CN109242922B - Terrain synthesis method based on radial basis function network - Google Patents
Terrain synthesis method based on radial basis function network Download PDFInfo
- Publication number
- CN109242922B CN109242922B CN201810942069.4A CN201810942069A CN109242922B CN 109242922 B CN109242922 B CN 109242922B CN 201810942069 A CN201810942069 A CN 201810942069A CN 109242922 B CN109242922 B CN 109242922B
- Authority
- CN
- China
- Prior art keywords
- terrain
- training
- height
- point
- network
- 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
- 238000001308 synthesis method Methods 0.000 title claims abstract description 9
- 238000000034 method Methods 0.000 claims abstract description 21
- 230000015572 biosynthetic process Effects 0.000 claims abstract description 13
- 238000003786 synthesis reaction Methods 0.000 claims abstract description 13
- 230000006870 function Effects 0.000 claims description 29
- 239000013598 vector Substances 0.000 claims description 16
- 239000002245 particle Substances 0.000 claims description 12
- 238000004422 calculation algorithm Methods 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 3
- 210000002569 neuron Anatomy 0.000 claims description 3
- 238000001514 detection method Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000003628 erosive effect Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 230000002194 synthesizing effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T11/00—2D [Two Dimensional] image generation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a terrain synthesis method based on a radial basis function network, which utilizes a terrain data block of a Digital Elevation model (DEM for short) and adopts the radial basis function network to learn and extract terrain features; during terrain synthesis, according to a user hand-drawn sketch, the customized terrain corresponding to the user sketch can be synthesized by combining input elevation data and utilizing the characteristics learned in advance. The method has the characteristics of simplicity and effectiveness, and can synthesize a specific terrain according to the hand-drawn sketch of the user.
Description
Technical Field
The invention relates to the technical field of virtual simulation, in particular to a terrain synthesis method based on a radial basis function network, which adopts a terrain data block of a Digital Elevation model (DEM for short) and the learning and extraction of terrain features by the radial basis function network; during terrain synthesis, according to a user hand-drawn sketch, the customized terrain corresponding to the user sketch can be synthesized by combining with input elevation data and utilizing the characteristics learned in advance. The method has the characteristics of simplicity and effectiveness, and can synthesize a specific terrain according to the user hand-drawn sketch.
Background
The terrain is a main visual element in a three-dimensional virtual scene, and due to the wide application value of the terrain in real life, the terrain synthesis technology becomes a research hotspot in the field of computer vision. The realistic terrain can better improve the use experience of users in natural disaster prevention and movie game creation. Terrain synthesis technologies can be broadly divided into three categories: the modeling method based on the process, the modeling method based on the physical erosion and the modeling method based on the user sketch. In recent years, with the popularization and use of DEM topographic data, people begin to research a modeling method based on a sketch by combining with real topographic data characteristics, and realize a customized function of a user through sketch control.
Meanwhile, in order to realize intelligent synthesis of the terrain, machine learning and deep learning are rapidly developed in recent years, researchers try to understand implicit characteristics in real terrain data by using an artificial intelligence method, a terrain synthesis method based on a convolutional neural network appears, depth information prediction of a user input mountain contour line is realized, and the rationality of the synthesized terrain is improved. However, the deep learning method for synthesizing the terrain has the main problems that the network structure is complex, and the network parameter training is difficult to converge, which are the problems existing in the current intelligent terrain synthesis research.
Disclosure of Invention
The invention aims to overcome the defects in the prior art and solve the practical problems in terrain synthesis, and provides a terrain synthesis method based on a radial basis function network. The method has the characteristics of simplicity and effectiveness, and can synthesize a specific terrain according to the hand-drawn sketch of the user.
The specific technical scheme for realizing the purpose of the invention is as follows:
a terrain synthesis method based on a radial basis function network is characterized by comprising the following specific steps:
step 1: preparing blocks of elevation data
Downloading an elevation data block D of a WGS84 coordinate system from an SRTM website http:// SRTM. Csi. Cgiar. Org, wherein the spatial resolution of D is between 90m multiplied by 90m and 200m multiplied by 200m, the number of pixel points in D is N, and the height of any point A is recorded as H A ；
And 2, step: establishing a training data set:
(1) First, the entropy E of an arbitrary point A is calculated A ：
Wherein p is k Is the kth point (with height H) in the 3X 3 neighborhood T of any point A k ) K =1,2 \ 82309, 9,p k The calculation is as follows:
wherein,
the height statistics of 9 neighborhood points in T is carried out, and eta is the height minimum value of 9 points in T; delta is a constant value, and H is taken as 0.0001 t (T =1,2, \8230;, 9) represents the height of the T-th neighborhood point in T;
(2) Establishing a feature vector of A: v A ＝(E A ,H A ) Further, a feature vector set S = { (E) is constructed by using the feature vector of each point in D s ,H s )|1≤s≤N}；
(3) Randomly selecting M eigenvectors C from S e ＝(E e ,H e ) (e =1,2 \8230;, M) as M cluster centers, M is 3 ≦ 5, and the feature vectors in S are grouped into M classes using the K-means method: l is f (f＝1,2…,M)；
(4) Respectively for each type L f (f =1,2 \8230;, M) the entropy of the pixel particle is summed up statistically, and the class with the largest statistical entropy is denoted as L m (M is more than or equal to 1 and less than or equal to M), adding L m Considered as a salient region;
(5) With L m Establishing a data set: l is m And (3) marking the pixel particle set corresponding to the medium feature vector as G, calculating a skeleton B of the G by using a 3 × 3 window by using a D8 algorithm, specifically, calculating a difference between a center and 8 pixels around the center and the neighborhood by using a 3 × 3 neighborhood around the pixel, and taking the direction with the largest difference as the skeleton direction. The feature vector set of the particles in G is U = { (d) g ,h g )|1≤g≤Q}，d g Is the shortest distance, h, from any particle j in G to B g Is the height of j, Q represents the number of data elements in G;
and step 3: constructing a radial basis function network for learning and training
Establishing a three-layer forward propagation radial basis function network X, which comprises an input layer, a hidden layer and an output layer: during training, a supervised learning method is adopted, and the network input is d g Output ofThe supervision data is h g G =1,2 \ 8230; in the hidden layer, the radial basis function adopts a Gaussian function, and the training parameters of the model comprise: the number of gaussian kernels, the center of the hidden layer neurons (the center of the basis functions), the variance, and the weight between the hidden layer and the output layer; the loss function of the training is defined as:
wherein, P g Is the predicted altitude result of the X network output, h g As high supervision data, g is more than or equal to 1 and less than or equal to Q;
training process: initializing the number K of Gaussian kernels to 5; (b) Training X, calculating a loss function F by using a formula (3), if F is larger than a threshold value of 0.01, the number K = K +5 of Gaussian kernels, and turning to the step (b) to perform next training; otherwise, if F is less than or equal to the threshold 0.01, the training is terminated;
and 4, step 4: terrain synthesis
Inputting a binary image V as a hand-drawn sketch by a user, drawing the sketch of the user by adopting drawing software, storing the sketch as a binary image, taking the V as a framework, and calculating the distance J from any point on the V to the framework b (b =1,2 \8230;, x), x representing the number of pixels in the synthetic terrain;
will J b (b =1,2 \8230;, X) is input into the network X, and the height value of each point of the synthetic terrain is obtained by predicting by using the trained parameters in the X network, so that the synthetic terrain is obtained.
The method has the characteristics of simplicity and practicality, and can synthesize the terrain customized by a specific user according to the user hand-drawn sketch and the input elevation data.
Drawings
FIG. 1 is a graph of the results of salient region detection according to the present invention;
fig. 2 is a graph showing the results of the terrain synthesis of the present invention.
Detailed Description
Examples
The invention is further described below with reference to the accompanying drawings.
The embodiment is implemented under a Windows10 64-bit operating system on a PC, and the hardware configuration is a processor
Core TM I5-7500 3.4GHz CPU,8GB memory, matlab 2015b software environment, python language adopted by programming, and combination of a visual open source library OpenCV 2.4.4 and an open source grid space data conversion library GDAL.
The specific embodiment scheme of the invention is as follows:
a terrain synthesis method based on a radial basis function network comprises the following specific steps:
step 1: preparing blocks of elevation data
Downloading an elevation data block D of a WGS84 coordinate system from an SRTM website http:// SRTM. Csi. Cgiar. Org, wherein the spatial resolution of D is 90m multiplied by 90m, the number of pixel points in D is N, and the height of any point A is recorded as H A ；
Step 2: establishing a training data set:
(1) First, the entropy E of A is calculated A ：
Wherein p is k Is the kth point in the 3X 3 neighborhood T of A (height H) k ) K =1,2 \ 82309, 9,p k The calculation is as follows:
wherein,
the height statistics of 9 neighborhood points in T is carried out, and eta is the minimum height value of 9 points in T; delta is a constant value, and H is taken as 0.0001 t (T =1,2, \8230;, 9) represents the height of the T-th neighborhood point in T;
(2) Characteristics of ASign vector: v A ＝(E A ,H A ) And further utilizing the feature vector of each point in D to construct a feature vector set S = (E) s ,H s )|1≤s≤N}；
(3) Randomly selecting 3 eigenvectors C from S e ＝(E e ,H e ) (e =1,2,3), as 3 clustering centers, feature vectors in S are clustered into 3 classes using the K-means method: l is f (f＝1,2,3)；
(4) Respectively for each type L f (f =1,2,3) the entropy of the pixel particles is summed up and the class with the largest statistical entropy is denoted as L m (1. Ltoreq. M. Ltoreq.3), adding L m Considered as a salient region;
(5) By L m Establishing a data set: l is m Marking the pixel particle set corresponding to the medium eigenvector as G, and calculating the skeleton B of the G by using a 3 multiplied by 3 window by using a D8 algorithm, wherein the eigenvector set of the particles in the G is U = { (D) g ,h g )|1≤g≤Q}，d g Is the shortest distance, h, from any particle j in G to B g Q represents the number of data elements in G, which is the height of j;
and step 3: constructing a radial basis function network for learning and training
Establishing a three-layer forward propagation radial basis function network X, which comprises an input layer, a hidden layer and an output layer: during training, a supervised learning method is adopted, and the network input is d g The output supervision data is h g G =1,2 \ 8230; in the hidden layer, the radial basis function adopts a Gaussian function, and the training parameters of the model comprise: the number of Gaussian kernels, the center of hidden layer neurons (the center of the basis function), the variance, and the weight between the hidden layer and the output layer; the loss function of the training is defined as:
wherein, P g Is the predicted altitude result of the X network output, h g As high supervision data, g is more than or equal to 1 and less than or equal to Q;
training process: initializing the number K of Gaussian kernels to 5; (b) Training X, calculating a loss function F by using a formula (3), if F is larger than a threshold value of 0.01, the number K = K +5 of Gaussian kernels, and turning to the step (b) to perform next training; otherwise, if F is less than or equal to the threshold 0.01, training is terminated;
and 4, step 4: terrain synthesis
Inputting a binary image V as a hand-drawn sketch by a user, drawing the sketch of the user by adopting drawing software, storing the sketch as a binary image, taking the V as a skeleton, and calculating the distance J from any point on the V to the skeleton b (b =1,2 \ 8230;, x), x representing the number of pixels in the synthetic terrain;
will J b (b =1,2 \8230;, X) is input into the network X, and the height value of each point of the synthetic terrain is obtained by prediction using the trained parameters in the X network, thus obtaining the synthetic terrain.
Fig. 1 shows the result of salient region detection based on DEM topographic data. The figure shows 3 experimental results, one for each action; in each row, the 1 st column on the left side is an input elevation data block, the result on the right side shows the result of detection of a salient region, as can be seen from the figure, the salient region corresponds to a detail region of the DEM terrain block, and features are extracted from the salient region to create a data set;
fig. 2 is a result of terrain synthesis based on a user input sketch. The figure shows 3 experimental results, one for each action; in each row, the 1 st column on the left side is an input elevation data block, and the middle column is a sketch input by a user; the results in the 1 st column on the right side are synthetic terrain results, and it is obvious from the synthetic results that the method provided by the invention can realize the synthesis of real terrain according to the user sketch, and the method is simple and convenient.
Claims (1)
1. A terrain synthesis method based on a radial basis function network is characterized by comprising the following specific steps:
step 1: preparing blocks of elevation data
Downloading an elevation data block D of a WGS84 coordinate system from an SRTM website http:// SRTM. Csi. Cgiar. Org, wherein the spatial resolution of D is 90m multiplied by 90m to200m is multiplied by 200m, the number of pixel points in D is N, and the height of any point A is marked as H A ；
Step 2: establishing a training data set:
(1) First, the entropy E of an arbitrary point A is calculated A ：
Wherein p is k Is the height distribution of the kth point in a 3X 3 neighborhood T of an arbitrary point A, the height of the kth point is H k ；k＝1,2…,9，p k The calculation is as follows:
wherein,
the height statistics of 9 neighborhood points in T is carried out, and eta is the height minimum value of 9 points in T; delta is a constant value, and H is taken as 0.0001 t (T =1,2, \8230;, 9) represents the height of the tth neighborhood point in T;
(2) Establishing a feature vector of A: v A ＝(E A ,H A ) And further utilizing the feature vector of each point in D to construct a feature vector set S = (E) s ,H s )|1≤s≤N}；
(3) Randomly selecting M eigenvectors C from S e ＝(E e ,H e ) (e =1,2 \8230;, M) as M cluster centers, M is 3 ≦ 5, and the feature vectors in S are grouped into M classes using the K-means method: l is f (f＝1,2…,M)；
(4) Respectively for each type L f (f =1,2 \8230;, M) the entropy of the pixel particle is summed up statistically, and the class with the largest statistical entropy is denoted as L m (M is more than or equal to 1 and less than or equal to M), adding L m Considered as a salient region;
(5) By L m Establishing a data set: l is m Marking the pixel particle set corresponding to the medium feature vector as G, calculating the skeleton B of G by using a 3 multiplied by 3 window by using a D8 algorithm,the feature vector set of the particles in G is U = { (d) g ,h g )|1≤g≤Q}，d g Is the shortest distance, h, from any particle j in G to B g Is the height of j, Q represents the number of data elements in G;
and step 3: constructing a radial basis function network for learning and training
Establishing a three-layer forward propagation radial basis function network X, which comprises an input layer, a hidden layer and an output layer: during training, a supervised learning method is adopted, and the network input is d g The output supervision data is h g G =1,2 \ 8230; in the hidden layer, the radial basis function adopts a Gaussian function, and the training parameters of the model comprise: the number of Gaussian kernels, the center of a hidden layer neuron, namely the center of a base function, the variance and the weight between a hidden layer and an output layer; the loss function of the training is defined as:
wherein, P g Is the predicted altitude result of the X network output, h g As high supervision data, g is more than or equal to 1 and less than or equal to Q;
training: initializing the number K of Gaussian kernels to 5; (b) Training X, calculating a loss function F by using a formula (3), if F is larger than a threshold value of 0.01, the number K = K +5 of Gaussian kernels, and turning to the step (b) to perform next training; otherwise, if F is less than or equal to the threshold 0.01, training is terminated;
and 4, step 4: terrain synthesis
Inputting a binary image V as a hand-drawn sketch by a user, drawing the sketch of the user by adopting drawing software, storing the sketch as a binary image, taking the V as a framework, and calculating the distance J from any point on the V to the framework b (b =1,2 \8230;, x), x representing the number of pixels in the synthetic terrain;
will J b (b =1,2 \8230;, X) is input into the network X, and the height value of each point of the synthesized terrain is obtained by predicting by using the trained parameters in the X network, so that the synthesized terrain is obtained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810942069.4A CN109242922B (en) | 2018-08-17 | 2018-08-17 | Terrain synthesis method based on radial basis function network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810942069.4A CN109242922B (en) | 2018-08-17 | 2018-08-17 | Terrain synthesis method based on radial basis function network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109242922A CN109242922A (en) | 2019-01-18 |
| CN109242922B true CN109242922B (en) | 2023-04-07 |
Family
ID=65071520
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810942069.4A Active CN109242922B (en) | 2018-08-17 | 2018-08-17 | Terrain synthesis method based on radial basis function network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109242922B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112632869B (en) * | 2020-12-23 | 2022-09-27 | 中国空气动力研究与发展中心计算空气动力研究所 | Grid frame-based non-structural boundary layer grid generation method |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107122396A (en) * | 2017-03-13 | 2017-09-01 | 西北大学 | Three-dimensional model searching algorithm based on depth convolutional neural networks |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140270489A1 (en) * | 2013-03-12 | 2014-09-18 | Microsoft Corporation | Learned mid-level representation for contour and object detection |
-
2018
- 2018-08-17 CN CN201810942069.4A patent/CN109242922B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107122396A (en) * | 2017-03-13 | 2017-09-01 | 西北大学 | Three-dimensional model searching algorithm based on depth convolutional neural networks |
Non-Patent Citations (1)
| Title |
|---|
| 赵鹏 ; 刘杨 ; 刘慧婷 ; 姚晟 ; .基于深度卷积-递归神经网络的手绘草图识别方法.计算机辅助设计与图形学学报.2018,(第02期),全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109242922A (en) | 2019-01-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113469356B (en) | Improved VGG16 network pig identity recognition method based on transfer learning | |
| CN110119703B (en) | Human body action recognition method fusing attention mechanism and spatio-temporal graph convolutional neural network in security scene | |
| Zhang et al. | Relational attention network for crowd counting | |
| Clouâtre et al. | Figr: Few-shot image generation with reptile | |
| CN109740419A (en) | A kind of video behavior recognition methods based on Attention-LSTM network | |
| Xie et al. | Learning energy-based spatial-temporal generative convnets for dynamic patterns | |
| CN114049381A (en) | Twin cross target tracking method fusing multilayer semantic information | |
| Hua et al. | Depth estimation with convolutional conditional random field network | |
| CN115147891A (en) | System, method, and storage medium for generating synthesized depth data | |
| CN113221663A (en) | Real-time sign language intelligent identification method, device and system | |
| Ha et al. | Deep neural networks using capsule networks and skeleton-based attentions for action recognition | |
| CN109902613A (en) | A kind of human body feature extraction method based on transfer learning and image enhancement | |
| Ding et al. | Skeleton-based square grid for human action recognition with 3D convolutional neural network | |
| CN114638408A (en) | Pedestrian trajectory prediction method based on spatiotemporal information | |
| CN109598771B (en) | Terrain synthesis method of multi-landform feature constraint | |
| CN114581918A (en) | Text recognition model training method and device | |
| CN110111365A (en) | Training method and device and method for tracking target and device based on deep learning | |
| Zhang et al. | Multiresolution attention extractor for small object detection | |
| Salem et al. | Semantic image inpainting using self-learning encoder-decoder and adversarial loss | |
| Salami et al. | Motion pattern recognition in 4d point clouds | |
| Shariff et al. | Artificial (or) fake human face generator using generative adversarial network (GAN) machine learning model | |
| Zhou et al. | Image2GIF: Generating cinemagraphs using recurrent deep q-networks | |
| Du et al. | Adaptive visual interaction based multi-target future state prediction for autonomous driving vehicles | |
| CN109242922B (en) | Terrain synthesis method based on radial basis function network | |
| CN109658508B (en) | Multi-scale detail fusion terrain synthesis method |
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 |