Remote sensing image building polygon contour extraction model based on end-to-end learning
By introducing end-to-end learning and frame field learning models in remote sensing images, the problem of error accumulation in building polygonal contour extraction is solved, and the precise extraction of complex building shapes and the improvement of building contour extraction accuracy in high-score remote sensing images is achieved.
Patent Information
- Application Number
- CN202411540384.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-31
- Publication Date
- 2025-03-21
AI Technical Summary
The prior art has problems of error accumulation in the extraction of polygonal contours of buildings, especially under the influence of shadows and spectral diversity in complex buildings and remote sensing images, and the extraction accuracy is insufficient.
The remote sensing image building polygonal contour extraction model is adopted based on end-to-end learning, combined with the frame field learning model and the building polygonization module, and the segmentation results are optimized through error backpropagation to achieve accurate extraction of polygonal contours.
The accuracy of building profile extraction in high-score remote sensing images is improved, error accumulation is reduced, and the ability to identify complex building shapes is improved.
Smart Images
Figure CN119672514A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of remote sensing image information extraction, and in particular to a remote sensing image building polygonal contour extraction model based on end-to-end learning. Background Art
[0002] In the study of building polygonal contour extraction, the method of directly extracting building polygonal contours from remote sensing images can certainly obtain its vector data more conveniently. However, its reliance on key points and key edges limits its extraction accuracy for more complex buildings. The segmentation method obtains raster data of buildings, which requires post-processing operations, which also affects the accuracy of the results. This problem can be solved by vectorizing the segmentation results. However, most of the vectorization extraction of segmentation results nowadays decomposes segmentation and vectorization into separate tasks, which is computationally intensive and requires careful adjustment of parameters to obtain the desired results. Most importantly, the segmentation of buildings will produce errors due to the influence of factors such as the shadows and spectral diversity of the remote sensing images themselves, and in the polygonization step, due to the regularity of the building shape, the vector results need to be simplified or regularized. The multi-task operation of the two will cause the accumulation of errors and affect the final extraction accuracy. Summary of the invention
[0003] In view of the technical problems existing in the prior art, the present invention provides a remote sensing image building polygonal contour extraction model based on end-to-end learning, introduces a frame field learning model to realize the polygonization of building contours, constructs an end-to-end learning framework, and utilizes error back propagation to ultimately achieve accurate extraction of complex building polygonal contours.
[0004] According to a first aspect of the present invention, there is provided a remote sensing image building polygonal contour extraction model based on end-to-end learning, comprising: a building segmentation network, a frame field learning network, a building polygonization module and a loss calculation module; The input of the building segmentation network is the original image, and the output is the segmentation result map of the boundary and interior of the building; The input of the frame field learning network is the features output by the backbone network and the splicing features of the segmentation result image, and the output is the frame field corresponding to each pixel of the image; the frame field represents the edge direction and tangent direction at the pixel; The input of the building polygonization module is the segmentation result map and the frame field, and the output is the building polygon outline; The loss calculation module calculates the end-to-end loss based on the difference between the polygon outline and the true value, and back-propagates the end-to-end loss to the building segmentation network. The building segmentation network segments the original image based on the end-to-end loss.
[0005] Based on the above technical solution, the present invention can also make the following improvements.
[0006] Optionally, the backbone network includes: a backbone network and a fully convolutional network; The backbone network includes: a U-Net network and a Resnet network, and each pooling layer and upsampling layer in the U-Net network is followed by a residual learning module of a Resnet network; The fully convolutional network is attached to the backbone network, and the input of the fully convolutional network is the output of the backbone network.
[0007] Optionally, the frame field is composed of two pairs of symmetric π vectors, To represent the four directions of each pixel, and , and Represent two pairs of opposite directions, and They represent the same direction as the building boundary and the direction perpendicular to it, respectively.
[0008] Optionally, the process of the building polygonization module outputting the building polygonal outline includes: The raster data is used to generate polygonal contours using the Marching Square algorithm, and the resulting point sequence is composed into the initial contour; The initial contour is deformed using an active skeleton model, and the active skeleton model constrains the deformation of the initial contour through an energy function.
[0009] Optionally, the energy function is obtained by , and It is composed of a linear combination of three functions; ; represents the internal segmentation image of the building, Indicates the set threshold range. P Represents the position of the contour vertex along the row and column axes, and pos represents the contour point set; ; is the value obtained by subtracting the coordinates of the corresponding pixel on the frame field from the skeleton contour vertex, E is the set of valid edge vectors, is the direction vector obtained by the two coordinates, is the center point, is the parameter pair output by the frame field learning network, f Used to distinguish inclined walls and corners and make the outline more regular; .
[0010] Optionally, the building polygonization module further includes: using a DP simplification algorithm and a regularization algorithm to simplify and regularize the polygonal outline of the building.
[0011] Optionally, the loss calculation to segment the original image based on the end-to-end loss includes: The output of the building segmentation network is the building boundary segmentation result map And the internal segmentation result map ; Using cross entropy loss and Dice loss The linear combination of and The difference from the true value; through the loss function To adjust and The consistency of the output results.
[0012] Optional, cross entropy loss ; Among them, mean means average, sum means sum, and reduction means loss. , For this sample The loss of each category, N represents the total number of loss categories; Dice loss ; X and Y represent the true value and predicted value of the sample respectively; and The linear combination of ; a and b are the set weight parameters.
[0013] Optionally, the loss calculation module calculates a loss function of the loss of the frame field learning network: ; in, is the line domain obtained by segmentation, is the vector tangent, represents a vector, so It represents the direction angle of the vector. is the direction aligned with the line domain tangent, so = is the direction to align with the contour domain.
[0014] Optionally, the loss calculation module calculates the end-to-end loss based on the difference between the polygon outline and the true value, including: calculating the coupling loss of the building segmentation network and the frame field learning network; The coupling loss includes: coupling loss at the edge of the building and the coupling loss between the frame field and the interior area of the building ; ; ; and The gradient operator is calculated for the segmentation result images of the interior and boundary of the building respectively. H and W are the height and width of the original image, x represents the point in the image, and I represents the image. Parameter pairs output by the frame field learning network.
[0015] The present invention provides a remote sensing image building polygonal contour extraction model based on end-to-end learning. Aiming at the problem of error accumulation in the current polygonal extraction method using building segmentation results, a frame field learning model is introduced to realize the polygonization of building contours, an end-to-end learning framework is constructed, and error back propagation is used to back propagate the accumulated errors in the building segmentation to polygonization steps, so that the errors in the next step can reversely optimize the segmentation of the previous step, and finally the error is minimized. The obtained results are used to optimize the effect of building segmentation, and finally the results are regularized according to the shape characteristics of the building itself, thereby improving the accuracy of building contour extraction from high-resolution remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 A flowchart of extracting polygonal contours of buildings from remote sensing images based on end-to-end learning provided by an embodiment of the present invention; Figure 2 A building segmentation network structure diagram provided by an embodiment of the present invention; Figure 3 A frame field learning network structure diagram provided by an embodiment of the present invention; FIG4( a) is a vector result diagram of the unoptimized embodiment 1 in the process of polygonizing a building provided by the present invention; FIG4( b ) is a vector result diagram of the unoptimized embodiment 2 in the process of polygonizing a building provided by the present invention; FIG. 4( c ) is a vector result diagram of the unoptimized embodiment 3 in the process of polygonizing a building provided by the present invention. DETAILED DESCRIPTION
[0017] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.
[0018] As the main source of building data, high-resolution remote sensing images can distinguish finer components inside objects with their clear spatial structure characteristics and surface texture information, make object edge information clearer, and make object pixel representation richer. The image performance is more in line with the real surface, providing an effective way to extract artificial objects and facilitating corresponding research in many fields. Therefore, using high-resolution remote sensing images to extract buildings has always been a direction of continuous exploration in the field of remote sensing.
[0019] Figure 1 The extraction flow chart of a remote sensing image building polygon contour extraction model based on end-to-end learning provided by an embodiment of the present invention is as follows: Figure 1 As shown, the present invention integrates segmentation and vectorization into one model, and uses the final result to optimize the segmentation result, and finally realizes a more refined extraction of the building outline. Specifically, the extraction model includes: a building segmentation network, a frame field learning network, a building polygonization module and a loss calculation module.
[0020] The input of the building segmentation network is the original image, and the output is the segmentation result map of the building boundary and interior.
[0021] The input of the frame field learning network is the features output by the backbone network and the concatenated features of the segmentation result image, and the output is the frame field corresponding to each pixel of the image; the frame field represents the edge direction and tangent direction at the pixel.
[0022] The input of the building polygonization module is the segmentation result map and frame field, and the output is the building polygon outline.
[0023] The loss calculation module calculates the end-to-end loss based on the difference between the polygon contour and the true value, and back-propagates the end-to-end loss to the building segmentation network. The building segmentation network segments the original image based on the end-to-end loss.
[0024] The present invention provides a remote sensing image building polygonal contour extraction model based on end-to-end learning. In order to solve the problem of error accumulation in the current method of polygonizing the building segmentation results, a frame field learning model is introduced to realize the polygonization of the building contour, an end-to-end learning framework is constructed, and error back propagation is used to finally realize the accurate extraction of the polygonal contours of complex buildings.
[0025] Example 1 Embodiment 1 provided by the present invention is an embodiment of a remote sensing image building polygon outline extraction model based on end-to-end learning provided by the present invention, combined with Figure 2 It can be seen that an embodiment of the extraction model includes: a building segmentation network, a frame field learning network, a building polygonization module and a loss calculation module.
[0026] The input of the building segmentation network is the original image, and the output is the segmentation result map of the building boundary and interior.
[0027] like Figure 2 The figure shows a building segmentation network structure diagram provided by an embodiment of the present invention. In a specific implementation, the original image is Image, H and W represent height and width respectively, the output segmentation result map , contains two channels ,in This corresponds to the segmentation diagram inside the building. The corresponding image is the building boundary segmentation map. After the building segmentation, it is optimized through frame field learning and frame field polygonization processing, and finally a vector polygon result is obtained. The loss is calculated with the true value and returned to the model to optimize the segmentation result.
[0028] The input of the frame field learning network is the features output by the backbone network and the concatenated features of the segmentation result image, and the output is the frame field corresponding to each pixel of the image; the frame field represents the edge direction and tangent direction at the pixel.
[0029] like Figure 3 The figure shows a frame field learning network structure diagram provided by an embodiment of the present invention. In a possible implementation manner, the backbone network includes: a backbone network and a fully convolutional network.
[0030] The backbone network includes: U-Net network and Resnet network. Each pooling layer and upsampling layer in the U-Net network is followed by a residual learning module of the Resnet network.
[0031] The fully convolutional network is attached to the backbone network, and the input of the fully convolutional network is the output of the backbone network.
[0032] The fully convolutional network consists of Convolutional layer, batch normalization layer, ELU nonlinear layer, The convolutional layer and the Sigmoid nonlinear composition, the backbone network is used to extract Characteristics of Dimensions .
[0033] Since a building is a regular polygon with angles, at a corner point, a single direction is not accurate for calculating the angle of the point, so two directions are needed to capture the direction at the corner point. In a possible embodiment, the frame field consists of two pairs of symmetrical π vectors, To represent the four directions of each pixel, and , and Represent two pairs of opposite directions, and They represent the same direction as the building boundary and the direction perpendicular to it respectively; a framework field of a 4-PolyVector field is assigned to each pixel point on the plane.
[0034] The frame field contains four directions for each pixel, corresponding to two complex numbers , the edge direction and tangent direction can be easily deduced using this data.
[0035] The frame field learning network is composed of a Convolutional layer, batch normalization layer, ELU nonlinear layer, A fully convolutional network consisting of a convolutional layer and a Tanh nonlinear layer is attached to the backbone network.
[0036] The input of the building polygonization module is the segmentation result map and frame field, and the output is the building polygon outline.
[0037] In a possible embodiment, the process of outputting the polygonal outline of a building by the building polygonization module includes: The raster data is used to generate polygonal contours using the Marching Square algorithm, and the resulting point sequence is composed into the initial contour.
[0038] The active skeleton model is used to deform the initial contour, and the active skeleton model constrains the deformation of the initial contour through an energy function.
[0039] Since some pixels between the walls of the building are connected, the connected graph formed constitutes the skeleton. Therefore, this method uses the skeleton instead of the contour for deformation optimization, which can better align the edge with the frame field. Specifically, the active skeleton model (ASM) is used for optimization, which is similar to the active contour model. Using a given initial contour, the gradient descent method is used to minimize the energy function, forcing the initial contour to change its shape, and moving the contour points to the optimal position at the minimum energy.
[0040] In one possible embodiment, the energy function is expressed by , and It is composed of a linear combination of three functions.
[0041] Segment the image from the skeleton outline and the building interior Within a certain threshold In the specific implementation, the general setting is , the calculation formula of the energy function is (1): (1)
[0042] Prepresents the position of the contour vertices along the row and column axes, POS Represents a set of contour points.
[0043] The energy term aligns the frame domain with the skeleton contour, where It is obtained by subtracting the coordinates of the corresponding pixels on the frame field from the vertices of the skeleton contour. E is the set of valid edge vectors, is the direction vector obtained by the two coordinates, is the center point, is the parameter pair output by the frame field learning network, f It is used to distinguish between inclined walls and corners and make the outline look more regular. The calculation formula of this function is shown in (2): (2)
[0044] is to ensure that the nodes on the skeleton contour remain tightly and evenly distributed. Similarly, by subtracting the two coordinates, we get , in order to make the nodes on the contour more uniform and smooth, The absolute value is taken and then squared to constrain the contour smoothness. The formula is as follows (3): (3)
[0045] The loss calculation module calculates the end-to-end loss based on the difference between the polygon contour and the true value, and back-propagates the end-to-end loss to the building segmentation network. The building segmentation network segments the original image based on the end-to-end loss.
[0046] In a possible implementation manner, the building polygonization module further includes: using a DP simplification algorithm and a regularization algorithm to simplify and regularize the polygonal outline of the building.
[0047] In the specific implementation, the process of polygonization of buildings is determined by the direction of segmentation and frame field. However, in the actual polygonization process, since the results obtained by segmentation and frame field are not regular, the final result does not meet the expectations of the building, as shown in Figures 4(a), (b) and (c) below. In Figures 4(a) and 4(b), there are obvious errors caused by the imperfection of the segmentation map, and in Figure 4(c), there is a gap between the angle of the generated polygon and the angle of the real building. Therefore, this method uses the DP simplification algorithm and the regularization algorithm to eliminate the obvious errors in this process and generate a polygonal outline of the building with a regular shape.
[0048] The simplification is divided into two steps. First, the corner points of the building are detected. The contour will be divided into polylines at the corner points of the building. Then, The algorithm simplifies the polyline and eventually generates a regular polygonal outline. In particular, the extraction of corner points is based on the direction information of the frame field, which is always aligned with the direction of its edge or its tangent. For a point on the edge, it connects two edges, corresponding to two frame field directions. If there is no corner point, the two frame fields have the same direction. If there is a corner point, the two frame fields corresponding to the corner point position will have different directions. Based on this theory, combined with the information stored in the frame field, the corner points of the polygon can be extracted.
[0049] Regularization introduces a regularization algorithm, which is an algorithm composed of empirical rules. It uses the idea of the main direction, takes the longest side as the main direction, and adjusts the remaining sides to make them parallel or perpendicular to the main direction. The direction is adjusted by setting an angle threshold and rotating along the midpoint of the line segment to reach the appropriate direction. At the same time, the distance between parallel lines is set according to the area. The threshold is set to determine whether to delete or add sides. This method has a good effect on eliminating some obvious short sides, overly sharp or smooth angles, and false polygons.
[0050] In a possible implementation manner, the loss calculation module includes multiple loss functions for different learning tasks, namely: building segmentation and end-to-end learning loss, frame field learning loss, and coupling loss of frame field and segmented image.
[0051] The calculation process of building segmentation and end-to-end learning loss includes: the two outputs of the segmentation task are internal segmentation output and edge segmentation output , both use cross entropy loss and Dice loss The difference between the edge segmentation and the true value is calculated by the linear combination of the two. The end-to-end loss is also calculated by the linear combination of the two. However, in the process of calculating the end-to-end loss, only the edge loss is calculated, and the internal loss is not calculated. In addition, since the edge segmentation and the internal segmentation are two separate outputs, in order to increase the consistency between the output results, a loss function is designed To adjust the internal segmentation and edge segmentation Consistency of output results.
[0052] Among them, the cross entropy loss It is used to calculate the degree of proximity between the predicted output and the true value. The predicted value and the true value have a certain probability distribution. The closer the two are, the closer their probability distributions are. The present invention simultaneously segments the edge and the internal area during the segmentation task, and uses a binary cross loss function, whose formula is shown in (4): (4)
[0053] Among them, mean means average, sum means sum, and reduction means loss. , For this sample The loss of each category, N represents the total number of loss categories, is the weight of the category, and its calculation formula is (5): (5)
[0054] For this sample The true value of each category, For this sample The predicted value of each category.
[0055] Dice loss also calculates the similarity between the true value and the predicted value, using X and Y to represent the true value and predicted value of the sample respectively. Its function is defined as formula (6): (6)
[0056] Building edge segmentation and interior segmentation losses can be used and The linear combination of is expressed as formula (7), a and b are weight parameters set, and in the embodiment provided by the present invention, it can be set The value of is 1, The value of is 0.2: (7)
[0057] Edge and internal area adjustment function The function makes the edge area obtained by segmentation equal to the spatial gradient norm of the internal area. This function makes the predicted edge and internal area more consistent. Its calculation formula is shown in (8): (8)
[0058] Where x represents a point in the image and I represents the image. When both terms of are 0, that is Located inside the building, not outside or on the boundary, the loss is 0. When located at the boundary of the building or outside the boundary, the loss is not 0, and if there is a gradient change outside the house, it is optimized to reduce it, so that the gradient inside the house is finally flattened to 0, and the modulus of the gradient in other places is equal to the boundary probability map.
[0059] The calculation process of frame field learning loss includes: the learning of frame field finally outputs four channels, corresponding to the four directions of the frame field, and correspondingly uses ( ) coefficient pair, and the ground truth label is the angle of the tangent vector of the polygon outline. , used to train the frame field loss. The frame field learning loss includes three aspects: aligning the frame field with the contour direction loss, aligning the frame field with the contour tangent direction loss, and smoothing the frame field loss.
[0060] In order to prevent the collapse of the frame field, the loss function is designed To force the frame field direction to align with the contour domain direction, the specific function formula is shown in formula (9): (9)
[0061] in is the line domain obtained by segmentation, is the vector tangent, represents a vector, so It represents the direction angle of the vector. is the direction aligned with the line domain tangent, so = is the direction aligned with the contour domain. When , it means that the polynomial is There is a root nearby, and at least one direction of the frame domain is aligned with the direction of the line domain.
[0062] In addition to the loss that encourages alignment with the contour domain direction, a loss function is designed To align the frame field direction with the tangent direction of the building boundary, the specific formula is shown in formula (10): (10)
[0063] The last loss function is the Dirichlet energy function that adjusts the smoothness of the frame field, so that the model learns a smoother frame domain. The specific function is shown in formula (11): (11)
[0064] in = , It represents the degree of change of the function in a certain area, so the integral of the gradient modulus is used to adjust the smoothness of the frame field.
[0065] The calculation process of the coupling loss between the frame field and the segmented image includes: in the network model, the frame field result and the segmentation result are output separately. In order to make the match between the two more consistent and increase the coupling, the coupling loss between the frame field and the segmentation result is added. This coupling loss includes two aspects: one is the coupling loss between the frame field and the edge of the building. , and the other is the coupling loss between the frame field and the interior area of the building , these two coupling losses will be introduced below respectively.
[0066] The coupling loss between the building edge and the frame field is to align the frame field direction with the spatial gradient direction of the building edge. Similarly, the coupling loss between the building interior area and the frame field is to align the frame field direction with the spatial gradient direction of the building interior area. Similar to the edge alignment loss in the frame field learning loss, two coupling losses are designed as shown in formulas (12) and (13), where It is a gradient operator for calculating the segmentation map output by the network.
[0067] (12) (13)
[0068] There are multiple different loss functions when calculating losses, and different losses have different units. Therefore, in the experiment, a randomly initialized network is used to randomly sample the training data set and calculate the normalization coefficient of each loss to balance the losses in different value ranges.
[0069] The embodiment of the present invention provides a remote sensing image building polygonal contour extraction model based on end-to-end learning. In order to solve the problem of error accumulation in the current method of extracting polygons using building segmentation results, a frame field learning model is introduced to realize the polygonization of building contours, an end-to-end learning framework is constructed, and the error back propagation is used to optimize the effect of building segmentation with the obtained results. Finally, according to the shape characteristics of the building itself, the result is regularized, and the accumulated errors in the building segmentation to polygonization step are back propagated, so that the error in the next step can reversely optimize the segmentation of the previous step, and finally the error is minimized, thereby improving the accuracy of building contour extraction from high-resolution remote sensing images.
[0070] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and for parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.
[0071] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0072] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0073] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0074] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0075] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.
[0076] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.
Claims
1. A remote sensing image building polygon outline extraction model based on end-to-end learning, characterized in that: The extraction model includes: a building segmentation network, a frame field learning network, a building polygonization module and a loss calculation module; The input of the building segmentation network is the original image, and the output is the segmentation result map of the boundary and interior of the building; The input of the frame field learning network is the features output by the backbone network and the splicing features of the segmentation result image, and the output is the frame field corresponding to each pixel of the image; the frame field represents the edge direction and tangent direction at the pixel; The input of the building polygonization module is the segmentation result map and the frame field, and the output is the building polygon outline; The loss calculation module calculates the end-to-end loss based on the difference between the polygon outline and the true value, and back-propagates the end-to-end loss to the building segmentation network. The building segmentation network segments the original image based on the end-to-end loss.
2. The extraction model according to claim 1, characterized in that The backbone network includes: a backbone network and a fully convolutional network; The backbone network includes: a U-Net network and a Resnet network, and each pooling layer and upsampling layer in the U-Net network is followed by a residual learning module of a Resnet network; The fully convolutional network is attached to the backbone network, and the input of the fully convolutional network is the output of the backbone network.
3. The extraction model according to claim 1, characterized in that The frame field consists of two pairs of symmetrical π vectors, To represent the four directions of each pixel, and , and Represent two pairs of opposite directions, and They represent the same direction as the building boundary and the direction perpendicular to it, respectively.
4. The extraction model according to claim 1, characterized in that The process of outputting the polygonal outline of the building by the building polygonization module includes: The raster data is used to generate polygonal contours using the Marching Square algorithm, and the resulting point sequence is composed into the initial contour; The initial contour is deformed using an active skeleton model, and the active skeleton model constrains the deformation of the initial contour through an energy function.
5. The extraction model according to claim 4, characterized in that The energy function is expressed by , and It is composed of a linear combination of three functions; ; represents the internal segmentation image of the building, Indicates the threshold range to be set. P Represents the position of the contour vertex along the row and column axes, and pos represents the contour point set; ; is the value obtained by subtracting the coordinates of the corresponding pixel on the frame field from the skeleton contour vertex, E is the set of valid edge vectors, is the direction vector obtained by the two coordinates, is the center point, is the parameter pair output by the frame field learning network, f Used to distinguish inclined walls and corners and make the outline more regular; 。 6. The extraction model according to claim 1, characterized in that The building polygonization module also includes: using DP simplification algorithm and regularization algorithm to simplify and regularize the polygonal outline of the building.
7. The extraction model according to claim 1, characterized in that The loss calculation for segmenting the original image based on the end-to-end loss includes: The output of the building segmentation network is the building boundary segmentation result map And the internal segmentation result map ; Using cross entropy loss and Dice loss The linear combination of and The difference from the true value; through the loss function To adjust and The consistency of the output results.
8. The extraction model according to claim 7, characterized in that Cross Entropy Loss ; Among them, mean means average, sum means sum, and reduction means loss. , For this sample The loss of each category, N represents the total number of loss categories; Dice loss ; X and Y represent the true value and predicted value of the sample respectively; and The linear energy combination is expressed as ; a and b are the set weight parameters.
9. The extraction model according to claim 1, characterized in that The loss calculation module calculates the loss function of the frame field learning network: ； Among them, H and W are the height and width of the original image, is the line domain obtained by segmentation, is the vector tangent, represents a vector, It represents the direction angle of the vector. = is the direction aligned with the contour field, x represents a point in the image, and I represents the image.
10. The extraction model according to claim 1, characterized in that The loss calculation module calculates the end-to-end loss based on the difference between the polygon outline and the true value, including: calculating the coupling loss of the building segmentation network and the frame field learning network; The coupling loss includes: coupling loss at the edge of the building and the coupling loss between the frame field and the interior area of the building ; ； ； in, and The gradient operator is calculated for the segmentation result images of the interior and boundary of the building respectively. H and W are the height and width of the original image, x represents a point in the image, and I represents the image. Parameter pairs output by the frame field learning network.