CN106600553A - DEM super-resolution method based on convolutional neural network - Google Patents
DEM super-resolution method based on convolutional neural network Download PDFInfo
- Publication number
- CN106600553A CN106600553A CN201611159517.0A CN201611159517A CN106600553A CN 106600553 A CN106600553 A CN 106600553A CN 201611159517 A CN201611159517 A CN 201611159517A CN 106600553 A CN106600553 A CN 106600553A
- Authority
- CN
- China
- Prior art keywords
- resolution
- dem
- data
- convolutional neural
- super
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 89
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 52
- 238000012549 training Methods 0.000 claims abstract description 18
- 238000005070 sampling Methods 0.000 claims description 22
- 230000008569 process Effects 0.000 claims description 19
- 230000005284 excitation Effects 0.000 claims description 10
- 238000012545 processing Methods 0.000 claims description 6
- 238000011478 gradient descent method Methods 0.000 claims description 4
- 230000000593 degrading effect Effects 0.000 claims description 2
- 239000000284 extract Substances 0.000 claims description 2
- 238000000605 extraction Methods 0.000 abstract description 3
- 230000006870 function Effects 0.000 description 9
- 238000005457 optimization Methods 0.000 description 9
- 241001269238 Data Species 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 238000004519 manufacturing process Methods 0.000 description 3
- 238000013507 mapping Methods 0.000 description 3
- 230000001537 neural effect Effects 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 235000013399 edible fruits Nutrition 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000007123 defense Effects 0.000 description 1
- 238000013178 mathematical model Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000007935 neutral effect Effects 0.000 description 1
- 238000013441 quality evaluation Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/73—Deblurring; Sharpening
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a DEM super-resolution method based on a convolutional neural network. The method includes following steps: (1) obtaining a super-resolution convolutional neural network through training according to low-resolution image data and high-resolution image data corresponding to each other in advance; (2) extending to-be-processed low-resolution DEM data by employing an interpolation method, and obtaining quasi-high-resolution DEM data having the same scale with expected high-resolution DEM data; (3) obtaining a gradient map of the quasi-high-resolution DEM data by employing en edge extraction operator; (4) inputting the gradient map to the super-resolution convolutional neural network, and obtaining an estimation gradient map of the high-resolution DEM data; and (5) reconstructing a height map of the high-resolution DEM based on constraint of the estimation gradient map and the to-be-processed low-resolution DEM data. According to the super-resolution method, the robustness is high, and the precision of the reconstruction result is high.
Description
Technical field
The invention belongs to terrain mapping technology field, more particularly, to a kind of DEM oversubscription based on convolutional neural networks
Resolution method.
Background technology
Digital elevation model (Digital Elevation Model, DEM) is a branch of digital terrain model, it
It is a kind of mathematical model that ground elevation is represented with one group of orderly array of values form.It has wide in economy and national defense construction
In general application demand, such as trajectory planning, it is desirable to using more accurate terrain data, as far as possible program results most
Excellent, pursuit of the old friends to high accuracy relief model is an eternal theme.To obtain high-precision relief model, generally adopt
Use two methods.A kind of method generally adopts for example high-resolution remote sensing image of measuring apparatus of higher precision, and by intensive
Measurement, to reach high-precision purpose is put forward.This method production cost is high, especially for the mapping of sea-floor relief, relative to
For earth's surface is using satellite remote-sensing image mapping, cost increases especially pronounced.Second method takes the strategy of super-resolution, leads to
The process to high-resolution data is crossed, the precision and resolution of DEM is improved, the cost that high-resolution data is obtained is reduced, therefore
The concern of a large number of researchers is attracted.
Analyze according to more than, the strategy of existing super-resolution mainly takes a kind of non-local learning method, the method
Using traditional super-resolution method based on study, according to the principle of manifold learning, by calculating test dem data and sample
The similarity of high-resolution data in storehouse, according to similarity and the corresponding high-resolution datas of similar DEM superresolution processing is carried out.
The method improves to a certain extent the precision of high resolution DEM data, but has the disadvantage that the method adaptability relatively
It is little, and depend critically upon the measure of similarity.In addition, to set up based on the DEM super-resolution methods of convolutional neural networks,
Need using substantial amounts of DEM samples, and high-precision DEM is difficult to what is obtained relative to image pattern, because most height
Resolution DEM is secrecy.Therefore, set up one kind and do not rely on high-precision dem data and the preferable DEM oversubscription of robustness
Resolution method tool is of great significance.
The content of the invention
For the disadvantages described above or Improvement requirement of prior art, it is an object of the invention to provide a kind of be based on convolutional Neural
The DEM super-resolution methods of network, wherein by low resolution dem data processing procedure, the convolutional neural networks crucial to its
Training learning style etc. be improved so that the super-resolution method strong robustness, reconstructed results high precision;Also, we
Method is relatively small by the gradient dynamic range by means of DEM, meets the precision of the convolutional neural networks trained by image pattern,
High accuracy DEM sample data is obtained so as to avoid, production cost is greatly reduced.
For achieving the above object, it is proposed, according to the invention, there is provided a kind of DEM super-resolution sides based on convolutional neural networks
Method, it is characterised in that comprise the following steps:
(1) according to mutual corresponding low resolution image data and high resolution image data in advance, what training was obtained
Super-resolution convolutional neural networks;
(2) pending low resolution dem data is expanded into s times using interpolation method, is obtained and desired high-resolution
The quasi- high resolution DEM data of the same scale size of dem data;
(3) gradient map of the quasi- high resolution DEM data is obtained using arithmetic operators;
(4) gradient map obtained in the step (3) is input to the super-resolution obtained in the step (1)
In rate convolutional neural networks, obtain the estimation gradient map of high resolution DEM data and estimate Grad accordingly;
(5) based on the estimation gradient map obtained in the step (4) and the pending low resolution dem data
Constraint, reconstruct the height map of high resolution DEM.
Used as present invention further optimization, the step (1) specifically includes following steps:
(1-1) high resolution image data is carried out degrading processing and obtains corresponding low resolution image data, then
The gradient for obtaining the high resolution image data and the low resolution image data is extracted using arithmetic operators
Figure, then each gradient map is divided into into multiple gradient blocks, then choose corresponding high resolution image data gradient block,
And the gradient block of low resolution image data is used as training sample;
(1-2) convolutional neural networks are built and model parameter is set;
(1-3) volume in the step (1-2) is trained according to the training sample obtained in the step (1-1)
Product neutral net, obtains super-resolution convolutional neural networks.
As present invention further optimization, the convolutional neural networks in the step (1-2), including multiple convolution
Layer, the adjacent convolutional layer of any two is connected by excitation layer;Wherein, the first convolutional layer is used to be input into low resolution
View data, the end convolutional layer is used to export high resolution image data.
Used as present invention further optimization, the step (5) specifically includes following steps:
(5-1) one is built based on the estimation gradient map and the pending constraint of the low resolution dem data, pass
In the least square function of target high-resolution dem data;
(5-2) method for solving updated using iteration finds the optimum of least square function described in the step (5-1)
Solution;
If (5-3) carrying out the data that obtain after down-sampled process and the pending low resolution to the optimal solution
Average reconstruction error between dem data exceedes threshold value Th set in advance or iterationses are not up to set in advance wanting
Ask, then return the step (5-2);Otherwise, the optimal solution corresponds to final high resolution DEM data, and the optimal solution is
For rebuilding the height map of the high resolution DEM.
Used as present invention further optimization, in the step (5-1), the least square function isWherein, X is the pending low resolution dem data;Y is
The high resolution DEM data of target, ▽ Y are the corresponding gradients of Y,For the high-resolution obtained in the step (4)
The estimation Grad of dem data;↓sS times of down-sampled process of expression, Y ↓sIt is that the data that s times of down-sampled process is obtained are carried out to Y;β
It is weight factor set in advance.
Used as present invention further optimization, the method for solving that iteration described in the step (5-2) updates is under gradient
Drop method;Preferably, the Y after the t+1 time iteration updatest+1Meet:
Wherein, YtIt is the data after the t time iteration updates;τ is iteration step length set in advance;Y be the t+1 time iteration more
Y after newt+1;X is the pending low resolution dem data;↓sS times of down-sampled process of expression, Y ↓sIt is to carry out s times to Y to drop
The data that sampling processing is obtained;↑sRepresent that s times of up-sampling is processed, (X-Y ↓s)↑sBe to (X-Y ↓s) carry out s times of up-sampling and process
The data for arriving;▽ Y are the corresponding gradients of Y,The estimation ladder of the high resolution DEM data to obtain in the step (4)
Angle value;β is weight factor set in advance;Div is divergence computing.
Used as present invention further optimization, in the step (5-3), Th is preferably 5;
In the step (5-1) and the step (5-2), β is preferably 0.03.
Used as present invention further optimization, in the step (5-2), maximum iteration time is preferably 150 times, iteration step
Long τ is preferably 0.2.
Used as present invention further optimization, the up-sampling is processed and the down-sampled process, is to take interpolation method
Or the dot interlace method of sampling;Preferably, the interpolation method is nearest-neighbor interpolation, bilinear interpolation or bicubic interpolation;
The arithmetic operators can be Sobel operators, Roberts operators, Prewitt operators or Canny operators.
As present invention further optimization, in the step (3), the gradient map bag of the quasi- high resolution DEM data
Containing the edge graph in X-direction and Y-direction, the X-direction and the Y-direction are mutually perpendicular to.
By the contemplated above technical scheme of the present invention, compared with prior art, due to according to mutually corresponding in advance
(wherein, low resolution image data can be by high resolution graphics for low resolution image data and high resolution image data
As data are obtained by down-sampled process), training obtains super-resolution convolutional neural networks, recycles the super-resolution to use
Convolutional neural networks process initial low resolution dem data (that is, pending low resolution dem data), can reconstruct height
Resolution dem data.
DEM super-resolution methods in the present invention based on convolutional neural networks, for summary, mainly include the following steps that:
First, by obtaining great amount of images high-resolution data, all images in image data base are carried out down-sampled, fuzzy etc. to degrade
Operation obtains corresponding high-resolution data, then extracts the gradient map of high-low resolution data respectively using arithmetic operators
And be cut into small pieces, using the gradient block of corresponding high-low resolution data as training sample;Then, based on the training sample for obtaining
This, sets up and trains the super-resolution convolutional neural networks mapped from low resolution to high-resolution;Then, using slotting
Low resolution dem data is expanded value method the quasi- high-resolution obtained with the same scale size of desired high resolution DEM data
Rate DEM;Then, the gradient map of quasi- high resolution DEM data is obtained using arithmetic operators;Then, gradient map is input to
Obtaining the estimation gradient map of high resolution DEM data in the convolutional neural networks for having trained;Finally, based on estimation gradient
The constraint of figure and original low-resolution dem data, reconstructs the height map of high resolution DEM.The present invention can utilize view data
To rebuild high resolution DEM data, so as to solve the problems, such as that high resolution DEM data are difficult to a large amount of acquisitions and reconstructed results are clear
Clear, accuracy is high.
In general, by the contemplated above technical scheme of the present invention compared with prior art, with following beneficial effect
Really:
1st, will be incorporated in DEM reconstructions based on the learning method of convolutional neural networks, as long as one is trained, just can be used
In the super-resolution of all dem datas, so as to improve the robustness of algorithm.
2nd, the CNN networks (that is, convolutional neural networks) trained using image pattern realize DEM gradient super-resolutions, so as to
Carry out DEM super-resolution rebuildings, it is to avoid collect the job content of high accuracy DEM sample data, greatly reduce production cost.
3rd, the caused knot using the dynamic range avoided based on the thought of gradient because of the dynamic range of DEM higher than image
The larger shortcoming of fruit error, so as to improve the degree of accuracy of algorithm so that reconstructed results are clear, and accuracy is higher.
Description of the drawings
Fig. 1 is the flow chart of the DEM super-resolution methods of the embodiment of the present invention;
Fig. 2 is super-resolution convolutional neural networks (that is, CNN) structure chart;
Fig. 3 is the DEM super-resolution method and traditional method comparison diagram of the embodiment of the present invention.
Specific embodiment
In order that the objects, technical solutions and advantages of the present invention become more apparent, it is right below in conjunction with drawings and Examples
The present invention is further elaborated.It should be appreciated that specific embodiment described herein is only to explain the present invention, and
It is not used in the restriction present invention.As long as additionally, technical characteristic involved in invention described below each embodiment
Not constituting conflict each other just can be mutually combined.
As shown in figure 1, the DEM super-resolution methods based on convolutional neural networks of the embodiment of the present invention, including following step
Suddenly:
(1) low resolution dem data is expanded into s times using interpolation method, to obtain and desired high resolution DEM data
The quasi- high resolution DEM data of same scale size (that is, picture size is identical).
Wherein, the interpolation method is nearest-neighbor interpolation, bilinear interpolation or bicubic interpolation.
(2) gradient map of quasi- high resolution DEM data is obtained using arithmetic operators, gradient map now includes X and Y
Edge graph on direction.
Wherein, arithmetic operators can be Sobel operators, Roberts operators, Prewitt operators or Canny operators.
(3) gradient map is input to the super-resolution convolutional Neural net obtained according to the training of high-low resolution view data
In network, the estimation gradient map of high resolution DEM data is obtained;
Wherein, the super-resolution convolutional neural networks that the high-low resolution view data training is obtained, according to as follows
Step is obtained:
(3-1) multiple images high-resolution data (quantity is The more the better) is obtained, all images in image data base is entered
Row elder generation down-sampling, the operation that degrades such as up-sample again and obtain corresponding low resolution image data, by high-low resolution view data
Carry out extracting gradient map using arithmetic operators and being cut into small pieces, the gradient block of corresponding high-low resolution view data is made
For training sample, wherein up-sampling and down-sampling can take bicubic interpolation method or the dot interlace method of sampling;
(3-2) convolutional neural networks are built and model parameter is set, Fig. 2 illustrates the model of convolutional neural networks, including
Convolutional layer, excitation layer, convolutional layer, excitation layer and the convolutional layer being sequentially connected (wherein each excitation layer Fig. 2 is not showed that).Finally
The as corresponding high-resolution data of convolutional layer output.Then model parameter is set, and wherein learning rate is preferably arranged to
0.001, often train 10000 learning rates to drop to the 1/5 of former learning rate;
(3-3) obtain in gradient block training step (3-2) of the high-low resolution view data of acquisition according in step (3-1)
The convolutional neural networks for obtaining, obtain described super-resolution convolutional neural networks;
Preferably, the convolutional neural networks model described in step (3-2) include be sequentially connected convolutional layer, excitation layer,
Convolutional layer, excitation layer and convolutional layer.The as corresponding high-resolution data of last convolutional layer output.
(4) by the constraint of the estimation gradient map that obtains in step (3) and original low-resolution dem data, height is reconstructed
The height map of resolution DEM.
Wherein, the height process of reconstruction concrete steps of described high resolution DEM are preferably:
(4-1) least square function constrained based on gradient map and low resolution dem data is built;
(4-2) method for solving updated using iteration is found and obtains optimal solution;
If (4-3) downsampled version of optimal solution and the average reconstruction error of original low-resolution dem data exceed certain
Threshold value Th or iterationses t are inadequate, then return to step (4-2).
Wherein, described least square function is preferablyX is
Original low-resolution data, Y is desired high resolution DEM, and ▽ Y are its correspondence gradient,To obtain in step (3)
Estimation gradient, s is sample rate, ↓ represent it is down-sampled, β is weight, for controlling the balance of original DEM weights and gradient region
Coefficient.The strategy that iteration updates is preferably gradient descent method.Such as:
Wherein t is iterationses, and τ is iteration step length, ↑ it is up-sampling operation, div is divergence computing.
Parameter threshold Th is preferably 5 during the height of described high resolution DEM is rebuild, and weight beta is 0.03.Iterationses t
For 150 times, iteration step length τ is 0.2.
Preferably, the up-sampling and down-sampling can take interpolation method or the dot interlace method of sampling.
Preferably, the interpolation method is nearest-neighbor interpolation, bilinear interpolation or bicubic interpolation.
Preferably, described arithmetic operators can for Sobel operators, Roberts operators, Prewitt operators or
Canny operators.
It is below specific embodiment:
Embodiment 1
The present embodiment 1 is comprised the following steps:
(1) low resolution dem data is expanded into s times using bicubic interpolation method, is obtained and desired high resolution DEM
The quasi- high resolution DEM data of the same scale size of data.
(2) edge graph on the X and Y-direction of quasi- high resolution DEM data is obtained using Sobel operator extractions.
(3) gradient map is input to the super-resolution convolutional Neural net obtained according to the training of high-low resolution view data
In network, the estimation gradient map of high resolution DEM data is obtained;
Wherein, the super-resolution convolutional neural networks that the high-low resolution view data training is obtained, according to as follows
Step is obtained:
(3-1) great amount of images high-resolution data is obtained, all images in image data base is carried out being adopted under first bicubic
Sample, the again operation of bicubic up-sampling obtain corresponding low resolution image data, and high-low resolution view data is utilized
Sobel operator extractions gradient map is simultaneously cut into small pieces, using the gradient block of corresponding high-low resolution view data as training sample
Method, wherein low resolution gradient block size are 33, and high-resolution gradient block size is 21;
(3-2) convolutional neural networks are built, Fig. 2 illustrates the model of convolutional neural networks, including the convolution being sequentially connected
Layer, excitation layer, convolutional layer, excitation layer and convolutional layer.The as corresponding high-resolution data of last convolutional layer output.
Arranging model parameter is:First convolutional layer template size is set to 1 × 9 × 9 × 64, second convolutional layer template
64 × 1 × 1 × 32 are dimensioned to, last convolutional layer template size is set to 32 × 5 × 5 × 1.Excitation layer is adopted
ReLU (max (0, x)).
(3-3) obtain in gradient block training step (3-2) of the high-low resolution view data of acquisition according in step (3-1)
The convolutional neural networks for obtaining, obtain described super-resolution convolutional neural networks;
It is 0.001 to arrange learning rate, often trains 10000 learning rates to drop to the 1/5 of former learning rate.Using stochastic gradient
Descent method carries out the iteration of weight and updates.
(4) by the constraint of the estimation gradient map that obtains in step (3) and original low-resolution dem data, height is reconstructed
The height map of resolution DEM.
Wherein, the height process of reconstruction of described high resolution DEM is concretely comprised the following steps:
(4-1) least square function constrained based on gradient map and low resolution dem data is built;
(4-2) method for solving updated using iteration finds optimal solution;
If (4-3) downsampled version of optimal solution and the average reconstruction error of original low-resolution dem data exceed certain
Threshold value Th or iterationses t are inadequate, then return to step (4-2).
Wherein, described least square function is preferablyX is
Original low-resolution data, Y is desired high resolution DEM, and ▽ Y are its correspondence gradient,To obtain in step (3)
Estimation gradient, s is sample rate, ↓ represent it is down-sampled, β is weight, for controlling the balance of original DEM weights and gradient region
Coefficient.The strategy that iteration updates is preferably gradient descent method.Such as:
Wherein t is iterationses, and τ is iteration step length, ↑ it is up-sampling operation, div is divergence computing.
Parameter threshold Th is set to 5 during the height of described high resolution DEM is rebuild, and weight beta is 0.03.Iterationses t
For 150 times, iteration step length τ is 0.2.
Table 1 illustrates the knot of the result under 2 groups of test datas are amplified at different s times and traditional bicubic interpolation method
Fruit carries out qualitatively quality evaluation.Fig. 3 is the result schematic diagram under DEM2 amplifies at 4 times.Wherein Fig. 3 (a) is true figure, Fig. 3
B () is bicubic interpolation result, Fig. 3 (c) is test result.
Table 1
Can significantly find out from table 1 and Fig. 3, the more traditional interpolation method under different data of the result that we obtain
There is significant precision improvement, and details becomes apparent from.Therefore the reasonability of our methods is also demonstrated.
Unspecified each function (such as various arithmetic operators, argmin (Y) function) can be in the present invention
The usual definition of art of mathematics;Up-sampling, the processing data method (such as interpolation method, the dot interlace method of sampling) of down-sampling
The prior art of reference picture process field.For example,Representing matrix (X-Y ↓s) two normal forms.
As it will be easily appreciated by one skilled in the art that the foregoing is only presently preferred embodiments of the present invention, not to
The present invention, all any modification, equivalent and improvement made within the spirit and principles in the present invention etc. are limited, all should be included
Within protection scope of the present invention.
Claims (10)
1. a kind of DEM super-resolution methods based on convolutional neural networks, it is characterised in that comprise the following steps:
(1) according to mutual corresponding low resolution image data and high resolution image data in advance, the oversubscription for obtaining is trained
Resolution convolutional neural networks;
(2) pending low resolution dem data is expanded into s times using interpolation method, is obtained and desired high resolution DEM number
According to the quasi- high resolution DEM data of same scale size;
(3) gradient map of the quasi- high resolution DEM data is obtained using arithmetic operators;
(4) gradient map obtained in the step (3) is input to the super-resolution obtained in the step (1) to use
In convolutional neural networks, obtain the estimation gradient map of high resolution DEM data and estimate Grad accordingly;
(5) based on the estimation gradient map obtained in the step (4) and the pact of the pending low resolution dem data
Beam, reconstructs the height map of high resolution DEM.
2. the DEM super-resolution methods of convolutional neural networks are based on as claimed in claim 1, it is characterised in that the step (1)
Specifically include following steps:
(1-1) high resolution image data is carried out degrading processing and obtains corresponding low resolution image data, then utilized
Arithmetic operators extract the gradient map for obtaining the high resolution image data and the low resolution image data, connect
And each gradient map be divided into into multiple gradient blocks, then choose corresponding high resolution image data gradient block and
The gradient block of low resolution image data is used as training sample;
(1-2) convolutional neural networks are built and model parameter is set;
(1-3) convolution god in the step (1-2) is trained according to the training sample obtained in the step (1-1)
Jing networks, obtain super-resolution convolutional neural networks.
3. the DEM super-resolution methods of convolutional neural networks are based on as claimed in claim 2, it is characterised in that the step (1-
2) convolutional neural networks in, including multiple convolutional layers, the adjacent convolutional layer of any two is connected by excitation layer;
Wherein, the first convolutional layer is used to be input into low resolution image data, and the end convolutional layer is used to export high resolution graphics
As data.
4. the DEM super-resolution methods of convolutional neural networks are based on as claimed in claim 1, it is characterised in that the step (5)
Specifically include following steps:
(5-1) build one based on the estimation gradient map and the pending low resolution dem data constraint, with regard to mesh
The least square function of absolute altitude resolution dem data;
(5-2) method for solving updated using iteration finds the optimal solution of least square function described in the step (5-1);
If (5-3) carrying out the data that obtain after down-sampled process and the pending low resolution DEM number to the optimal solution
Average reconstruction error according between exceedes threshold value Th set in advance or iterationses are not up to requirement set in advance, then return
Return the step (5-2);Otherwise, the optimal solution corresponds to final high resolution DEM data, and the optimal solution is used to rebuild
The height map of the high resolution DEM.
5. the DEM super-resolution methods of convolutional neural networks are based on as claimed in claim 4, it is characterised in that the step (5-
1) in, the least square function isWherein, X waits to locate for described
The low resolution dem data of reason;Y is the high resolution DEM data of target,For the corresponding gradients of Y,For the step
(4) the estimation Grad of the high resolution DEM data obtained in;↓sS times of down-sampled process of expression, Y ↓sIt is that s times is carried out to Y
It is down-sampled to process the data for obtaining;β is weight factor set in advance.
6. the DEM super-resolution methods of convolutional neural networks are based on as claimed in claim 4, it is characterised in that the step (5-
2) method for solving that iteration described in updates is gradient descent method;Preferably, the Y after the t+1 time iteration updatest+1Meet:
Wherein, YtIt is the data after the t time iteration updates;τ is iteration step length set in advance;Y is after the t+1 time iteration updates
Yt+1;X is the pending low resolution dem data;↓sS times of down-sampled process of expression, Y ↓sBe Y is carried out s times it is down-sampled
The data that process is obtained;↑sRepresent that s times of up-sampling is processed, (X-Y ↓s)↑sBe to (X-Y ↓s) carry out what s times of up-sampling process was obtained
Data;For the corresponding gradients of Y,The estimation gradient of the high resolution DEM data to obtain in the step (4)
Value;β is weight factor set in advance;Div is divergence computing.
7. the DEM super-resolution methods of convolutional neural networks are based on as described in claim 5 or 6, it is characterised in that the step
(5-3) in, Th is preferably 5;
In the step (5-1) and the step (5-2), β is preferably 0.03.
8. the DEM super-resolution methods of convolutional neural networks are based on as claimed in claim 6, it is characterised in that the step (5-
2) in, maximum iteration time is preferably 150 times, and iteration step length τ is preferably 0.2.
9. the DEM super-resolution methods of convolutional neural networks are based on as described in claim 1-8 any one, it is characterised in that
The up-sampling is processed and the down-sampled process, is to take interpolation method or the dot interlace method of sampling;Preferably, the interpolation side
Method is nearest-neighbor interpolation, bilinear interpolation or bicubic interpolation;
The arithmetic operators can be Sobel operators, Roberts operators, Prewitt operators or Canny operators.
10. the DEM super-resolution methods of convolutional neural networks are based on as claimed in claim 1, it is characterised in that the step
(3) in, the gradient map of the quasi- high resolution DEM data includes the edge graph in X-direction and Y-direction, the X-direction and described
Y-direction is mutually perpendicular to.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611159517.0A CN106600553B (en) | 2016-12-15 | 2016-12-15 | DEM super-resolution method based on convolutional neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611159517.0A CN106600553B (en) | 2016-12-15 | 2016-12-15 | DEM super-resolution method based on convolutional neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106600553A true CN106600553A (en) | 2017-04-26 |
CN106600553B CN106600553B (en) | 2019-12-17 |
Family
ID=58801511
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611159517.0A Expired - Fee Related CN106600553B (en) | 2016-12-15 | 2016-12-15 | DEM super-resolution method based on convolutional neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106600553B (en) |
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107578375A (en) * | 2017-08-21 | 2018-01-12 | 北京陌上花科技有限公司 | Image processing method and device |
CN107689034A (en) * | 2017-08-16 | 2018-02-13 | 清华-伯克利深圳学院筹备办公室 | A kind of training method of neutral net, denoising method and device |
CN107945134A (en) * | 2017-11-30 | 2018-04-20 | 北京小米移动软件有限公司 | Image processing method and device |
WO2018205627A1 (en) * | 2017-05-08 | 2018-11-15 | 京东方科技集团股份有限公司 | Image processing system and method, and display apparatus |
CN109584164A (en) * | 2018-12-18 | 2019-04-05 | 华中科技大学 | Medical image super-resolution three-dimensional rebuilding method based on bidimensional image transfer learning |
CN110321913A (en) * | 2018-03-30 | 2019-10-11 | 杭州海康威视数字技术股份有限公司 | A kind of text recognition method and device |
CN110706166A (en) * | 2019-09-17 | 2020-01-17 | 中国科学院遥感与数字地球研究所 | Image super-resolution reconstruction method and device for sharpening label data |
CN110827375A (en) * | 2019-10-31 | 2020-02-21 | 湖北大学 | Infrared image true color coloring method and system based on low-light-level image |
CN111008640A (en) * | 2019-10-17 | 2020-04-14 | 平安科技(深圳)有限公司 | Image recognition model training and image recognition method, device, terminal and medium |
CN112200751A (en) * | 2020-10-23 | 2021-01-08 | 华强方特(深圳)电影有限公司 | Image enhancement method |
CN112419146A (en) * | 2019-08-20 | 2021-02-26 | 武汉Tcl集团工业研究院有限公司 | Image processing method and device and terminal equipment |
CN112669209A (en) * | 2020-12-24 | 2021-04-16 | 华中科技大学 | Three-dimensional medical image super-resolution reconstruction method and system |
CN112907733A (en) * | 2021-02-23 | 2021-06-04 | 北京华清易通科技有限公司 | Method and device for reconstructing three-dimensional model and three-dimensional model acquisition and reconstruction system |
CN113362384A (en) * | 2021-06-18 | 2021-09-07 | 安徽理工大学环境友好材料与职业健康研究院(芜湖) | High-precision industrial part measurement algorithm of multi-channel sub-pixel convolution neural network |
WO2021258529A1 (en) * | 2020-06-22 | 2021-12-30 | 北京大学深圳研究生院 | Image resolution reduction and restoration method, device, and readable storage medium |
CN114331842A (en) * | 2021-12-28 | 2022-04-12 | 武汉大学 | DEM super-resolution reconstruction method combined with topographic features |
CN116485641A (en) * | 2023-01-03 | 2023-07-25 | 南京大学 | Unsupervised DEM super-resolution reconstruction method integrating priori and terrain constraints |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103390266A (en) * | 2013-07-31 | 2013-11-13 | 广东威创视讯科技股份有限公司 | Image super-resolution method and device |
CN103455709A (en) * | 2013-07-31 | 2013-12-18 | 华中科技大学 | Super-resolution method and system for digital elevation model |
CN106228512A (en) * | 2016-07-19 | 2016-12-14 | 北京工业大学 | Based on learning rate adaptive convolutional neural networks image super-resolution rebuilding method |
-
2016
- 2016-12-15 CN CN201611159517.0A patent/CN106600553B/en not_active Expired - Fee Related
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103390266A (en) * | 2013-07-31 | 2013-11-13 | 广东威创视讯科技股份有限公司 | Image super-resolution method and device |
CN103455709A (en) * | 2013-07-31 | 2013-12-18 | 华中科技大学 | Super-resolution method and system for digital elevation model |
CN106228512A (en) * | 2016-07-19 | 2016-12-14 | 北京工业大学 | Based on learning rate adaptive convolutional neural networks image super-resolution rebuilding method |
Non-Patent Citations (2)
Title |
---|
JINGXU CHEN等: "Single Image Super-Resolution Based on Deep Learning and Gradient Transformation", 《2016 IEEE 13TH INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING(ICSP)》 * |
ZIXUAN CHEN等: "CONVOLUTIONAL NEURAL NETWORK BASED DEM SUPER RESOLUTION", 《THE INTERNATIONAL ARCHIVES OF THE PHOTOGRAMMETRY, REMOTE SENSING AND SPATIAL INFORMATION SCIENCES》 * |
Cited By (29)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2018205627A1 (en) * | 2017-05-08 | 2018-11-15 | 京东方科技集团股份有限公司 | Image processing system and method, and display apparatus |
US11216910B2 (en) | 2017-05-08 | 2022-01-04 | Boe Technology Group Co., Ltd. | Image processing system, image processing method and display device |
CN107689034A (en) * | 2017-08-16 | 2018-02-13 | 清华-伯克利深圳学院筹备办公室 | A kind of training method of neutral net, denoising method and device |
CN107689034B (en) * | 2017-08-16 | 2020-12-01 | 清华-伯克利深圳学院筹备办公室 | Denoising method and denoising device |
CN107578375A (en) * | 2017-08-21 | 2018-01-12 | 北京陌上花科技有限公司 | Image processing method and device |
CN107578375B (en) * | 2017-08-21 | 2020-10-02 | 北京陌上花科技有限公司 | Image processing method and device |
CN107945134B (en) * | 2017-11-30 | 2020-10-09 | 北京小米移动软件有限公司 | Image processing method and device |
CN107945134A (en) * | 2017-11-30 | 2018-04-20 | 北京小米移动软件有限公司 | Image processing method and device |
CN110321913B (en) * | 2018-03-30 | 2023-07-25 | 杭州海康威视数字技术股份有限公司 | Text recognition method and device |
CN110321913A (en) * | 2018-03-30 | 2019-10-11 | 杭州海康威视数字技术股份有限公司 | A kind of text recognition method and device |
CN109584164A (en) * | 2018-12-18 | 2019-04-05 | 华中科技大学 | Medical image super-resolution three-dimensional rebuilding method based on bidimensional image transfer learning |
CN112419146A (en) * | 2019-08-20 | 2021-02-26 | 武汉Tcl集团工业研究院有限公司 | Image processing method and device and terminal equipment |
CN112419146B (en) * | 2019-08-20 | 2023-12-29 | 武汉Tcl集团工业研究院有限公司 | Image processing method and device and terminal equipment |
CN110706166A (en) * | 2019-09-17 | 2020-01-17 | 中国科学院遥感与数字地球研究所 | Image super-resolution reconstruction method and device for sharpening label data |
CN110706166B (en) * | 2019-09-17 | 2022-03-18 | 中国科学院空天信息创新研究院 | Image super-resolution reconstruction method and device for sharpening label data |
WO2021052261A1 (en) * | 2019-09-17 | 2021-03-25 | 中国科学院空天信息创新研究院 | Image super-resolution reconstruction method and apparatus for sharpening of label data |
CN111008640A (en) * | 2019-10-17 | 2020-04-14 | 平安科技(深圳)有限公司 | Image recognition model training and image recognition method, device, terminal and medium |
CN111008640B (en) * | 2019-10-17 | 2024-03-19 | 平安科技(深圳)有限公司 | Image recognition model training and image recognition method, device, terminal and medium |
CN110827375A (en) * | 2019-10-31 | 2020-02-21 | 湖北大学 | Infrared image true color coloring method and system based on low-light-level image |
CN110827375B (en) * | 2019-10-31 | 2023-05-30 | 湖北大学 | Infrared image true color coloring method and system based on low-light-level image |
WO2021258529A1 (en) * | 2020-06-22 | 2021-12-30 | 北京大学深圳研究生院 | Image resolution reduction and restoration method, device, and readable storage medium |
CN112200751A (en) * | 2020-10-23 | 2021-01-08 | 华强方特(深圳)电影有限公司 | Image enhancement method |
CN112669209A (en) * | 2020-12-24 | 2021-04-16 | 华中科技大学 | Three-dimensional medical image super-resolution reconstruction method and system |
CN112907733A (en) * | 2021-02-23 | 2021-06-04 | 北京华清易通科技有限公司 | Method and device for reconstructing three-dimensional model and three-dimensional model acquisition and reconstruction system |
CN113362384A (en) * | 2021-06-18 | 2021-09-07 | 安徽理工大学环境友好材料与职业健康研究院(芜湖) | High-precision industrial part measurement algorithm of multi-channel sub-pixel convolution neural network |
CN114331842A (en) * | 2021-12-28 | 2022-04-12 | 武汉大学 | DEM super-resolution reconstruction method combined with topographic features |
CN114331842B (en) * | 2021-12-28 | 2024-04-05 | 武汉大学 | DEM super-resolution reconstruction method combining topographic features |
CN116485641A (en) * | 2023-01-03 | 2023-07-25 | 南京大学 | Unsupervised DEM super-resolution reconstruction method integrating priori and terrain constraints |
CN116485641B (en) * | 2023-01-03 | 2024-06-21 | 南京大学 | Unsupervised DEM super-resolution reconstruction method integrating priori and terrain constraints |
Also Published As
Publication number | Publication date |
---|---|
CN106600553B (en) | 2019-12-17 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106600553A (en) | DEM super-resolution method based on convolutional neural network | |
CN110119780B (en) | Hyper-spectral image super-resolution reconstruction method based on generation countermeasure network | |
CN106796716B (en) | For providing the device and method of super-resolution for low-resolution image | |
CN105279790B (en) | Fracture network 3-dimensional digital rock core modeling method | |
CN108550115A (en) | A kind of image super-resolution rebuilding method | |
CN106600683B (en) | A kind of adaptive method for reconstructing of grid model towards Bone CT sequence image | |
DE60130680T2 (en) | METHOD FOR ORIENTING A POINT GRILLE ON THE BASIS OF PICTURE CHARACTERISTICS | |
Shahzad et al. | TecDEM: A MATLAB based toolbox for tectonic geomorphology, Part 2: Surface dynamics and basin analysis | |
CN105678757B (en) | A kind of ohject displacement measuring method | |
CN109003229B (en) | Magnetic resonance super-resolution reconstruction method based on three-dimensional enhanced depth residual error network | |
CN102136142B (en) | Nonrigid medical image registration method based on self-adapting triangular meshes | |
CN107730451A (en) | A kind of compressed sensing method for reconstructing and system based on depth residual error network | |
CN110187143B (en) | Chromatography PIV reconstruction method and device based on deep neural network | |
CN106067161A (en) | A kind of method that image is carried out super-resolution | |
CN105006018A (en) | Three-dimensional CT core image super-resolution reconstruction method | |
CN106780458B (en) | Point cloud framework extraction method and device | |
CN102903103B (en) | Migratory active contour model based stomach CT (computerized tomography) sequence image segmentation method | |
CN107045580A (en) | A kind of shale mechanics parameter quick calculation method based on digital cores | |
CN109584164A (en) | Medical image super-resolution three-dimensional rebuilding method based on bidimensional image transfer learning | |
CN110070567A (en) | A kind of ground laser point cloud method for registering | |
CN114331842B (en) | DEM super-resolution reconstruction method combining topographic features | |
CN106600557A (en) | PSF estimation method based on hybrid Gaussian model and sparse constraints | |
CN109242771A (en) | Super-resolution image reconstruction method and device, computer-readable storage medium and computer equipment | |
CN103077559A (en) | Cluster three-dimensional rebuilding method based on sequence image | |
CN107563963B (en) | Super-resolution reconstruction method based on single depth map |
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 | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20191217 Termination date: 20201215 |
|
CF01 | Termination of patent right due to non-payment of annual fee |