CN109903304B - Automatic building contour extraction algorithm based on convolutional neural network and polygon regularization - Google Patents
Automatic building contour extraction algorithm based on convolutional neural network and polygon regularization Download PDFInfo
- Publication number
- CN109903304B CN109903304B CN201910136835.2A CN201910136835A CN109903304B CN 109903304 B CN109903304 B CN 109903304B CN 201910136835 A CN201910136835 A CN 201910136835A CN 109903304 B CN109903304 B CN 109903304B
- Authority
- CN
- China
- Prior art keywords
- building
- polygon
- neural network
- regularization
- edge
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Abstract
The invention discloses a building outline automatic extraction algorithm based on a convolutional neural network and polygon regularization, which comprises the following steps: constructing a sample library according to the existing image and building coverage vector files; constructing a multi-scale fusion full convolution neural network, training the multi-scale fusion full convolution neural network through a sample library, and predicting the remote sensing image by using a trained network model to obtain a segmentation result of the remote sensing image surface building coverage; initializing the edge of the building based on the semantic segmentation result of the building, and obtaining an initial vector polygon; eliminating wrong polygons and wrong edges and nodes of the polygons by using a coarse adjustment algorithm; and utilizing a regularization algorithm to regularize the vector polygon to obtain a regular building vector edge. The multi-scale fusion full convolution neural network has strong scale robustness, and the regularization algorithm can adapt to vector edges under various conditions, thereby greatly reducing the workload of manually drawing the building edges.
Description
Technical Field
The invention relates to a deep learning method for remote sensing image building extraction and a regularization algorithm of a building polygon outline, which can be used for remote sensing image building extraction, building vector edge generation, building change detection and the like.
Background
The automatic extraction of the remote sensing image buildings has very important significance in the applications of urban planning, population estimation, map making and updating and the like. Traditionally, the main work of extracting buildings from aerial/aerospace imagery has focused on: an appropriate feature is empirically designed to express "what is a building" and corresponding features are created for automatic identification and extraction of buildings. Common metrics include pixel, spectrum, length, edge, shape, texture, shading, height, semantic, and the like. These indicators can vary significantly with season, lighting, atmospheric conditions, sensor quality, scale, building style, and environment. Therefore, this method of empirically designing features often can only process specific data, and cannot be truly automated. The convolutional neural network in deep learning shows strong performance in image retrieval, image classification and target detection. The convolutional neural network can automatically learn a multi-layer feature expression and map the original input image into a univariate or multivariate label. This ability to learn features by itself surpasses and gradually replaces the traditional approach of designing features through manual experience. Building extraction is not only a classification and semantic segmentation problem, but also a target detection and instance segmentation problem. The object extracted by the building is not concerned with whether a certain pixel is a building, but is more concerned with the number, position and shape of the building. One of the major and burdensome tasks for mappers around the world is to manually outline architectural vector maps on aerial/aerospace imagery, thereby producing various topographic and thematic maps. Therefore, vector data extraction research for buildings is of great importance.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a neural network with strong scale robustness, which can adapt to the extraction of remote sensing image buildings with different scales, and adds prior knowledge to the building edge obtained by semantic segmentation to carry out regularization processing to obtain a regular building polygon with high quality.
The technical scheme adopted for realizing the aim of the invention is that the building outline automatic extraction algorithm based on the convolutional neural network and the polygon regularization comprises the following steps:
step 2, constructing a multi-scale fusion full convolution neural network, training the multi-scale fusion full convolution neural network through a sample library, and predicting the remote sensing image by using a trained network model to obtain a segmentation result of the remote sensing image surface building coverage;
step 3, initializing the edge of the building based on the semantic segmentation result of the building, and obtaining an initial vector polygon;
step 4, eliminating wrong polygons and wrong edges and nodes of the polygons by using a coarse adjustment algorithm;
and 5, regularizing the vector polygon by using a regularization algorithm to obtain a regular building vector edge.
Further, the multi-scale fusion full convolution neural network in the step 2 comprises 3 parts of encoding (encoding stage), decoding (decoding stage) and outputting (output); wherein the coding part consists of 5 convolutional layers (Convolution Layer), 4 Max Pooling layers (Max Pooling Layer); the decoding part consists of 4 convolution layers and 4 Deconvolution layers (Deconvolition layers); the output section consists of 4 sub-outputs and 1 main output.
Further, the first two convolutional layers of the coding part are composed of two groups of continuously stacked convolutions, modified Linear units (relus) and a Batch Normalization layer (BN); the last three convolutional layers consist of three groups of continuously stacked convolutions, modified Linear units (relus) and two Batch Normalization layers (BN); the first three convolutional layers and the last convolutional layer of the decoding part are respectively composed of 4 groups and 3 groups of Convolution (Convolution) and modified Linear Unit (ReLU) which are stacked continuously; each decoding part finally obtains a sub-output, the main output obtains a 4-channel characteristic diagram through the final 1-channel characteristic diagram on each scale of the serial decoding part, and the final prediction result is obtained through the activation of a Sigmoid function.
Further, in step 3, edge initialization is performed by using a Douglas-Peucker algorithm, and different tolerance thresholds D are set according to the number of pixels of each connected domain to obtain an initial vector polygon.
Further, the specific implementation manner of step 4 is as follows,
41) the sides of the building should have a certain length: rejecting polygons smaller than TdThe edge of (1);
42) building corners cannot be too sharp or too shallow: rejecting angles smaller than alpha and larger than beta;
43) the building should have a certain area: deleting polygons with the areas smaller than S;
44) the building nodes should not be excessive: the ratio of the rejection area to the number of the nodes is less than TaThe ratio of the sum perimeter to the number of the nodes is less than TpIs used for the polygon of (1).
Further, the specific implementation manner of step 5 is as follows,
51) determining the main direction of the building: setting different side length thresholds W according to the area of the building, wherein the sides with the length larger than W are long sides, the rest are short sides, and reducing W in sequence for the building with the smaller area and the length which cannot find at least one long side until at least one long side is found; firstly, taking the longest side as the initial main direction, adding the longest side into a main direction list, comparing all the other long sides with the main direction list, and if the included angle between the long side and all the main directions is more than deltaminAnd is less than deltamaxAdding the direction of the long side into the main direction queue;
52) adjusting the edges of the polygon according to the main direction: firstly, the long edge is adjusted, the long edge is compared with the main direction and the vertical direction of the main direction to obtain the minimum included angle, and if the included angle is less than deltaminRotating the long edge to a direction parallel or perpendicular to the main direction with the minimum included angle by taking the middle point as a rotation center; then, the short edge is adjusted, the short edge is compared with the main direction and the vertical direction of the main direction to obtain the minimum included angle, and if the included angle is smaller than TsRotating the short side to a parallel or vertical direction of the minimum included angle corresponding to the main direction by taking the middle point as a rotation center;
53) redundant parallel line culling and adding vertical lines: two continuous edges possibly exist in the adjusted edge and are parallel lines, different distance thresholds d are set according to the area of the building, if the distance between the two continuous parallel lines is larger than d, a perpendicular line is added between the two parallel lines, otherwise, the edge with the smaller length is removed;
54) and sequentially calculating intersection points between two continuous edges to serve as nodes of the vector polygon, and sequentially connecting all the nodes to obtain the adjusted vector polygon.
Further, the specific implementation manner of step 1 is as follows,
firstly, converting a building coverage vector file corresponding to an image into a binary raster image, wherein the pixel value of the raster image is a label of an image pixel, wherein 0 represents that a background 1 is a building, and cutting an original image and a corresponding raster label into tiles to construct a sample library.
The invention has the following advantages: the multi-scale fusion full convolution neural network has strong scale robustness, can adapt to the extraction of remote sensing image buildings with different scales, and can continuously and iteratively optimize; the regularization algorithm can adapt to vector edges under various conditions, and the workload of manually drawing the building edges is greatly reduced.
Drawings
FIG. 1 is a flow chart of the sample library construction of the present invention.
Fig. 2 is a schematic diagram of the neural network structure of the present invention.
FIG. 3 is a data range diagram of an embodiment of the present invention.
FIG. 4 is a schematic diagram of a data building in accordance with an embodiment of the present invention.
FIG. 5 is a graph showing the results of an example of the present invention.
Detailed Description
The technical scheme of the invention is further specifically described by the following embodiments and the accompanying drawings.
The invention provides a building outline automatic extraction algorithm based on a convolutional neural network and polygon regularization, which comprises the following steps: step 1, firstly, converting a building coverage vector file corresponding to an image into a binary raster image, wherein pixel values of the raster image are labels of image pixels, 0 represents that a background 1 is a building, and cutting an original image and a corresponding raster label into tiles to construct a sample library; step 2, training a Multi-scale integration Full convolution Neural Network (MA-FCN) to learn the characteristics of the building in the remote sensing image; after the training of the network model is finished, predicting the remote sensing image by using the trained network model to obtain a segmentation result of the remote sensing image surface building coverage; step 3, initializing the edge of the building by using a Douglas-Peucker algorithm based on the semantic segmentation result of the building; step 4, eliminating wrong polygons and wrong edges and nodes of the polygons by using a coarse adjustment algorithm; and 5, regularizing the vector polygon by using a regularization algorithm to obtain a regular building vector edge.
The multi-scale fusion full convolution neural network comprises 3 parts of encoding (encoding stage), decoding (decoding stage) and outputting (output). The coding part consists of 5 convolutional layers (Convolution Layer) and 4 Max Pooling layers (Max Pooling Layer); the decoding part consists of 4 convolution layers and 4 Deconvolution layers (Deconvolition layers); the output section consists of 4 sub-outputs and 1 main output.
The first two convolutional layers of the coding part consist of two groups of convolution which are stacked continuously, a modified Linear Unit (ReLU) and a Batch Normalization layer (BN); the last three convolutional layers consist of three sets of successively stacked convolutions, modified Linear units (relus) and two Batch Normalization layers (BN). The convolution kernel sizes are all 3 × 3, and the convolution step sizes are all 1.
The maximum pooling layer step size of the encoding part is 2 x 2, and after passing through the pooling layer, the height and width of the output feature map become one half of the input.
The input of each convolution layer of the decoding part is the series connection of the characteristic diagram obtained after deconvolution and the corresponding size characteristic diagram of the coding part.
The first three convolutional layers and the last convolutional layer of the above decoding part are respectively composed of 4 and 3 groups of sequentially stacked convolutions (Convolution), modified Linear units (ReLU). The convolution kernel sizes are all 3 × 3, and the convolution step sizes are all 1. Each decoding section finally results in a sub-output. And activating the characteristic diagram of the 1 channel finally obtained by the convolutional layer by a Sigmoid function to obtain a sub-prediction diagram with a corresponding scale.
And the main output part obtains a 4-channel characteristic diagram through the last 1-channel characteristic diagram on each scale of the serial decoding part, and obtains the final prediction result through the activation of a Sigmoid function.
In the step 3, after a building semantic segmentation result is obtained, edge initialization is carried out by using a Douglas-Peucker algorithm, different tolerance thresholds D are set according to the number of pixels of each connected domain, and an initial vector polygon is obtained.
In step 4, the initial polygon is adjusted by using a coarse adjustment algorithm:
41) the sides of the building should have a certain length: rejecting polygons smaller than TdThe edge of (1);
42) building corners cannot be too sharp or too shallow: rejecting angles smaller than alpha and larger than beta;
43) the building should have a certain area: deleting polygons with the area smaller than S;
44) the building nodes should not be excessive: the ratio of the rejection area to the number of the nodes is less than TaThe ratio of the sum perimeter to the number of the nodes is less than TpIs used for the polygon of (1).
In step 5, fine adjustment is performed on the coarsely adjusted polygon:
51) the main direction of the building is determined. Different side length thresholds W are set according to the area of the building, the side with the length larger than W is a long side, and the rest are short sides. And sequentially reducing W for the buildings with smaller areas and at least one long edge which cannot be found until at least one long edge is found. First, the longest side is taken as the initial main direction and added into the main direction list. Comparing all the other long edges with the main direction list, and if the included angle between the long edge and all the main directions is larger than deltaminAnd is less than deltamaxThen the direction of the long edge is added to the main direction queue.
52) The edges of the polygon are adjusted according to the main direction. The long side is first adjusted. And comparing the long edge with the main direction and the vertical direction of the main direction to obtain the minimum included angle. If the angle is less than deltaminThen, the long side is rotated to the parallel or vertical direction of the minimum included angle corresponding to the main direction by taking the middle point as the rotation center. Then the short edge adjustment is performed. And comparing the short edge with the main direction and the vertical direction of the main direction to obtain the minimum included angle. If the angle is less than TsThen the short side is rotated to the parallel or vertical direction of the minimum included angle corresponding to the main direction by taking the middle point as the rotation center.
53) Redundant parallel lines are culled and added to the vertical lines. The adjusted edge may have two continuous sides which are parallel lines. Different distance thresholds d are set according to the area of the building, and if the distance between the continuous parallel lines is larger than d, a perpendicular line is added between the two parallel lines. Otherwise, the side with smaller length is removed.
54) And sequentially calculating the intersection points between two continuous edges as the nodes of the vector polygon. And connecting all the nodes in sequence to obtain the adjusted vector polygon.
Example (b):
referring to fig. 1 and 2, the invention learns the building features in the high-resolution remote sensing image by using a Multi-scale fusion Full convolution Neural Network (MA-FCN), and then performs pixel-level prediction on the building coverage of the remote sensing image. In order to train the neural network model, training samples are first acquired, and fig. 1 shows a process of constructing a training sample library. Firstly, the remote sensing image is cut and resampled to obtain an image range with proper resolution and building coverage data. And rasterizing the corresponding building vector data in the image range to make the building vector data consistent with the image resolution. Finally, combining with factors such as computer performance and ground feature size, the remote sensing image and the corresponding label data are divided into sample blocks with proper size (such as 256 × 256 pixels or 512 × 512 pixels). FIG. 3 is a sample library data ensemble in which the dashed box image is used for training and the solid box image is used for testing. Fig. 4 is a partial building example.
After the training data are obtained, iterative training is carried out on the neural network until the model is optimal. And after the model training is finished, cutting the remote sensing image to be classified into image blocks with the sizes consistent with those of the training samples, and extracting the building from the image blocks by using the trained model to obtain the building segmentation result of the remote sensing image. And finally, splicing the prediction results of all the image blocks to obtain a building segmentation graph of a complete image.
After the semantic segmentation graph of the building is obtained, appropriate parameters are set according to the graph, and the Douglas-Peucker algorithm is used for initializing the edge vector of the building. And setting parameters to perform coarse adjustment on the initialized vector polygon. And setting appropriate parameters to finely adjust the coarsely adjusted polygon to obtain a final regularized vector result. On the test data, letSetting alpha pi/6, beta pi/18, S20, Ta＝1,Tp＝1.5,Td＝0.5,w＝0.2,δmin＝π/12,δmax＝5π/12,TsPi/4, D is set to:
wherein N is the number of pixels; w and d are:
where S is the building area.
Fig. 5 is a final result example, where the first column is an original image and a building edge example, the second column is a segmentation result obtained by multi-scale fusion of the image with a full convolution neural network, the third column is an initial test edge obtained by a Douglas-Peucker algorithm, and the fourth column is a regularization edge obtained by a regularization algorithm.
The specific embodiments described herein are merely illustrative of the spirit of the invention. Various modifications or additions may be made to the described embodiments or alternatives may be employed by those skilled in the art without departing from the spirit or ambit of the invention as defined in the appended claims.
Claims (6)
1. An automatic building contour extraction algorithm based on a convolutional neural network and polygon regularization is characterized by comprising the following steps:
step 1, constructing a sample library according to the existing image and building coverage vector file;
step 2, constructing a multi-scale fusion full convolution neural network, training the multi-scale fusion full convolution neural network through a sample library, and predicting the remote sensing image by using a trained network model to obtain a segmentation result of the remote sensing image surface building coverage;
step 3, initializing the edge of the building based on the semantic segmentation result of the building, and obtaining an initial vector polygon;
step 4, eliminating wrong polygons and wrong edges and nodes of the polygons by using a coarse adjustment algorithm;
step 5, regularizing the vector polygon by using a regularization algorithm to obtain a regular building vector edge; the concrete implementation mode is as follows,
51) determining the main direction of the building: setting different side length thresholds W according to the area of the building, wherein the sides with the length larger than W are long sides, the rest are short sides, and reducing W in sequence for the building with the smaller area and the length which cannot find at least one long side until at least one long side is found; firstly, taking the longest side as the initial main direction, adding the longest side into a main direction list, comparing all the other long sides with the main direction list, and if the included angle between the long side and all the main directions is more than deltaminAnd is less than deltamaxAdding the direction of the long side into the main direction list;
52) adjusting the edges of the polygon according to the main direction: firstly, the long edge is adjusted, the long edge is compared with the main direction and the vertical direction of the main direction to obtain the minimum included angle, and if the included angle is less than deltaminRotating the long edge to a direction parallel or perpendicular to the main direction with the minimum included angle by taking the middle point as a rotation center; then, the short edge is adjusted, the short edge is compared with the main direction and the vertical direction of the main direction to obtain the minimum included angle, and if the included angle is smaller than TsRotating the short side to a parallel or vertical direction of the minimum included angle corresponding to the main direction by taking the middle point as a rotation center;
53) redundant parallel line culling and adding vertical lines: two continuous edges possibly exist in the adjusted edge and are parallel lines, different distance thresholds d are set according to the area of the building, if the distance between the two continuous parallel lines is larger than d, a perpendicular line is added between the two parallel lines, otherwise, the edge with the smaller length is removed;
54) and sequentially calculating intersection points between two continuous edges to serve as nodes of the vector polygon, and sequentially connecting all the nodes to obtain the adjusted vector polygon.
2. The building outline automatic extraction algorithm based on the convolutional neural network and the polygon regularization as claimed in claim 1, characterized in that: the multi-scale fusion full convolution neural network in the step 2 comprises 3 parts of encoding, decoding and outputting; wherein the coding part consists of 5 convolutional layers and 4 maximum pooling layers; the decoding part consists of 4 convolution layers and 4 deconvolution layers; the output section consists of 4 sub-outputs and 1 main output.
3. The building outline automatic extraction algorithm based on the convolutional neural network and the polygon regularization as claimed in claim 2, characterized in that: the first two convolutional layers of the coding part consist of two groups of continuously stacked convolutions, modified linear units and a batch normalization layer; the last three convolutional layers consist of three groups of continuously stacked convolution, correction linear units and two batch normalization layers; the first three convolutional layers and the last convolutional layer of the decoding part are respectively composed of 4 groups and 3 groups of convolution and correction linear units which are stacked continuously; each decoding part finally obtains a sub-output, the main output obtains a 4-channel characteristic diagram through the final 1-channel characteristic diagram on each scale of the serial decoding part, and the final prediction result is obtained through the activation of a Sigmoid function.
4. The building outline automatic extraction algorithm based on the convolutional neural network and the polygon regularization as claimed in claim 1, characterized in that: and 3, initializing edges by using a Douglas-Peucker algorithm, and setting different tolerance thresholds D according to the number of pixels of each connected domain to obtain an initial vector polygon.
5. The building outline automatic extraction algorithm based on the convolutional neural network and the polygon regularization as claimed in claim 1, characterized in that: the specific implementation of step 4 is as follows,
41) the sides of the building should have a certain length: rejecting polygons smaller than TdThe edge of (1);
42) building corners cannot be too sharp or too shallow: rejecting angles smaller than alpha and larger than beta;
43) the building should have a certain area: deleting polygons with the areas smaller than S;
44) the building nodes should not be excessive: the ratio of the rejection area to the number of the nodes is less than TaThe ratio of the sum perimeter to the number of the nodes is less than TpIs used for the polygon of (1).
6. The building outline automatic extraction algorithm based on the convolutional neural network and the polygon regularization as claimed in claim 1, characterized in that: the specific implementation of step 1 is as follows,
firstly, converting a building coverage vector file corresponding to an image into a binary raster image, wherein the pixel value of the raster image is a label of an image pixel, wherein 0 represents that a background 1 is a building, and cutting an original image and a corresponding raster label into tiles to construct a sample library.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910136835.2A CN109903304B (en) | 2019-02-25 | 2019-02-25 | Automatic building contour extraction algorithm based on convolutional neural network and polygon regularization |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910136835.2A CN109903304B (en) | 2019-02-25 | 2019-02-25 | Automatic building contour extraction algorithm based on convolutional neural network and polygon regularization |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109903304A CN109903304A (en) | 2019-06-18 |
| CN109903304B true CN109903304B (en) | 2021-03-16 |
Family
ID=66945394
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910136835.2A Active CN109903304B (en) | 2019-02-25 | 2019-02-25 | Automatic building contour extraction algorithm based on convolutional neural network and polygon regularization |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109903304B (en) |
Families Citing this family (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110909602A (en) * | 2019-10-21 | 2020-03-24 | 广联达科技股份有限公司 | Two-dimensional vector diagram sub-domain identification method and device |
| CN111080587B (en) * | 2019-12-04 | 2023-05-26 | 中山大学 | Resampling image detection method based on convolutional neural network |
| CN111127538B (en) * | 2019-12-17 | 2022-06-07 | 武汉大学 | Multi-view image three-dimensional reconstruction method based on convolution cyclic coding-decoding structure |
| CN111489387B (en) * | 2020-04-09 | 2023-10-20 | 湖南盛鼎科技发展有限责任公司 | Remote sensing image building area calculation method |
| CN111652892A (en) * | 2020-05-02 | 2020-09-11 | 王磊 | Remote sensing image building vector extraction and optimization method based on deep learning |
| CN111652251B (en) * | 2020-06-09 | 2023-06-27 | 星际空间（天津）科技发展有限公司 | Remote sensing image building feature extraction model construction method, device and storage medium |
| CN111652250B (en) * | 2020-06-09 | 2023-05-26 | 星际空间（天津）科技发展有限公司 | Remote sensing image building extraction method and device based on polygons and storage medium |
| CN112288647B (en) * | 2020-10-13 | 2022-06-03 | 武汉大学 | Remote sensing image cloud and shadow restoration method based on gating convolution |
| CN112348836B (en) * | 2020-11-06 | 2024-03-12 | 二十一世纪空间技术应用股份有限公司 | Method and device for automatically extracting building outline |
| US11836223B2 (en) * | 2020-11-13 | 2023-12-05 | Meta Platforms, Inc. | Systems and methods for automated detection of building footprints |
| CN112700464B (en) * | 2021-01-15 | 2022-03-29 | 腾讯科技（深圳）有限公司 | Map information processing method and device, electronic equipment and storage medium |
| CN112949407B (en) * | 2021-02-02 | 2022-06-14 | 武汉大学 | Remote sensing image building vectorization method based on deep learning and point set optimization |
| CN113160258B (en) * | 2021-03-31 | 2022-11-29 | 武汉汉达瑞科技有限公司 | Method, system, server and storage medium for extracting building vector polygon |
| CN113129323A (en) * | 2021-04-27 | 2021-07-16 | 西安微电子技术研究所 | Remote sensing ridge boundary detection method and system based on artificial intelligence, computer equipment and storage medium |
| CN113190639B (en) * | 2021-05-13 | 2022-12-13 | 重庆市勘测院 | Comprehensive drawing method for residential area |
| CN113920147B (en) * | 2021-12-14 | 2022-03-01 | 星际空间（天津）科技发展有限公司 | Remote sensing image building extraction method and device based on deep learning |
| CN114387329B (en) * | 2022-01-16 | 2024-03-22 | 四川轻化工大学 | Building contour progressive regularization method based on high-resolution remote sensing image |
| CN114863288B (en) * | 2022-07-05 | 2022-09-23 | 航天宏图信息技术股份有限公司 | Building contour extraction and regularization method and device |
| CN115797592B (en) * | 2023-02-08 | 2023-04-14 | 航天宏图信息技术股份有限公司 | Method and device for automatically generating building block based on oblique photography three-dimensional model |
| CN116343070B (en) * | 2023-05-22 | 2023-10-13 | 武汉天地鹰测绘科技有限公司 | Intelligent interpretation method for aerial survey image ground object elements |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105608458A (en) * | 2015-10-20 | 2016-05-25 | 武汉大学 | High-resolution remote sensing image building extraction method |
| CN106464320A (en) * | 2014-05-28 | 2017-02-22 | 高通股份有限公司 | Method and apparatus for orientation based beam selection |
| CN106548479A (en) * | 2016-12-06 | 2017-03-29 | 武汉大学 | A kind of multi-level laser point cloud building border rule method |
| CN106650749A (en) * | 2016-11-16 | 2017-05-10 | 桂林理工大学 | Method for plotting right-angled building in high-resolution optical image |
| CN108154209A (en) * | 2016-12-02 | 2018-06-12 | 杭州海康机器人技术有限公司 | A kind of Quick Response Code generation, recognition methods and device for localization for Mobile Robot |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101726255B (en) * | 2008-10-24 | 2011-05-04 | 中国科学院光电研究院 | Method for extracting interesting buildings from three-dimensional laser point cloud data |
| CN102521884B (en) * | 2011-12-16 | 2014-04-09 | 南京大学 | Three-dimensional roof reconstruction method based on LiDAR data and ortho images |
| CN102938066B (en) * | 2012-12-07 | 2015-10-28 | 南京大学 | A kind of based on the polygonal method of multivariate data reconstruction buildings outline |
| CN105608691B (en) * | 2015-12-17 | 2018-01-12 | 武汉大学 | A kind of High-resolution SAR Images single building extracting method |
| CN106127857B (en) * | 2016-06-24 | 2018-11-16 | 南京林业大学 | The on-board LiDAR data modeling method of integrated data driving and model-driven |
-
2019
- 2019-02-25 CN CN201910136835.2A patent/CN109903304B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106464320A (en) * | 2014-05-28 | 2017-02-22 | 高通股份有限公司 | Method and apparatus for orientation based beam selection |
| CN105608458A (en) * | 2015-10-20 | 2016-05-25 | 武汉大学 | High-resolution remote sensing image building extraction method |
| CN106650749A (en) * | 2016-11-16 | 2017-05-10 | 桂林理工大学 | Method for plotting right-angled building in high-resolution optical image |
| CN108154209A (en) * | 2016-12-02 | 2018-06-12 | 杭州海康机器人技术有限公司 | A kind of Quick Response Code generation, recognition methods and device for localization for Mobile Robot |
| CN106548479A (en) * | 2016-12-06 | 2017-03-29 | 武汉大学 | A kind of multi-level laser point cloud building border rule method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109903304A (en) | 2019-06-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109903304B (en) | Automatic building contour extraction algorithm based on convolutional neural network and polygon regularization | |
| US10528812B1 (en) | Distributed and self-validating computer vision for dense object detection in digital images | |
| CN107609525B (en) | Remote sensing image target detection method for constructing convolutional neural network based on pruning strategy | |
| CN108647655B (en) | Low-altitude aerial image power line foreign matter detection method based on light convolutional neural network | |
| CN114120102A (en) | Boundary-optimized remote sensing image semantic segmentation method, device, equipment and medium | |
| CN111126359B (en) | High-definition image small target detection method based on self-encoder and YOLO algorithm | |
| US20210027532A1 (en) | Primitive-based 3d building modeling, sensor simulation, and estimation | |
| CN113516135B (en) | Remote sensing image building extraction and contour optimization method based on deep learning | |
| CN111898507A (en) | Deep learning method for predicting earth surface coverage category of label-free remote sensing image | |
| CN111652892A (en) | Remote sensing image building vector extraction and optimization method based on deep learning | |
| CN111951284B (en) | Optical remote sensing satellite image refined cloud detection method based on deep learning | |
| CN114510772B (en) | Rapid generation method of building surface contour based on oblique scanning data | |
| CN108154158B (en) | Building image segmentation method for augmented reality application | |
| CN111125397B (en) | Cloth image retrieval method based on convolutional neural network | |
| CN112101364B (en) | Semantic segmentation method based on parameter importance increment learning | |
| CN110443279B (en) | Unmanned aerial vehicle image vehicle detection method based on lightweight neural network | |
| CN111353396A (en) | Concrete crack segmentation method based on SCSEOCUnet | |
| CN111178438A (en) | ResNet 101-based weather type identification method | |
| CN112233129A (en) | Deep learning-based parallel multi-scale attention mechanism semantic segmentation method and device | |
| CN113159215A (en) | Small target detection and identification method based on fast Rcnn | |
| CN114820668A (en) | End-to-end building regular outline automatic extraction method based on concentric ring convolution | |
| CN114283343A (en) | Map updating method, training method and equipment based on remote sensing satellite image | |
| CN111161159B (en) | Image defogging method and device based on combination of priori knowledge and deep learning | |
| CN116403109A (en) | Building identification and extraction method and system based on improved neural network | |
| CN114612450B (en) | Image detection segmentation method and system based on data augmentation machine vision and electronic equipment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |