtain a post-processing segmentation image; on the post-processing segmentation image, the pixels of the continuous region are merged into a vector polygon to generate the disaster vector map.

6. The building damage assessment method based on high-resolution remote sensing images according to claim 5, wherein, In the step 4, the generated disaster vector map is input into a geographic information system, superimposed and fused with the construction vector data, and the fused result is outputted.

7. A structure disaster damage assessment system based on high-resolution remote sensing images, characterized by, It includes: A construction extraction dataset acquisition module for acquiring a high-resolution remote sensing image construction extraction dataset, selecting a fixed number of samples as a training set, constructing a dense scene construction sample library, and selecting a fixed number of samples as a test set; A boundary buffer zone parsing network construction module for constructing a boundary buffer zone parsing network, performing dense construction extraction from two aspects of dense target boundary separation and fuzzy target boundary optimization, and obtaining a final construction raster result; A regularized vector fusion model construction module for constructing a regularized vector fusion model, which can be used to convert the final construction raster result into a construction vector through training set training and optimization. A multi-source vector fusion module is configured to utilize the boundary buffer resolution network after training convergence and a regularized vector generation technique to predict a test set to obtain a test set structure vector map, and combine a disaster vector map extracted from multi-source disaster data to perform structure disaster damage assessment analysis through structure vector map and disaster vector map fusion, and output a structure disaster damage assessment result. The structure disaster damage assessment system based on high-resolution remote sensing images is configured to perform the steps of the structure disaster damage assessment method based on high-resolution remote sensing images in any one of claims 1-6.

8. A structure disaster damage assessment system based on high-resolution remote sensing imagery according to claim 7, characterized in that, The boundary buffer resolution network comprises: A dual-task interactive learning architecture is configured to utilize dual-task interactive learning to achieve joint optimization inference of the target internal mask and the boundary buffer; A buffer resolution module is configured to resolve the boundary buffer into a structure target internal mask and a target boundary mask, and output a structure target raster result after boundary separation; A high-frequency signal optimization module is configured to achieve fuzzy structure boundary optimization.

Images