A method for extracting base vectors of high-rise buildings from remote sensing images based on deep learning
Through deep learning technology, the dual-path U-Net model with adaptive multi-scale frequency domain enhancement and dynamic context fusion, combined with curvature optimization algorithm, the accuracy and efficiency problems of substrate vector extraction of high-rise buildings in remote sensing images are solved, achieving high Accurate and automated vector data generation, suitable for complex urban environments.
Patent Information
- Application Number
- CN202411273268.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-12
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-09-12
AI Technical Summary
The prior art is difficult to effectively extract the base vector data of high-rise buildings in remote sensing images, resulting in low position accuracy, low efficiency of traditional methods and inability to meet the rapid development of social needs and large-scale data processing requirements.
Using a deep learning-based method, combining adaptive multi-scale frequency domain enhancement algorithm, a dual-path U-Net model with dynamic context fusion and a vectorization algorithm based on curvature optimization, the building substrate is automatically extracted and optimized from remote sensing images to generate high Vector data with high accuracy, automation and strong adaptability.
It significantly improves the accuracy and robustness of building boundary recognition, overcomes the limitations of traditional models when processing multi-scale information, and maintains high accuracy at different scales, which is suitable for different application scenarios, and improves processing efficiency. and extraction accuracy.
Smart Images

Figure CN118799739B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a method for extracting base vectors of high-rise buildings in remote sensing images based on deep learning. Background Art
[0002] The automatic extraction of buildings in remote sensing images has always been an important research topic in the fields of urban planning, communication resource allocation, land and space detection, basic geographic information updating, and urban three-dimensional real-scene construction. This field has broad application prospects. With the increasing number of high-rise buildings in cities, remote sensing images obtained through aviation and aerospace platforms often have projection deviations, especially in the images of high-rise buildings. Since this deviation relies on extracting the roof outline of the building and using it as vector data, it will lead to low position accuracy. In contrast, the base outline of the high-rise building can more accurately represent the actual location of the building.
[0003] Traditional extraction methods usually rely on manual visual interpretation of images. By manually drawing the vector outline of the roof, and then correcting the projection deviation based on experience and measurement, the actual base position of the building is determined. This method is not only inefficient, but also slow in production due to its reliance on manual operation. Its accuracy and timeliness are difficult to meet the rapidly developing social needs and large-scale data processing requirements. Traditional building extraction methods usually use the pixel, shape and spectral characteristics of the building as the basis for judgment. These methods can work in some cases, but they are limited by multiple factors such as season, light, image scale, building shape, etc., and show poor feature representation capabilities and can only be effective in specific environments.
[0004] With the rapid development of deep learning technology, neural network algorithms based on deep learning have been widely used in building target detection and image classification. Existing remote sensing image interpretation methods based on deep learning mostly focus on the extraction of building roof contours. Due to the limitations of remote sensing image acquisition methods, changes in image resolution, the complexity of the shape of high-rise buildings, and factors such as building shadows and occlusions, the automatic extraction of high-rise building base vectors still faces huge challenges.
[0005] Therefore, how to provide a method for extracting base vectors of high-rise buildings in remote sensing images based on deep learning is an urgent problem that technicians in this field need to solve. Summary of the invention
[0006] One purpose of the present invention is to propose a method for extracting base vectors of high-rise buildings from remote sensing images based on deep learning. The present invention introduces an adaptive multi-scale frequency domain enhancement algorithm, a dual-path U-Net model with dynamic context fusion, and a vectorization algorithm based on curvature optimization, and describes in detail the whole process of automatically extracting and optimizing the building base from remote sensing images. The method has the advantages of high processing accuracy, high degree of automation, and strong adaptability.
[0007] The method for extracting base vectors of high-rise buildings from remote sensing images based on deep learning according to an embodiment of the present invention comprises the following steps:
[0008] S1. Obtain remote sensing image data, use adaptive multi-scale frequency domain enhancement algorithm, perform multi-scale processing on the image in the frequency domain, and highlight the edges and detail features of the building;
[0009] S2. Based on the pre-processed image data, automatically label the high-rise building sample set, generate labeled training data, and perform data expansion processing on it;
[0010] S3, Framework based, using Feature extraction backbone network, building dual paths with dynamic context fusion Model;
[0011] S4. Use dual paths The model extracts features from the input image, captures large-scale features of buildings through global paths, and captures edge details of buildings through local paths. Based on the extracted top contours of buildings and combined with height information, the model calculates the preliminary position of the building base through geometric deduction, and uses a dynamic context fusion mechanism to combine large-scale features with edge detail information to generate a building base feature map.
[0012] S5. Based on the fused building base feature map, multi-stage boundary refinement and segmentation are performed to identify and refine the boundaries of buildings, especially in areas covered by shadows and with complex shapes, by adjusting the connection points of boundary segments and iteratively correcting the building geometry;
[0013] S6, processing the segmentation results, optimizing the outline of the building base, removing unnecessary details, improving the smoothness and consistency of the overall boundary, and forming a binary image of the building base;
[0014] S7. The binary image of the building base is converted into vector data through a vectorization algorithm based on curvature optimization, and multi-level vector optimization is performed. The structure of the vector data is adjusted to make it suitable for analysis and application at different scales, and finally the vector data of the building base is generated.
[0015] Optionally, the S1 specifically includes:
[0016] S11, acquiring remote sensing image data, the image data including satellite images and aerial images of multi-spectral and panchromatic bands, and converting the remote sensing image data into frequency domain for multi-scale analysis;
[0017] S12, perform multi-scale decomposition on the image data converted to the frequency domain, and use a multi-resolution sub-band decomposition algorithm to convert the image data Decomposed into sub-bands of different resolutions and frequencies, each sub-band contains frequency domain information at different scales, among which the low-frequency components Preserve global features and high-frequency components Reflect the edges and details of buildings:
[0018] ;
[0019] in, is the subband coefficient, is the corresponding subband basis function;
[0020] S13, at different scales, for high frequency components Perform adaptive enhancement processing through adaptive enhancement function Adaptively adjust the intensity of high-frequency components according to local image features to highlight the edges and details of buildings:
[0021] ;
[0022] in, and According to different scales Adaptive adjustment parameters;
[0023] S14. At each scale, the enhanced high-frequency component is recombined with the low-frequency component to generate enhanced image data. ;
[0024] S15. Convert the enhanced image data of each scale from the frequency domain back to the spatial domain, perform multi-scale fusion, and generate integrated multi-scale enhanced remote sensing image data .
[0025] Optionally, the S2 specifically includes:
[0026] S21. Obtain remote sensing image data after multi-scale frequency domain enhancement processing , the image data contains the edges and detail features of the building;
[0027] S22. In the image data The spectral feature analysis of the sample set data of the building is performed to extract the spectral information of each band, and the attributes of the building samples are annotated based on the analysis results to generate the spectral feature annotation data of the building. , including building top outlines and other relevant attribute information:
[0028] ;
[0029] in, The corresponding band The weight of Indicates in band The image data on Indicates the total number of bands;
[0030] S23. Labeling data based on spectral features , extract the top outline information of the building , the top profile information is obtained by analyzing the spectral feature data and edge detection;
[0031] S24, the marked building spectral characteristic data The corresponding image data Combined to form a preliminary annotated training data set :
[0032] ;
[0033] in, represents the index of the data sample, is the number of samples initially labeled;
[0034] S25. Preliminary labeled training data set Data expansion is performed, and a data expansion method based on generative adversarial networks is used to generate diverse training samples. The expanded training data set is represented as , and used for model training:
[0035] ;
[0036] in, is the coordinate after the expansion operation, is the number of samples after expansion.
[0037] Optionally, the S4 specifically includes:
[0038] S41, the training image dataset after data expansion Input to the dual-path U-Net model, where each image data sample As input to the model, the global path and the local path simultaneously receive input image data;
[0039] S42, input image data through the global path For large-scale feature extraction, the global path captures the overall structural characteristics of the building through a series of convolutional layers and activation functions to generate a large-scale feature map ;
[0040] S43, input image data through local path Perform edge detail feature extraction. The local path extracts the edge details of the building through the receptive field and convolution operation to generate an edge detail feature map ;
[0041] S44, based on the extracted building top outline and spectral characteristic data , combined with the building height information , the preliminary position of the building base is calculated by geometric deduction and spectral feature correlation analysis:
[0042] ;
[0043] in, is a Gaussian kernel function, which is used to smooth the mapping relationship between height information and spectral features. and They represent the partial derivatives of the height information, is the mode length of the spectral feature;
[0044] Using the dynamic context fusion mechanism, the large-scale feature map generated by the global path Edge detail feature map generated by local path Combined to generate a fusion feature map :
[0045] ;
[0046] in, and is the adaptive weight coefficient of the dynamic context fusion mechanism, is a scale parameter that controls the spatial correlation between global and local features. The spatial position coordinates considered during fusion;
[0047] S45, generate the fusion feature map The input is into the adaptive multi-scale aggregation network, which adaptively aggregates features of different scales, nonlinearly combines the multi-scale features of the building, and finally outputs the building base feature map.
[0048] Optionally, the S5 specifically includes:
[0049] S51, in the fusion feature map By calculating the gradient information, the preliminary boundary of the building is identified. The result of the gradient calculation is used to locate potential boundary points and generate a preliminary boundary map. ;
[0050] S52, the initially generated building boundary map Refine the boundary points and generate a refined boundary map by iteratively correcting the position of the boundary points , in each iteration, the positions of the boundary points are adjusted according to the local geometric features to describe the outline of the actual building:
[0051] ;
[0052] in, is the adjustment factor, The correction vector representing the boundary point;
[0053] S53, identify the refined boundary map The key boundary segment on the left side adjusts the connection point of the key boundary segment:
[0054] ;
[0055] in, are the adjusted boundary point coordinates, represents the Euclidean distance metric, is the boundary point before adjustment;
[0056] S54, iteratively correcting the geometric shape of the adjusted boundary segment, using an adaptive boundary optimization algorithm, adaptively adjusting the shape of the boundary segment according to the local geometric features of the boundary, and generating the geometric shape of the building through multiple optimizations;
[0057] S55. After boundary identification, refinement, adjustment and optimization, the final building boundary map is generated .
[0058] Optionally, the S6 specifically includes:
[0059] S61. The refined and adjusted building boundary map Convert to binary image , where the pixel value is Represents the building boundary, and the pixel value is Indicates background;
[0060] S62, Binarized Image Perform morphological processing to remove isolated noise points and fill small gaps in the boundaries. Morphological processing includes corrosion and expansion operations to generate processed images. ;
[0061] S63. Image after morphological processing For edge smoothing, a smoothing algorithm based on adaptive edge fusion is used. This algorithm adaptively adjusts the smoothing intensity according to the texture features and neighborhood information of the boundary area, so that the boundary can maintain the necessary details and achieve smooth transition in complex areas:
[0062] ;
[0063] in, represents the image after smoothing. represents the image after morphological processing, is the adaptive weight function;
[0064] S64, smoothed image Perform boundary optimization by adjusting the position of boundary points, optimizing the overall smoothness of the boundary and the rationality of the geometric shape, and generating an optimized boundary map ;
[0065] S65. After morphological processing, smoothing and boundary optimization, the building boundary map generates the final building base binary image. .
[0066] Optionally, the S7 specifically includes:
[0067] S71, binary image of the building base Input into the vectorization algorithm, perform edge tracking on the binary image, and generate preliminary vector boundary lines :
[0068] ;
[0069] in, Indicates the pixel points belonging to the boundary in the binary image. is a parameter, indicating the position on the boundary line;
[0070] S72, the vector boundary line generated initially Perform curvature analysis by calculating the curvature of each vector segment , identify curvature areas, and smooth these areas to generate optimized vector boundary lines :
[0071] ;
[0072] in, and For the boundary line The derivative of position, and is the second-order derivative;
[0073] S73, optimized vector boundary line Perform multi-level vector optimization to generate multi-level vector data suitable for analysis at different scales by simplifying redundant nodes and adjusting the connection of line segments. ;
[0074] S74, vector data after multi-level optimization Make structural adjustments to make it suitable for analysis requirements in different application scenarios. The adjustments include optimizing the topology, connectivity and geometry of the vector data to generate the final vector data. ;
[0075] S75, the vector data after structural adjustment Convert and export formats.
[0076] The beneficial effects of the present invention are:
[0077] The present invention introduces an adaptive multi-scale frequency domain enhancement algorithm, which is significantly innovative in remote sensing image processing. Traditional image enhancement methods often cannot take into account both the details and the overall outline of the building at the same time. The present invention effectively retains the global structural characteristics of the building by performing multi-scale processing of the image in the frequency domain, while significantly enhancing the detail information, especially the edges of the building. This innovation enables the present invention to cope with challenges such as diverse morphologies and lighting changes when dealing with building extraction in complex urban environments, greatly improving the accuracy and robustness of building boundary recognition.
[0078] The present invention constructs a dual-path U-Net model with dynamic context fusion. Different from the traditional single-path feature extraction model, the dual-path structure of the present invention can extract the global large-scale features and local detail features of the building respectively. Through the dynamic context fusion mechanism, the weight distribution of these two types of features is adaptively adjusted to generate a more accurate building base feature map, which significantly improves the building base extraction accuracy in densely packed buildings and complex urban scenes, and overcomes the limitations of traditional models in processing multi-scale information.
[0079] During the vectorization stage of the building base, the present invention adopts a vectorization algorithm based on curvature optimization. By analyzing and optimizing the curvature of the building boundary, not only the boundary line is smoothed, but also multi-level vector optimization technology is used to ensure that the generated vector data can maintain high precision at different scales. This curvature optimization method solves the problem of insufficient precision of traditional vectorization technology when dealing with complex boundaries. The generated vector data is more accurate and concise, and is suitable for analysis needs in different application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0080] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:
[0081] Figure 1 A flowchart of the method for extracting base vectors of high-rise buildings from remote sensing images based on deep learning proposed by the present invention;
[0082] Figure 2 A schematic diagram of processing remote sensing images using the adaptive multi-scale frequency domain enhancement algorithm of the present invention;
[0083] Figure 3 Schematic diagram of feature extraction using a dual-path U-Net model with dynamic context fusion in the present invention. DETAILED DESCRIPTION
[0084] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.
[0085] refer to Figure 1-3 , a method for extracting base vectors of high-rise buildings from remote sensing images based on deep learning, comprising the following steps:
[0086] S1. Obtain remote sensing image data, use adaptive multi-scale frequency domain enhancement algorithm, perform multi-scale processing on the image in the frequency domain, and highlight the edges and detail features of the building;
[0087] S2. Based on the pre-processed image data, automatically label the high-rise building sample set, generate labeled training data, and perform data expansion processing on it;
[0088] S3, Framework based, using Feature extraction backbone network, building dual paths with dynamic context fusion Model;
[0089] S4. Use dual paths The model extracts features from the input image, captures large-scale features of buildings through global paths, and captures edge details of buildings through local paths. Based on the extracted top contours of buildings and combined with height information, the model calculates the preliminary position of the building base through geometric deduction, and uses a dynamic context fusion mechanism to combine large-scale features with edge detail information to generate a building base feature map.
[0090] S5. Based on the fused building base feature map, multi-stage boundary refinement and segmentation are performed to identify and refine the boundaries of buildings, especially in areas covered by shadows and with complex shapes, by adjusting the connection points of boundary segments and iteratively correcting the building geometry;
[0091] S6, processing the segmentation results, optimizing the outline of the building base, removing unnecessary details, improving the smoothness and consistency of the overall boundary, and forming a binary image of the building base;
[0092] S7. The binary image of the building base is converted into vector data through a vectorization algorithm based on curvature optimization, and multi-level vector optimization is performed. The structure of the vector data is adjusted to make it suitable for analysis and application at different scales, and finally the vector data of the building base is generated.
[0093] In this implementation, S1 specifically includes:
[0094] S11, acquiring remote sensing image data, the image data including satellite images and aerial images of multi-spectral and panchromatic bands, and converting the remote sensing image data into frequency domain for multi-scale analysis;
[0095] S12, perform multi-scale decomposition on the image data converted to the frequency domain, and use a multi-resolution sub-band decomposition algorithm to convert the image data Decomposed into sub-bands of different resolutions and frequencies, each sub-band contains frequency domain information at different scales, among which the low-frequency components Preserve global features and high-frequency components Reflect the edges and details of buildings:
[0096] ;
[0097] in, is the subband coefficient, is the corresponding subband basis function;
[0098] S13, at different scales, for high frequency components Perform adaptive enhancement processing through adaptive enhancement function Adaptively adjust the intensity of high-frequency components according to local image features to highlight the edges and details of buildings:
[0099] ;
[0100] in, and According to different scales Adaptive adjustment parameters;
[0101] S14. At each scale, the enhanced high-frequency component is recombined with the low-frequency component to generate enhanced image data. ;
[0102] S15. Convert the enhanced image data of each scale from the frequency domain back to the spatial domain, perform multi-scale fusion, and generate integrated multi-scale enhanced remote sensing image data .
[0103] In this implementation, S2 specifically includes:
[0104] S21. Obtain remote sensing image data after multi-scale frequency domain enhancement processing , the image data contains the edges and detail features of the building;
[0105] S22. In the image data The spectral feature analysis of the sample set data of the building is performed to extract the spectral information of each band, and the attributes of the building samples are annotated based on the analysis results to generate the spectral feature annotation data of the building. , including building top outlines and other relevant attribute information:
[0106] ;
[0107] in, The corresponding band The weight of Indicates in band The image data on Indicates the total number of bands;
[0108] S23. Labeling data based on spectral features , extract the top outline information of the building , the top profile information is obtained by analyzing the spectral feature data and edge detection;
[0109] S24, the marked building spectral characteristic data The corresponding image data Combined to form a preliminary annotated training data set :
[0110] ;
[0111] in, represents the index of the data sample, is the number of samples initially labeled;
[0112] S25. Preliminary labeled training data set Data expansion is performed, and a data expansion method based on generative adversarial networks is used to generate diverse training samples. The expanded training data set is represented as , and used for model training:
[0113] ;
[0114] in, is the coordinate after the expansion operation, is the number of samples after expansion.
[0115] In this implementation manner, the S4 specifically includes:
[0116] S41, the training image dataset after data expansion Input to the dual-path U-Net model, where each image data sample As input to the model, the global path and the local path simultaneously receive input image data;
[0117] S42, input image data through the global path For large-scale feature extraction, the global path captures the overall structural characteristics of the building through a series of convolutional layers and activation functions to generate a large-scale feature map ;
[0118] S43, input image data through local path Perform edge detail feature extraction. The local path extracts the edge details of the building through the receptive field and convolution operation to generate an edge detail feature map ;
[0119] S44, based on the extracted building top outline and spectral characteristic data , combined with the building height information , the preliminary position of the building base is calculated by geometric deduction and spectral feature correlation analysis:
[0120] ;
[0121] in, is a Gaussian kernel function, which is used to smooth the mapping relationship between height information and spectral features. and They represent the partial derivatives of the height information, is the mode length of the spectral feature;
[0122] Using the dynamic context fusion mechanism, the large-scale feature map generated by the global path Edge detail feature map generated by local path Combined to generate a fusion feature map :
[0123] ;
[0124] in, and is the adaptive weight coefficient of the dynamic context fusion mechanism, is a scale parameter that controls the spatial correlation between global and local features. The spatial position coordinates considered during fusion;
[0125] S45, generate the fusion feature map The input is into the adaptive multi-scale aggregation network, which adaptively aggregates features of different scales, nonlinearly combines the multi-scale features of the building, and finally outputs the building base feature map.
[0126] In this implementation manner, S5 specifically includes:
[0127] S51, in the fusion feature map By calculating the gradient information, the preliminary boundary of the building is identified. The result of the gradient calculation is used to locate potential boundary points and generate a preliminary boundary map. ;
[0128] S52, the initially generated building boundary map Refine the boundary points and generate a refined boundary map by iteratively correcting the position of the boundary points , in each iteration, the positions of the boundary points are adjusted according to the local geometric features to describe the outline of the actual building:
[0129] ;
[0130] in, is the adjustment factor, The correction vector representing the boundary point;
[0131] S53, identify the refined boundary map The key boundary segment on the left side adjusts the connection point of the key boundary segment:
[0132] ;
[0133] in, are the adjusted boundary point coordinates, represents the Euclidean distance metric, is the boundary point before adjustment;
[0134] S54, iteratively correcting the geometric shape of the adjusted boundary segment, using an adaptive boundary optimization algorithm, adaptively adjusting the shape of the boundary segment according to the local geometric features of the boundary, and generating the geometric shape of the building through multiple optimizations;
[0135] S55. After boundary identification, refinement, adjustment and optimization, the final building boundary map is generated .
[0136] In this implementation manner, S6 specifically includes:
[0137] S61. The refined and adjusted building boundary map Convert to binary image , where the pixel value is Represents the building boundary, and the pixel value is Indicates background;
[0138] S62, Binarized Image Perform morphological processing to remove isolated noise points and fill small gaps in the boundaries. Morphological processing includes corrosion and expansion operations to generate processed images. ;
[0139] S63. Image after morphological processing For edge smoothing, a smoothing algorithm based on adaptive edge fusion is used. This algorithm adaptively adjusts the smoothing intensity according to the texture features and neighborhood information of the boundary area, so that the boundary can maintain the necessary details and achieve smooth transition in complex areas:
[0140] ;
[0141] in, represents the image after smoothing. represents the image after morphological processing, is the adaptive weight function;
[0142] S64, smoothed image Perform boundary optimization by adjusting the position of boundary points, optimizing the overall smoothness of the boundary and the rationality of the geometric shape, and generating an optimized boundary map ;
[0143] S65. After morphological processing, smoothing and boundary optimization, the building boundary map generates the final building base binary image. .
[0144] In this implementation manner, the S7 specifically includes:
[0145] S71, binary image of the building base Input into the vectorization algorithm, perform edge tracking on the binary image, and generate preliminary vector boundary lines :
[0146] ;
[0147] in, Indicates the pixel points belonging to the boundary in the binary image. is a parameter, indicating the position on the boundary line;
[0148] S72, the vector boundary line generated initially Perform curvature analysis by calculating the curvature of each vector segment , identify curvature areas, and smooth these areas to generate optimized vector boundary lines :
[0149] ;
[0150] in, and For the boundary line The derivative of position, and is the second-order derivative;
[0151] S73, optimized vector boundary line Perform multi-level vector optimization to generate multi-level vector data suitable for analysis at different scales by simplifying redundant nodes and adjusting the connection of line segments. ;
[0152] S74, vector data after multi-level optimization Make structural adjustments to make it suitable for analysis requirements in different application scenarios. The adjustments include optimizing the topology, connectivity and geometry of the vector data to generate the final vector data. ;
[0153] S75, the vector data after structural adjustment Convert and export formats.
[0154] Embodiment 1:
[0155] In order to verify the feasibility of the present invention in implementation, the present invention is applied to disaster emergency response and urban planning projects in a coastal city. There is a need to accurately extract the bases of high-rise buildings. Due to the complex terrain, dense buildings and various shapes in the city, traditional building base extraction methods are difficult to meet the requirements of accuracy and efficiency in such scenarios. Especially after disasters such as typhoons and earthquakes, quickly and accurately obtaining building base information is crucial for assessing the degree of building damage and conducting subsequent urban planning. Existing technologies often show problems of insufficient accuracy and low processing efficiency when dealing with such complex environments, and it is difficult to provide high-quality data support for disaster response in a short period of time.
[0156] In response to this actual demand, the present invention has been applied and verified in the disaster emergency response scenario of the city. We have obtained high-resolution remote sensing image data of the city, which covers the core area of the city. The image resolution reaches 0.3 meters, and can clearly present the details and edges of high-rise buildings. Since the city is located in the coastal area, typhoons occur frequently, and the building shapes are complex, there are a large number of occlusions, shadows and varying degrees of illumination changes in the image data. We first performed adaptive multi-scale frequency domain enhancement processing on the remote sensing image data, and performed multi-scale processing on the image in the frequency domain to effectively highlight the edges and detail features of the buildings. Through this processing step, we have significantly improved the visibility of the building boundaries, especially in areas with severe shadows and occlusions, the building outlines have been significantly enhanced.
[0157] After obtaining the enhanced image data, we used the dual-path U-Net model with dynamic context fusion to extract features from the image. The model captures the large-scale features of the building through the global path and the detailed features of the building through the local path, and finally integrates these two types of features organically through the dynamic context fusion mechanism. Thanks to the advantages of the dual-path U-Net model, we can effectively distinguish the subtle differences between buildings and the surrounding complex background in a complex urban environment, and ensure the extraction accuracy of the building base. During the training of the model, we used a large amount of labeled data including high-rise buildings and low-rise buildings. The data covers different building forms and urban areas, which gives the model good generalization ability.
[0158] After feature extraction, we performed boundary recognition and multi-stage refined segmentation on the building base. Especially when dealing with occluded and complex-shaped buildings, through multiple rounds of iterative correction, we effectively eliminated the boundary misalignment and discontinuity problems common in traditional methods. Through this refined processing, we generated a binary image of the building base, providing high-quality basic data for the subsequent vectorization step.
[0159] We applied a vectorization algorithm based on curvature optimization to convert the binary building base image into vector data. The algorithm smoothed the boundary lines, eliminated noise and irregularities through precise analysis of the building boundary curvature, and further performed multi-level vector optimization. The optimized vector data is not only highly accurate, but also able to maintain consistency in analysis and applications at different scales, which is particularly important for rapid decision-making in disaster emergency response; for example, in the assessment of building damage after a typhoon, we were able to use these high-precision vector data to quickly identify severely affected building areas and provide an accurate basis for rescue and reconstruction work.
[0160] In order to verify the effectiveness of the present invention, we applied the method to multiple core areas of the city and compared it with the traditional base extraction method. In practical applications, we selected 50 representative areas where the buildings in these areas are complex in shape and densely distributed. The experimental results show that the extraction accuracy and processing efficiency are significantly better than the traditional methods. In terms of accuracy, the degree of consistency between the extracted building base and the actual surveying and mapping data is more than 95%, while the average accuracy of the traditional method is only 85%; in terms of processing efficiency, the high-resolution image data of 50 areas can be processed within 3 hours, while the traditional method takes more than 8 hours to complete the same workload.
[0161] In the building damage assessment after the typhoon, we successfully identified hundreds of damaged buildings through the vector data generated by the present invention, providing key support for the rescue work. Especially in a certain area, due to the high-precision base extraction of the present invention, we were able to quickly assess the risk of building collapse in the area, helping relevant departments to quickly formulate evacuation and rescue plans, avoiding the occurrence of secondary disasters.
[0162] Table 1 Comparison of extraction accuracy based on the present invention and traditional methods
[0163]
[0164] Table 2 Comparison of treatment efficiency based on the present invention and traditional methods
[0165]
[0166] Table 3 Application effect in post-disaster building damage assessment
[0167]
[0168] It can be seen from the above table that in terms of extraction accuracy, the actual mapping data consistency of the present invention reaches an average of 95.4%, which is significantly higher than the 85.6% of the traditional method. This shows that the present invention can more accurately capture the true outline of the building and reduce errors and deviations when dealing with building base extraction in complex urban environments, especially in areas with dense buildings and complex shapes. In terms of processing efficiency, the present invention takes an average processing time of 3 hours and 25 minutes when processing image data of the same size, while the traditional method takes 7 hours and 40 minutes, which is almost twice the difference in processing time. This shows that the present invention greatly improves the processing speed by automating and optimizing the processing flow, and can complete the processing of large-scale data in a short time. This is particularly important for application scenarios that require rapid response, such as disaster emergency response. In terms of the application effect in post-disaster building damage assessment, it is not only significantly better than the traditional method in the number of damaged buildings identified, but also greatly shortens the time for key rescue decision support. For example, in one area, the present invention shortens the key rescue decision support time from 10 hours of the traditional method to 4 hours. This time advantage may greatly improve rescue efficiency and reduce loss of life and property in emergency situations.
[0169] In summary, through the application of the present invention in actual disaster emergency response and urban planning, we not only solve the problems of insufficient extraction accuracy and low processing efficiency of traditional methods in complex environments, but also provide efficient and accurate building base data, which greatly improves the decision-making efficiency of emergency response and urban planning; experimental data proves that the application of the present invention in the extraction of high-rise building bases has significant advantages and can provide strong technical support for urban management and disaster emergency response.
[0170] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.
Claims
1. A method for extracting base vectors of high-rise buildings from remote sensing images based on deep learning, characterized in that: The steps include: S1. Obtain remote sensing image data, use adaptive multi-scale frequency domain enhancement algorithm, perform multi-scale processing on the image in the frequency domain, and highlight the edges and detail features of the building; S2. Based on the pre-processed image data, automatically label the high-rise building sample set, generate labeled training data, and perform data expansion processing on it; S3, Framework based, using Feature extraction backbone network, building dual paths with dynamic context fusion Model; S4. Use dual paths The model extracts features from the input image, captures large-scale features of buildings through global paths, and captures edge details of buildings through local paths. Based on the extracted top contours of buildings and combined with height information, the model calculates the preliminary position of the building base through geometric deduction, and uses a dynamic context fusion mechanism to combine large-scale features with edge detail information to generate a building base feature map. S5. Based on the fused building base feature map, multi-stage boundary refinement and segmentation are performed to identify and refine the boundaries of the buildings. In areas covered by shadows and complex shapes, the connection points of the boundary segments are adjusted and the building geometry is iteratively corrected. The S5 specifically includes: S51, in the fusion feature map By calculating the gradient information, the preliminary boundary of the building is identified. The result of the gradient calculation is used to locate potential boundary points and generate a preliminary boundary map. ; S52, the initially generated building boundary map Refine the boundary points and generate a refined boundary map by iteratively correcting the position of the boundary points , in each iteration, the positions of the boundary points are adjusted according to the local geometric features to describe the outline of the actual building: ; in, is the adjustment factor, The correction vector representing the boundary point; S53, identify the refined boundary map The key boundary segment on the left side adjusts the connection point of the key boundary segment: ; in, are the adjusted boundary point coordinates, represents the Euclidean distance metric, is the boundary point before adjustment; S54, iteratively correcting the geometric shape of the adjusted boundary segment, using an adaptive boundary optimization algorithm, adaptively adjusting the shape of the boundary segment according to the local geometric features of the boundary, and generating the geometric shape of the building through multiple optimizations; S55. After boundary identification, refinement, adjustment and optimization, the final building boundary map is generated ; S6, processing the segmentation results, optimizing the outline of the building base, removing unnecessary details, improving the smoothness and consistency of the overall boundary, and forming a binary image of the building base; S7. The binary image of the building base is converted into vector data through a vectorization algorithm based on curvature optimization, and multi-level vector optimization is performed. The structure of the vector data is adjusted to make it suitable for analysis and application at different scales, and finally the vector data of the building base is generated.
2. The method for extracting base vectors of high-rise buildings from remote sensing images based on deep learning according to claim 1, characterized in that: The S1 specifically includes: S11, acquiring remote sensing image data, the image data including satellite images and aerial images of multi-spectral and panchromatic bands, and converting the remote sensing image data into frequency domain for multi-scale analysis; S12, perform multi-scale decomposition on the image data converted to the frequency domain, and use a multi-resolution sub-band decomposition algorithm to convert the image data Decomposed into sub-bands of different resolutions and frequencies, each sub-band contains frequency domain information at different scales, among which the low-frequency components Preserve global features and high-frequency components Reflect the edges and details of buildings: ; in, is the subband coefficient, is the corresponding subband basis function; S13, at different scales, for high frequency components Perform adaptive enhancement processing through adaptive enhancement function Adaptively adjust the intensity of high-frequency components according to local image features to highlight the edges and details of buildings: ; in, and According to different scales Adaptive adjustment parameters; S14. At each scale, the enhanced high-frequency component is recombined with the low-frequency component to generate enhanced image data. ; S15. Convert the enhanced image data of each scale from the frequency domain back to the spatial domain, perform multi-scale fusion, and generate integrated multi-scale enhanced remote sensing image data .
3. The method for extracting base vectors of high-rise buildings from remote sensing images based on deep learning according to claim 1, characterized in that: The S2 specifically includes: S21. Obtain remote sensing image data after multi-scale frequency domain enhancement processing , the image data contains the edges and detail features of the building; S22. In the image data The spectral feature analysis of the sample set data of the building is performed to extract the spectral information of each band, and the attributes of the building samples are annotated based on the analysis results to generate the spectral feature annotation data of the building. , including building top outlines and other relevant attribute information: ; in, The corresponding band The weight of Indicates in band The image data on Indicates the total number of bands; S23. Labeling data based on spectral features , extract the top outline information of the building , the top profile information is obtained by analyzing the spectral feature data and edge detection; S24, the marked building spectral characteristic data The corresponding image data Combined to form a preliminary annotated training data set : ; in, represents the index of the data sample, is the number of samples initially labeled; S25. Preliminary labeled training data set Data expansion is performed, and a data expansion method based on generative adversarial networks is used to generate diverse training samples. The expanded training data set is represented as , and used for model training: ; in, is the coordinate after the expansion operation, is the number of samples after expansion.
4. The method for extracting base vectors of high-rise buildings from remote sensing images based on deep learning according to claim 1, characterized in that: The S4 specifically includes: S41, the training image dataset after data expansion Input to the dual-path U-Net model, where each image data sample As input to the model, the global path and the local path simultaneously receive input image data; S42, input image data through the global path For large-scale feature extraction, the global path captures the overall structural characteristics of the building through a series of convolutional layers and activation functions to generate a large-scale feature map ; S43, input image data through local path Perform edge detail feature extraction. The local path extracts the edge details of the building through the receptive field and convolution operation to generate an edge detail feature map ; S44, based on the extracted building top outline and spectral characteristic data , combined with the building height information , the preliminary position of the building base is calculated by geometric deduction and spectral feature correlation analysis: ; in, is a Gaussian kernel function, which is used to smooth the mapping relationship between height information and spectral features. and They represent the partial derivatives of the height information, is the mode length of the spectral feature; Using the dynamic context fusion mechanism, the large-scale feature map generated by the global path Edge detail feature map generated by local path Combined to generate a fusion feature map : ; in, and is the adaptive weight coefficient of the dynamic context fusion mechanism, is a scale parameter that controls the spatial correlation between global and local features. The spatial position coordinates considered during fusion; S45, generate the fusion feature map The input is into the adaptive multi-scale aggregation network, which adaptively aggregates features of different scales, nonlinearly combines the multi-scale features of the building, and finally outputs the building base feature map.
5. The method for extracting base vectors of high-rise buildings from remote sensing images based on deep learning according to claim 1, characterized in that: The S6 specifically includes: S61. The refined and adjusted building boundary map Convert to binary image , where the pixel value is Represents the building boundary, and the pixel value is Indicates background; S62, Binarized Image Perform morphological processing to remove isolated noise points and fill small gaps in the boundaries. Morphological processing includes corrosion and expansion operations to generate processed images. ; S63. Image after morphological processing For edge smoothing, a smoothing algorithm based on adaptive edge fusion is used. This algorithm adaptively adjusts the smoothing intensity according to the texture features and neighborhood information of the boundary area, so that the boundary can maintain the necessary details and achieve smooth transition in complex areas: ; in, represents the image after smoothing. represents the image after morphological processing, is the adaptive weight function; S64, smoothed image Perform boundary optimization by adjusting the position of boundary points, optimizing the overall smoothness of the boundary and the rationality of the geometric shape, and generating an optimized boundary map ; S65. After morphological processing, smoothing and boundary optimization, the building boundary map generates the final building base binary image. .
6. The method for extracting base vectors of high-rise buildings from remote sensing images based on deep learning according to claim 1, characterized in that: The S7 specifically includes: S71, binary image of the building base Input into the vectorization algorithm, perform edge tracking on the binary image, and generate preliminary vector boundary lines : ; in, Indicates the pixel points belonging to the boundary in the binary image. is a parameter, indicating the position on the boundary line; S72, the vector boundary line generated initially Perform curvature analysis by calculating the curvature of each vector segment , identify curvature areas, and smooth these areas to generate optimized vector boundary lines : ; in, and For the boundary line The derivative of position, and is the second-order derivative; S73, optimized vector boundary line Perform multi-level vector optimization to generate multi-level vector data suitable for analysis at different scales by simplifying redundant nodes and adjusting the connection of line segments. ; S74, vector data after multi-level optimization Make structural adjustments to make it suitable for analysis requirements in different application scenarios. The adjustments include optimizing the topology, connectivity and geometry of the vector data to generate the final vector data. ; S75, the vector data after structural adjustment Convert and export formats.
Citation Information
Patent Citations
Deep learning-based orthoimage high-rise building base vector extraction method
CN113139453A
Remote sensing image building extraction and contour optimization method based on deep learning
CN113516135A
Building base contour refined extraction method based on live-action three-dimensional model
CN114417489A
Remote sensing image building extraction method based on deep learning and contour regularization
CN115984693A
