CN-117496183-B - Remote sensing image vector contour extraction method, system, storage medium and equipment

CN117496183BCN 117496183 BCN117496183 BCN 117496183BCN-117496183-B

Abstract

The invention discloses a remote sensing image vector building contour extraction method, which comprises the steps of constructing a building vector contour extraction model, obtaining a multidimensional feature map through a feature pyramid network, generating building position coordinates through a region suggestion network, generating integral contour features, corner features and edge features through a transform network, restraining the three features through a related loss function, obtaining point coordinates through the corner features, obtaining position features of a next node through combining the corner features, the edge features and the point coordinates, inputting the position features of the next node and Fourier shape description sub-features of the integral contour features into an RNN module, and carrying out iteration to obtain integral vector contour information. The invention can improve the misjudgment phenomenon of the ridge line of the building in the remote sensing image, effectively restrict the incomplete building shape, reduce the topology error in the outline of the building and reduce the error in the extraction process of the edge points of the building.

Inventors

HU ANNA
MIAO YUYANG
XU YONGYANG
WU LIANG
XIE ZHONG

Assignees

中国地质大学（武汉）

Dates

Publication Date: 20260508
Application Date: 20231127

Claims (10)

1. The remote sensing image vector building contour extraction method is characterized by comprising the following steps of: s1, acquiring a remote sensing image of a building, and dividing the remote sensing image into a training data set and a test data set, wherein the training data set comprises the remote sensing image and building corner coordinate information, and the test data set comprises the remote sensing image; S2, building a building vector contour extraction model, wherein the model comprises a feature pyramid network, a region suggestion network, a transducer network, a cyclic neural network, a Fourier transform module and a feature fusion module; S3, training the model by using a training data set to obtain contour information of a building, continuously iterating parameters in the model by taking a loss function in optimized contour information as a target until a building vector contour extraction model converges, and obtaining a trained building vector contour extraction model, wherein the method specifically comprises the following steps of: s31, inputting the training data set into a feature pyramid network to obtain a multi-dimensional feature map of the whole remote sensing image; s32, inputting the multi-dimensional feature map into a regional suggestion network to generate position coordinates of a building in the remote sensing image; s33, inputting the position coordinates of the single building into a transducer network to generate preliminary contour information of the building, wherein the preliminary contour information comprises detailed integral contour features of the building, key corner features of the building and edge features of the building; S34, taking the maximum value point in the key corner features as the initial corner point of the building edge to obtain a point coordinate, and inputting the point coordinate, the building key corner features and the building edge features into a circulating neural network to obtain the position feature of the next node; S35, inputting the overall outline features of the building into a Fourier transform module, converting the overall outline features of the building into boundary point coordinates, and performing Fourier transform to obtain Fourier shape description sub-features; S36, inputting the position features of the next node obtained in the step S34 and the Fourier shape description sub-features obtained in the step S35 into a feature fusion module, generating coordinates of the next node, generating building edge corner points one by one through continuous iteration, and forming complete building contour information, wherein the complete building contour information is constrained through a vector contour optimization loss function; s4, inputting the test data set into the trained building vector contour extraction model to obtain the vector contour information of the building.
2. The method for extracting the outline of the remote sensing image vector building according to claim 1, wherein the overall outline characteristics of the building are constrained by an overall outline optimization loss function, and the overall outline optimization loss function is as follows: Wherein, the In order to output the number of categories, The image class i segmentation mask predictions generated for the Transformer network, And dividing the mask real result for the ith class.
3. The method for extracting the outline of the remote sensing image vector building according to claim 1, wherein the key corner features of the building are constrained by a key corner optimization loss function, and the key corner optimization loss function is: Wherein, the In order to output the number of categories, Key corner predictors generated for the Transformer network, The key corner real information is obtained.
4. The method for extracting the outline of the remote sensing image vector building according to claim 1, wherein the edge features of the building are constrained by an edge optimization loss function, and the edge optimization loss function is: Wherein, the In order to output the number of categories, Edge predictors generated for the Transformer network, Is the edge real information.
5. The method for extracting a vector building contour from a remote sensing image according to claim 1, wherein the vector contour optimization loss function is: Wherein, the For a building outline tag, The predicted building profile for the Transformer network, For coordinates within the current building outline box, guass () is a gaussian function, used to give a specific weight to an area inside the building that does not belong to a boundary, Is a super parameter, a is a matrix of values 1, corresponding to the feature size, which sets the non-boundary pixels to 1, Representing the positive part of the interior of the building outline, refers to the set of coordinates within the building outline defined as the "building" part, Representing the negative part within the building outline refers to the set of coordinates within the building outline defined as the "non-building" part.
6. The method of claim 1, wherein the change of the coordinates of the points in step S34 with the number of iterations is (starting point ), (starting point, starting point next point), (starting point, starting point next two points, starting point next three points).
7. The method for extracting the outline of the remote sensing image vector building according to claim 1, wherein the calculating method of the fourier shape descriptor features is as follows: Wherein D is a normalized Fourier shape descriptor feature vector, For the first frequency component of the fourier descriptor, For the nth frequency component of the fourier descriptor, Representing the norm.
8. A remote sensing image vector building contour extraction system, comprising: The system comprises a data acquisition module, a data processing module and a data processing module, wherein the data acquisition module is used for acquiring a remote sensing image of a building and dividing the remote sensing image into a training data set and a test data set, the training data set comprises the remote sensing image and building corner coordinate information, and the test data set comprises the remote sensing image; the model construction module is used for constructing a building vector contour extraction model, and the model comprises a feature pyramid network, a region suggestion network, a transducer network, a cyclic neural network, a Fourier transform module and a feature fusion module; The model training module is used for training the model by using a training data set to obtain contour information of a building, and continuously iterating parameters in the model until the building vector contour extraction model converges by taking a loss function in the optimized contour information as a target to obtain a trained building vector contour extraction model, wherein the model training module comprises the following specific steps of: inputting the training data set into a feature pyramid network to obtain a multi-dimensional feature map of the whole remote sensing image; Inputting the multi-dimensional feature map into a regional suggestion network to generate position coordinates of a building in a remote sensing image; Inputting the position coordinates of a single building into a transducer network to generate preliminary contour information of the building, wherein the preliminary contour information comprises detailed integral contour features of the building, key corner features of the building and edge features of the building; taking the maximum value point in the key corner feature as the initial corner point of the building edge to obtain a point coordinate, and inputting the point coordinate, the building key corner feature and the building edge feature into a cyclic neural network to obtain the position feature of the next node; inputting the overall outline features of the building into a Fourier transform module, converting the overall outline features of the building into boundary point coordinates, and carrying out Fourier transform to obtain Fourier shape description sub-features; inputting the obtained position features of the next node and the obtained Fourier shape description sub-features into a feature fusion module, generating coordinates of the next node, generating building edge corner points one by one through continuous iteration, and forming complete building contour information, wherein the complete building contour information is constrained through a vector contour optimization loss function; and the detection module is used for inputting the test data set into the trained building vector contour extraction model to obtain the vector contour information of the building.
9. A computer-readable storage medium, in which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method according to any of the claims 1-7.
10. An electronic device comprising a processor and a memory, the processor being interconnected with the memory, wherein the memory is configured to store a computer program comprising computer readable instructions, the processor being configured to invoke the computer readable instructions to perform the method of any of claims 1-7.

Description

Remote sensing image vector contour extraction method, system, storage medium and equipment Technical Field The invention relates to the technical field of remote sensing application, in particular to a remote sensing image vector contour extraction method, a remote sensing image vector contour extraction system, a storage medium and a storage device. Background The earth observation by using remote sensing satellite data is an acquisition means for rapidly acquiring earth surface information. The massive remote sensing data resources enable all-round, all-day-time and multi-dimensional observation and monitoring in a large scale range to be possible. The target objects in the high-resolution remote sensing image are quite rich, wherein the building is the most typical artificial ground object. The high-resolution remote sensing image can capture abundant scene detail information, and the color, texture and structure topology information of the building can be better characterized. As the visual sense and the spectrum wave band of the high-resolution remote sensing image are similar to those of human visual sense, the high-resolution remote sensing image has better visibility and interpretability, has the characteristics of large scale, wide range and sufficient data quantity, and provides a new idea for large-scale building contour extraction research. The automatic and accurate extraction of the building by fully and timely utilizing the high-resolution visible light remote sensing image has wide practical significance. In recent years, as the resolution of the obtained visible light remote sensing image is continuously improved, higher requirements are also put on the outline precision of the extracted building. The traditional building contour extraction method based on morphological algorithm is not used any more because of the great difficulty in acquiring the prior knowledge of the shape. The deep learning algorithm-based building contour extraction method undergoes a transition from a semantic segmentation method, an instance segmentation method, a single building vector contour extraction method to a multi-building vector contour extraction method. The building contour extraction method based on semantic segmentation can accurately position the building in the remote sensing image through strong image feature extraction capability, and extracts a relatively complete pixel-level building contour. However, semantic segmentation does not enable individual discrimination of buildings in dense building scenarios, which would lead to reduced accuracy of remote sensing interpretation applications based on building contours. Thus, to accurately distinguish the location of individual building contours in a remote sensing scene, an example segmentation algorithm was introduced into the building contour extraction study. The example segmentation method is used for extracting the building object at the pixel level by extracting the building outline and wrapping the rectangular frame and then carrying out example segmentation on the object in the frame. However, most of the smart city construction studies based on the building contour position are based on vector building contours, and a series of problems (the spatial correlation information between key corner points in the conversion process is lost) are caused by performing a grid-to-vector method on the pixel-level building information extracted by the example segmentation method. Therefore, the direct extraction of vector-level building corner points from the grid remote sensing image is the key point of current research. The single building vector contour extraction method also continuously develops from the practical application point of view, and the current building vector contour extraction method can realize simultaneous extraction of multiple building information. However, the method of forming the complete building contour by extracting key corner points of the building and then sequentially connecting them still cannot solve the problem of the topology error of the building (such as crossing of edge lines, overlapping of the building, non-closing of the building contour, etc.) generated by the point-to-edge connection. Disclosure of Invention In order to solve the above problems, the present invention provides a method for extracting a remote sensing image vector building contour, comprising the following steps: s1, acquiring a remote sensing image of a building, and dividing the remote sensing image into a training data set and a test data set, wherein the training data set comprises the remote sensing image and building corner coordinate information, and the test data set comprises the remote sensing image; S2, building a building vector contour extraction model, wherein the model comprises a feature pyramid network, a region suggestion network, a transducer network, a cyclic neural network, a Fourier transform module and a feature fusion module