CN107729922A - Remote sensing images method for extracting roads based on deep learning super-resolution technique - Google Patents
Remote sensing images method for extracting roads based on deep learning super-resolution technique Download PDFInfo
- Publication number
- CN107729922A CN107729922A CN201710857755.7A CN201710857755A CN107729922A CN 107729922 A CN107729922 A CN 107729922A CN 201710857755 A CN201710857755 A CN 201710857755A CN 107729922 A CN107729922 A CN 107729922A
- Authority
- CN
- China
- Prior art keywords
- resolution
- mrow
- image
- deep learning
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 57
- 238000013135 deep learning Methods 0.000 title claims abstract description 18
- 238000000605 extraction Methods 0.000 claims abstract description 38
- 239000013598 vector Substances 0.000 claims abstract description 37
- 239000000284 extract Substances 0.000 claims abstract description 6
- 230000015572 biosynthetic process Effects 0.000 claims abstract description 4
- 238000013136 deep learning model Methods 0.000 claims abstract description 4
- 238000003786 synthesis reaction Methods 0.000 claims abstract description 4
- 238000013507 mapping Methods 0.000 claims description 18
- 230000006870 function Effects 0.000 claims description 9
- 238000012549 training Methods 0.000 claims description 7
- 238000001228 spectrum Methods 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 4
- 238000007619 statistical method Methods 0.000 claims description 3
- 238000001914 filtration Methods 0.000 claims 2
- 238000002474 experimental method Methods 0.000 abstract description 3
- 230000001537 neural effect Effects 0.000 abstract description 3
- 238000001514 detection method Methods 0.000 abstract description 2
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 238000004422 calculation algorithm Methods 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 239000003795 chemical substances by application Substances 0.000 description 2
- 230000000052 comparative effect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000011158 quantitative evaluation Methods 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000012407 engineering method Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000005206 flow analysis Methods 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 238000003709 image segmentation Methods 0.000 description 1
- 210000000987 immune system Anatomy 0.000 description 1
- 238000003475 lamination Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 230000000877 morphologic effect Effects 0.000 description 1
- 230000007935 neutral effect Effects 0.000 description 1
- 238000002310 reflectometry Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4053—Scaling of whole images or parts thereof, e.g. expanding or contracting based on super-resolution, i.e. the output image resolution being higher than the sensor resolution
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Processing (AREA)
Abstract
The invention provides a kind of remote sensing images method for extracting roads based on deep learning super-resolution technique, it is characterised in that comprises the following steps：Step S1, extracts overlapping image block based on deep learning model from low-resolution image, and with higher-dimension vector representation；Step S2, each high n dimensional vector n is so mapped to another high n dimensional vector n, each high n dimensional vector n represents a high-resolution block；Step S3, the expression of each high-resolution block is integrated, synthesizes final high-definition picture；Step S4, road is extracted from the high-definition picture after synthesis.The present invention combines the super-resolution rebuilding based on convolutional Neural net and road extraction, and the validity of this method is demonstrated in the detection experiment of the road in satellite remote-sensing image Shang Zuo roads.
Description
Technical field
The present invention relates to image zooming-out technical field, and in particular to a kind of remote sensing based on deep learning super-resolution technique
Image method for extracting roads.
Background technology
Road extraction is in the database update of GIS-Geographic Information System, intelligent transportation system, traffic flow analysis, emergent risk
There is important value in many applications such as management and Map Service.Nearly 40 years of road extraction research.Existing method can be with
It is divided into：Knowledge based engineering method, image segmentation and sorting technique, Mathematical Morphology Method, active contour method, Dynamic Programming and
Group technology.Wherein, sorting technique includes neutral net and deep learning method, markov random file or condition random field side
Method and method based on graph theory etc..Over time, the performance of these methods is also constantly improving.
Unfortunately, on high resolution ratio satellite remote-sensing image, it is also difficult to extract road exactly, be particularly those
The road of only several pixel wides.On high spectrum image, each pixel with low resolution can be in sub-pix rank
Mix different Land cover types.In order to solve this problem, it is proposed that Decomposition of Mixed Pixels (pixel unmixing) and
The methods of sub-pixed mapping drawing (subpixel mapping, SPM), determines the relative scale of class and in each pixel
Position, obtain the terrain classification figure of sub-pixed mapping yardstick.In order to more effective and accurate, it is proposed that many methods are with going in prognostic chart picture
The spatial distribution of thing.
For low resolution remote sensing images, towards remote sensing images be based on more agent systems (multiagent system,
MAS) adaptive sub-pixed mapping drafting algorithm (multiagent system for remote-sensing imagery,
MASSM) solves sub-pixed mapping drawing problem using three kinds of agents.Sub-pixed mapping drawing (Spatial based on Spatial Distribution Pattern
Distribution pattern-based subpixel mapping, SPMS) calculate the spatial distribution mould of each geographic object
Formula.First, it classifies subjects into point Linear and face shape pattern according to space geometry feature, and then SPMS makes to corresponding object
With the different SPM algorithms based on matching.Finally, all images are integrated into sub-pixed mapping drawing.
For high-resolution remote sensing image, common method includes pixel exchange, sub-pixel/pixel space attracts model, line
Property optimisation technique, instruction cokriging, Markov random field, spatial regularization, method of geometry, artificial immune system, differential
Evolve and maximum a posteriori model.The potential hypothesis of these methods be between pixel and between spatial dependence be it is applicable, its
Reason is the fine granularity details that high-resolution remote sensing image can provide object of interest, and maximizes the space in class between class
Correlation.
Prior art does not combine the super-resolution rebuilding based on convolutional Neural net and road extraction.After reconstruction
Image has very high image resolution ratio.However, result images and the same high image of resolution ratio are different after all.If use resolution ratio
Sample training in same high image, is tested the image after super-resolution rebuilding, its result will be by a certain extent
Influence.
The content of the invention
High-resolution remote sensing image has reached sub-meter grade, and the resolution ratio highest of wherein satellite image can reach every pixel
0.4 meter or so.On the other hand, high-precision map precision prescribed is at least up to 0.1 meter.Therefore, if made according to satellite image
High-precision map is difficult to reach permissible accuracy.On the other hand, needing to do super-resolution rebuilding to image, carried from the image after reconstruction
It by way of road, can be only achieved with sub-pixel precision, the requirement of high-precision map could be met.
Based on this, present invention employs following technical scheme：
A kind of remote sensing images method for extracting roads based on deep learning super-resolution technique, comprises the following steps：
Step S1, extracts overlapping image block based on deep learning model from low-resolution image, and with high n dimensional vector n
Represent；
Step S2, each high n dimensional vector n is so mapped to another high n dimensional vector n, each high n dimensional vector n represents a height
Resolution ratio block；
Step S3, the expression of each high-resolution block is integrated, synthesizes final high-definition picture；
Step S4, road is extracted from the high-definition picture after synthesis.
Further, the super-resolution of low-resolution image is used for using SRCNN.
Further, the high n dimensional vector n in step S1 includes one group of characteristic pattern, and the wherein quantity of characteristic pattern is sweared equal to higher-dimension
The dimension of amount.
Further, in first layer SRCNN, image block operator F1Carried out by the set filter in first layer as follows
Convolution：
F1(Y)=max { 0, W1*Y+A1} (1)
Wherein, Y represents the image block of extraction, F1(Y) image processing operators, W are represented1Expression size is c × f1×f1N1
Individual wave filter, c are the port numbers of input picture, f1It is the size of wave filter, * represents convolution operation, A1It is biasing, max represents to take
Maximum.
Further, in step S2, based on operator F2, use n2The wave filter of individual 1 × 1 size is by each n1The vector of dimension
Be converted to n2The vector of dimension：
F2(Y)=max { 0, W2*F1(Y)+A2} (2)
Wherein, F2(Y) vector mapping operator, W are represented2It is n2Individual size is n1×f2×f2Wave filter, A2It is to bias, f2It is
The size of wave filter, each output size are n2One high-resolution block of vector representation of dimension.
Further, in step S3, convolutional layer and operator F are used3High-resolution block is optimized, obtains final height
Image in different resolution：
F3(Y)=W3*F2(Y)+A3 (3)
Wherein, F3(Y) Image Reconstruction operator, W are represented3It is that c size is n2×f3×f3Wave filter, A3It is to bias, f3
It is the size of wave filter.
Further, step S1, step S2 and step S3 are incorporated into SRCNN, mapping function F is by { F1, F2, F3Group
It is the parameter in SRCNN into, Θ, Θ={ W1, W2, W3, A1, A2, A3}；By minimizing reconstructed image F (Y；It is Θ) and corresponding
True high-definition picture { XiBetween loss optimize Θ, to make low-resolution image be { Yi, loss function mean square error
It is defined as：
Wherein, L (Θ) represents loss function, F (Yi；Θ) represent to low-resolution image { YiReconstruction result, i is sample
This sequence number, k are training samples numbers.
Further, high-definition picture is classified with OCSVM.
Further, step S4 specifically includes following steps：
Step S41, the information for inputting high-definition picture are mapped in high-dimensional feature space；
Step S42, then OCSVM search out that to separate the maximum border of training data super flat in high-dimensional feature space
Face；
Step S43, road is extracted according to the spectrum information of road；
Step S44, collect road pixel and carry out statistical analysis.
Further, the bounding hyperplane in the step S42 corresponds to following classifying rules：
Formula (5) needs to meet following condition：
(ω·φ(zj))≥ρ-ξj, and ξj≥0
Wherein, min represents to take minimum value, and w is normal vector,Represent total sample number, ξjIt is slack variable, j represents that sample exists
Sequence number in sample set, zjRepresent sample set in j-th of sample, v represent value (0,1] variable, ξjAlso it is variable with ρ,
φ represents Feature Mapping relation.
Then the beneficial effects of the present invention are be amplified to sub-pix rank again using super-resolution technique by original image
Extract road, on the one hand, reduce the error rate during extraction of road spectrum information；On the other hand, super-resolution also isolates some
Road waypoint connect into a line so that road extraction is more perfect.
Brief description of the drawings
Fig. 1 is the remote sensing images method for extracting roads frame diagram of the invention based on deep learning super-resolution technique.
Fig. 2 is the result of super-resolution method of the present invention, and wherein Fig. 2-1 is original image, and Fig. 2-2 is super-resolution figure
Picture.
Fig. 3 is piece image road extraction result, and wherein Fig. 3-1 is original image, and Fig. 3-2 is super-resolution image,
Fig. 3-3 is reference picture, and Fig. 3-4 is the road extraction of original image, and Fig. 3-5 is the road extraction of super-resolution image, Fig. 3-6
For the mark figure of the road extraction of original image, Fig. 3-7 is the road extraction mark figure of super-resolution image.
Fig. 4 is second image road extraction result, wherein, Fig. 4-1 is original image, and Fig. 4-2 is super-resolution image,
Fig. 4-3 is reference picture, and Fig. 4-4 is the road extraction of original image, and Fig. 4-5 is the road extraction of super-resolution image, Fig. 4-6
For the mark figure of the road extraction of original image, Fig. 4-7 is the road extraction mark figure of super-resolution image.
Embodiment
The present invention combines the super-resolution rebuilding based on convolutional Neural net and road extraction, in satellite remote-sensing image
The road detection experiment in Shang Zuo roads demonstrates the validity of this method.Hereinafter, the present invention is made into one in conjunction with the accompanying drawings and embodiments
Step illustrates.
Fig. 1 is remote sensing images method for extracting roads framework of the invention based on deep learning super-resolution technique
Figure, comprises the following steps：
Step S1, overlapping image block is extracted from low-resolution image (original image) based on deep learning model, and
With higher-dimension vector representation.
Step S2, is then mapped to another high n dimensional vector n by each high n dimensional vector n, and each vector represents a high score
Resolution block.
Step S3, then integrated each high-resolution piece expression, synthesize final high-definition picture.
Step S4, finally, road is extracted from the high-definition picture after synthesis.
1st, Image Super-resolution
The present invention uses super-resolution convolutional neural networks (Super-Resolution Convolutional Neural
Network, SRCNN) it is used for the super-resolution of original image.
1.1st, block extracts
Substantial amounts of overlapping image block is extracted from low-resolution image, wherein the higher-dimension vector representation of each image block.This
A little vectors include one group of characteristic pattern, and the wherein quantity of map is equal to the dimension of vector.In first layer SRCNN, image block is calculated
Sub- F1Following convolution is carried out by the set filter in first layer：
F1(Y)=max { 0, W1*Y+A1} (1)
Wherein F1(Y) image processing operators are represented, Y represents the image block of extraction, W1Expression size is c × f1×f1One
Group wave filter, n1Represent W1In number of filter, * represent convolution operation, A1It is biasing, c is the port number of input picture, f1
It is the size of wave filter.
1.2nd, non-linear drawing
Based on operator F2, use n2The wave filter of individual 1 × 1 size is by each n1The vector of dimension is converted to n2The vector of dimension.
F2(Y)=max { 0, W2*F1(Y)+A2} (2)
Wherein, F2(Y) vector mapping operator, W are represented2It is that one group of size is n1×f2×f2Wave filter, n2Represent W2Middle filter
The number of ripple device, A2It is to bias, f2It is the size of wave filter.Each output size is n2One dimensional images of vector representation of dimension
Block.In order to increase nonlinear degree, more convolutional layers are added.
1.3rd, reconstruct
After Nonlinear Mapping, resulting image block needs to be grouped together, for high resolution image reconstruction.
Traditionally, the small image block of high-resolution can be averaged, and finally be merged into a complete image.In the present invention, volume is used
Lamination and operator F3Small image block is optimized, its operation can obtain following final image：
F3(Y)=W3*F2(Y)+A3 (3)
Wherein, F3(Y) Image Reconstruction operator, W are represented3It is that one group of size is n2×f3×f3Wave filter, W3Median filter
Number be c, f3It is the size of wave filter, A3It is biasing.
1.4th, train
In order to optimize weight and deviation, the operation of above three step can be incorporated into super-resolution convolutional neural networks
In.Mapping function F is by { F1, F2, F3Form, the parameter in network is represented with symbol Θ, i.e. Θ={ W1, W2, W3, A1, A2, A3}。
Reconstructed image F (Y；Θ) represent.By minimizing reconstructed image and corresponding true high-definition picture { XiBetween damage
Lose to optimize Θ, wherein i represents sample sequence number.It is { Y to make low-resolution imagei, loss function is defined with mean square error (MSE)
For
Wherein, L (Θ) represents loss function, F (Yi；Θ) represent to low-resolution image YiReconstruction result, i is sample
Sequence number, k are training samples numbers.Above formula can be reduced by stochastic gradient descent method and lost.
2nd, road extraction
In the present invention, high-definition picture one-class support vector machines (One-Class Support Vector
Machine, OCSVM) classification, specifically include following steps：
Step S41, first, input high-resolution image information are mapped in high-dimensional feature space.
Step S42, then OCSVM searched out in high-dimensional feature space can the maximum border of perfect separation training data surpass
Plane.
In order to prevent OCSVM graders over-fitting in noise data, slack variable ξ can be introducedj.The hyperplane is corresponding
In following classifying rules：
And the formula need to meet following condition：
(ω·φ(zj))≥ρ-ξj, and ξj≥0
Wherein,Total sample number is represented, j represents sequence number of the sample in sample set, zjRepresent j-th of sample in sample set
This, w is normal vector, v represent value (0,1] variable, ξjAlso it is variable with ρ, φ represents Feature Mapping relation.
Step S43, then, road is extracted according to the spectrum information of road.
Step S44, finally, collect road pixel and carry out statistical analysis.Road information is obtained by morphological method.
3rd, test
3.1st, super-resolution experimental result
In this experiment, super-resolution remote sensing images are obtained using SRCNN methods.As shown in Fig. 2 the result of super-resolution
More preferable than traditional interpolation technique, edge contour information is apparent.
3.2nd, road extraction experimental result
Experimental image and its true value figure can be from http：//www.cse.iitm.ac.in/~vplab/
Satellite.html is downloaded.The super-resolution shown in Fig. 3-2 and Fig. 4-2 can be produced from original image Fig. 3-1 and Fig. 4-1
Image.Fig. 3-3 and Fig. 4-3 is the reference picture for representing road true value figure.The road extracted from original image such as Fig. 3-4 and figure
Shown in 4-4.Extraction road after super-resolution is as shown in Fig. 3-5 and Fig. 4-5.In order to facilitate comparative result, Image Adjusting is to identical
Width and height.The present invention is using conventional extraction evaluation index ----integrality and quality, for accurate comparative result.It is complete
Whole property is the fraction of correct extraction road pixel, and quality is the wellness of final extraction result.Road extraction result it is complete
Property and Mass Calculation are as follows：
Wherein, Completeness is integrality, and Quality is quality, and TP is true positives, i.e., the road length correctly extracted
Degree, FP is false positive, i.e., the link length of incorrect extraction, FN is false negative, i.e., does not extract the length of road.As Fig. 3-6,
Fig. 3-7, Fig. 4-6, shown in Fig. 4-7.
3.3rd, super-resolution is tested
As can be seen that some narrow roads can be identified using the road extraction of super-resolution from Fig. 3 and Fig. 4.
In terms of the integrality and quality of road extraction, the result obtained after super-resolution is better than original image.Reason is super-resolution
Reconstruction can increase information content：On the one hand, the error rate during extraction of road spectrum information is reduced.On the other hand, super-resolution
Some isolated road waypoints are connected into a line, so that road extraction is more perfect.
As shown in Figure 3-4, many building pixels close to road are divided into road, because building has and road phase
As spectral reflectivity.In figs. 3-5, after super-resolution, some adjacent architectural pixels are correctly extracted, because super-resolution
Building and road are distinguished by enlarged drawing details.In Fig. 4, some roads are covered by some automobiles and trees, some roads
Road is narrow.In this case, correctly extraction road is not easy to completely.As shown in Fig. 4-4, extracting section road seriously breaks
Open.After super-resolution, as illustrated in figures 4-5, road, particularly narrow road is generally connected.
According to the formula of calculation of integrity and quality, the result of each index is showed in Tables 1 and 2.From the number in table
There is higher integrality and extraction quality according to the road extraction result that can be seen that super-resolution image.In general, oversubscription
Road extraction in resolution image can obtain better result.
The quantitative evaluation of the Fig. 3 of table 1 road extraction
| Original image | Super-resolution image | |
| TP | 13313 | 12512 |
| FP | 13358 | 6781 |
| FN | 973 | 1774 |
| Integrality | 49.92% | 64.85% |
| Quality | 48.16% | 59.39% |
The quantitative evaluation of the Fig. 4 of table 2 road extraction
| Original image | Super-resolution image | |
| TP | 6071 | 6753 |
| FP | 1856 | 1738 |
| FN | 7926 | 7244 |
| Integrality | 76.59% | 79.53% |
| Quality | 38.30% | 42.92% |
Present invention preferably uses the Python of PyTorch platforms, processing speed can reach optimum efficiency.
Super-resolution reconstruction established model used in the present invention can select other super-resolution models, and grader can also be adopted
With other graders.
Although the present invention is disclosed as above with preferred embodiment, it is not for limiting the present invention, any this area
Technical staff without departing from the spirit and scope of the present invention, may be by the methods and technical content of the disclosure above to this hair
Bright technical scheme makes possible variation and modification, therefore, every content without departing from technical solution of the present invention, according to the present invention
Any simple modifications, equivalents, and modifications made to above example of technical spirit, belong to technical solution of the present invention
Protection domain.
Claims (10)
1. a kind of remote sensing images method for extracting roads based on deep learning super-resolution technique, it is characterised in that including following
Step：
Step S1, extracts overlapping image block based on deep learning model from low-resolution image, and with higher-dimension vector representation；
Step S2, each high n dimensional vector n is so mapped to another high n dimensional vector n, each high n dimensional vector n represents a high-resolution
Rate block；
Step S3, the expression of each high-resolution block is integrated, synthesizes final high-definition picture；
Step S4, road is extracted from the high-definition picture after synthesis.
2. a kind of remote sensing images method for extracting roads based on deep learning super-resolution technique as claimed in claim 1, its
It is characterised by, the super-resolution of low-resolution image is used for using SRCNN.
3. a kind of remote sensing images method for extracting roads based on deep learning super-resolution technique as claimed in claim 1, its
It is characterised by, the high n dimensional vector n in step S1 includes one group of characteristic pattern, and the wherein quantity of characteristic pattern is equal to the dimension of high n dimensional vector n.
4. a kind of remote sensing images method for extracting roads based on deep learning super-resolution technique as claimed in claim 3, its
It is characterised by, in first layer SRCNN, image block operator F1Following convolution is carried out by the set filter in first layer：
F1(Y)=max { 0, W1*Y+A1} (1)
Wherein, Y represents the image block of extraction, F1(Y) image processing operators, W are represented1Expression size is c × f1×f1N1Individual filter
Ripple device, c are the port numbers of input picture, f1It is the size of wave filter, * represents convolution operation, A1It is biasing, max represents to take maximum
Value.
5. a kind of remote sensing images method for extracting roads based on deep learning super-resolution technique as claimed in claim 4, its
It is characterised by, in step S2, based on operator F2, use n2The wave filter of individual 1 × 1 size is by each n1The vector of dimension is converted to n2
The vector of dimension：
F2(Y)=max { 0, W2*F1(Y)+A2} (2)
Wherein, F2(Y) vector mapping operator, W are represented2It is n2Individual size is n1×f2×f2Wave filter, A2It is to bias, f2It is filtering
The size of device, each output size are n2One high-resolution block of vector representation of dimension.
6. a kind of remote sensing images method for extracting roads based on deep learning super-resolution technique as claimed in claim 5, its
It is characterised by, in step S3, uses convolutional layer and operator F3High-resolution block is optimized, obtains final high resolution graphics
Picture：
F3(Y)=W3*F2(Y)+A3 (3)
Wherein, F3(Y) Image Reconstruction operator, W are represented3It is that c size is n2×f3×f3Wave filter, A3It is to bias, f3It is filtering
The size of device.
7. a kind of remote sensing images method for extracting roads based on deep learning super-resolution technique as claimed in claim 6, its
It is characterised by, step S1, step S2 and step S3 is incorporated into SRCNN, mapping function F is by { F1, F2, F3Form, Θ is
Parameter in SRCNN, Θ={ W1, W2, W3, A1, A2, A3}；By minimizing reconstructed image F (Y；Θ) and corresponding true high score
Resolution image { XiBetween loss optimize Θ, to make low-resolution image be { Yi, loss function is defined as with mean square error：
<mrow>
<mi>L</mi>
<mrow>
<mo>(</mo>
<mi>&Theta;</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mi>k</mi>
</mfrac>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>k</mi>
</munderover>
<mo>|</mo>
<mo>|</mo>
<mi>F</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>Y</mi>
<mi>i</mi>
</msub>
<mo>;</mo>
<mi>&Theta;</mi>
<mo>)</mo>
</mrow>
<mo>-</mo>
<msub>
<mi>X</mi>
<mi>i</mi>
</msub>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>4</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein, L (Θ) represents loss function, F (Yi；Θ) represent to low-resolution image { YiReconstruction result, i is sample sequence
Number, k is training samples number.
8. a kind of remote sensing images method for extracting roads based on deep learning super-resolution technique as claimed in claim 1, its
It is characterised by, high-definition picture is classified with OCSVM.
9. a kind of remote sensing images method for extracting roads based on deep learning super-resolution technique as claimed in claim 8, its
It is characterised by, step S4 specifically includes following steps：
Step S41, the information for inputting high-definition picture are mapped in high-dimensional feature space；
Step S42, then OCSVM is searched out in high-dimensional feature space can separate the maximum bounding hyperplane of training data；
Step S43, road is extracted according to the spectrum information of road；
Step S44, collect road pixel and carry out statistical analysis.
10. a kind of remote sensing images method for extracting roads based on deep learning super-resolution technique as claimed in claim 9, its
It is characterised by, the bounding hyperplane in the step S42 corresponds to following classifying rules：
<mrow>
<munder>
<mi>min</mi>
<mrow>
<mi>&omega;</mi>
<mo>,</mo>
<msub>
<mi>&xi;</mi>
<mi>j</mi>
</msub>
<mo>,</mo>
<mi>&rho;</mi>
</mrow>
</munder>
<mfrac>
<mn>1</mn>
<mn>2</mn>
</mfrac>
<mo>|</mo>
<mo>|</mo>
<mi>&omega;</mi>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
<mo>+</mo>
<mfrac>
<mn>1</mn>
<mrow>
<mi>v</mi>
<mi>l</mi>
</mrow>
</mfrac>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>l</mi>
</munderover>
<msub>
<mi>&xi;</mi>
<mi>j</mi>
</msub>
<mo>-</mo>
<mi>&rho;</mi>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>5</mn>
<mo>)</mo>
</mrow>
</mrow>
Formula (5) needs to meet following condition：
(ω·φ(zj))≥ρ-ξj, and ξj≥0
Wherein, min represents to take minimum value, and w is normal vector, and l represents total sample number, ξjIt is slack variable, j represents sample in sample
The sequence number of concentration, zjRepresent sample set in j-th of sample, v represent value (0,1] variable, ξjAlso it is variable with ρ, φ tables
Show Feature Mapping relation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710857755.7A CN107729922A (en) | 2017-09-20 | 2017-09-20 | Remote sensing images method for extracting roads based on deep learning super-resolution technique |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710857755.7A CN107729922A (en) | 2017-09-20 | 2017-09-20 | Remote sensing images method for extracting roads based on deep learning super-resolution technique |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107729922A true CN107729922A (en) | 2018-02-23 |
Family
ID=61207256
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710857755.7A Pending CN107729922A (en) | 2017-09-20 | 2017-09-20 | Remote sensing images method for extracting roads based on deep learning super-resolution technique |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107729922A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110298211A (en) * | 2018-03-21 | 2019-10-01 | 北京大学 | A kind of Methods Deriving Drainage Network based on deep learning and high-resolution remote sensing image |
| CN111157114A (en) * | 2019-12-26 | 2020-05-15 | 西安电子科技大学 | Long-wave infrared multispectral imaging method and device based on wavelength conversion |
| CN112613371A (en) * | 2020-12-16 | 2021-04-06 | 上海大学 | Hyperspectral image road extraction method based on dense connection convolution neural network |
| CN112767243A (en) * | 2020-12-24 | 2021-05-07 | 深圳大学 | Hyperspectral image super-resolution implementation method and system |
| CN113487900A (en) * | 2021-07-06 | 2021-10-08 | 北京邮电大学 | Asynchronous road information extraction system for satellite images and control method thereof |
| CN113538247A (en) * | 2021-08-12 | 2021-10-22 | 中国科学院空天信息创新研究院 | Super-resolution generation and conditional countermeasure network remote sensing image sample generation method |
| CN115100540A (en) * | 2022-06-30 | 2022-09-23 | 电子科技大学 | Method for automatically extracting high-resolution remote sensing image road |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104700100A (en) * | 2015-04-01 | 2015-06-10 | 哈尔滨工业大学 | Feature extraction method for high spatial resolution remote sensing big data |
| CN105989334A (en) * | 2015-02-12 | 2016-10-05 | 中国科学院西安光学精密机械研究所 | Monocular vision-based road detection method |
| CN106600538A (en) * | 2016-12-15 | 2017-04-26 | 武汉工程大学 | Human face super-resolution algorithm based on regional depth convolution neural network |
| CN206249426U (en) * | 2016-09-30 | 2017-06-13 | 宁波市东望智能***工程有限公司 | A kind of image restoration system |
-
2017
- 2017-09-20 CN CN201710857755.7A patent/CN107729922A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105989334A (en) * | 2015-02-12 | 2016-10-05 | 中国科学院西安光学精密机械研究所 | Monocular vision-based road detection method |
| CN104700100A (en) * | 2015-04-01 | 2015-06-10 | 哈尔滨工业大学 | Feature extraction method for high spatial resolution remote sensing big data |
| CN206249426U (en) * | 2016-09-30 | 2017-06-13 | 宁波市东望智能***工程有限公司 | A kind of image restoration system |
| CN106600538A (en) * | 2016-12-15 | 2017-04-26 | 武汉工程大学 | Human face super-resolution algorithm based on regional depth convolution neural network |
Non-Patent Citations (5)
| Title |
|---|
| CHAO DONG等: "Image Super-Resolution Using Deep Convolutional Networks", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
| 傅罡: "多源遥感数据的道路提取方法研究", 《中国博士学位论文全文数据库》 * |
| 冯爱民等: "基于核的单类分类器研究", 《南京师范大学学报（工程技术版）》 * |
| 尹传环等: "单类支持向量机的研究进展", 《计算机工程与应用》 * |
| 薄树奎等: "面向对象的遥感影像单类分类", 《现代电子技术》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110298211A (en) * | 2018-03-21 | 2019-10-01 | 北京大学 | A kind of Methods Deriving Drainage Network based on deep learning and high-resolution remote sensing image |
| CN111157114A (en) * | 2019-12-26 | 2020-05-15 | 西安电子科技大学 | Long-wave infrared multispectral imaging method and device based on wavelength conversion |
| CN111157114B (en) * | 2019-12-26 | 2022-01-28 | 西安电子科技大学 | Long-wave infrared multispectral imaging method and device based on wavelength conversion |
| CN112613371A (en) * | 2020-12-16 | 2021-04-06 | 上海大学 | Hyperspectral image road extraction method based on dense connection convolution neural network |
| CN112767243A (en) * | 2020-12-24 | 2021-05-07 | 深圳大学 | Hyperspectral image super-resolution implementation method and system |
| CN112767243B (en) * | 2020-12-24 | 2023-05-26 | 深圳大学 | Method and system for realizing super-resolution of hyperspectral image |
| CN113487900A (en) * | 2021-07-06 | 2021-10-08 | 北京邮电大学 | Asynchronous road information extraction system for satellite images and control method thereof |
| CN113538247A (en) * | 2021-08-12 | 2021-10-22 | 中国科学院空天信息创新研究院 | Super-resolution generation and conditional countermeasure network remote sensing image sample generation method |
| CN113538247B (en) * | 2021-08-12 | 2022-04-15 | 中国科学院空天信息创新研究院 | Super-resolution generation and conditional countermeasure network remote sensing image sample generation method |
| CN115100540A (en) * | 2022-06-30 | 2022-09-23 | 电子科技大学 | Method for automatically extracting high-resolution remote sensing image road |
| CN115100540B (en) * | 2022-06-30 | 2024-05-07 | 电子科技大学 | Automatic road extraction method for high-resolution remote sensing image |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107729922A (en) | Remote sensing images method for extracting roads based on deep learning super-resolution technique | |
| CN103679675B (en) | Remote sensing image fusion method oriented to water quality quantitative remote sensing application | |
| CN110287932B (en) | Road blocking information extraction method based on deep learning image semantic segmentation | |
| CN110929607B (en) | Remote sensing identification method and system for urban building construction progress | |
| Mesev | The use of census data in urban image classification | |
| CN111666918A (en) | Multi-factor-based coastline change identification method | |
| CN109657610A (en) | A kind of land use change survey detection method of high-resolution multi-source Remote Sensing Images | |
| Xu et al. | Leaf instance segmentation and counting based on deep object detection and segmentation networks | |
| CN110070518A (en) | It is a kind of based on dual path support under high spectrum image Super-resolution Mapping | |
| CN107944470A (en) | SAR image sorting technique based on profile ripple FCN CRF | |
| CN102110227A (en) | Compound method for classifying multiresolution remote sensing images based on context | |
| CN109522831B (en) | Real-time vehicle detection method based on micro-convolution neural network | |
| Peeters et al. | Automated recognition of urban objects for morphological urban analysis | |
| CN104732279A (en) | Improved cellular automaton traffic flow simulation analysis method based on geographic information system | |
| CN114332473A (en) | Object detection method, object detection device, computer equipment, storage medium and program product | |
| CN104809471A (en) | Hyperspectral image residual error fusion classification method based on space spectrum information | |
| CN109684910A (en) | A kind of method and system of network detection transmission line of electricity ground surface environment variation | |
| CN116662468A (en) | Urban functional area identification method and system based on geographic object space mode characteristics | |
| CN107392211A (en) | The well-marked target detection method of the sparse cognition of view-based access control model | |
| CN116704350B (en) | Water area change monitoring method and system based on high-resolution remote sensing image and electronic equipment | |
| CN104008383A (en) | Hyperspectral image characteristic extraction algorithm based on manifold learning linearization | |
| Engstrom et al. | Evaluating the Relationship between Contextual Features Derived from Very High Spatial Resolution Imagery and Urban Attributes: A Case Study in Sri Lanka | |
| CN116563649A (en) | Tensor mapping network-based hyperspectral image lightweight classification method and device | |
| Orun | Automated identification of man-made textural features on satellite imagery by Bayesian networks | |
| CN113486599A (en) | Method for calculating effective stroke of oil pumping unit |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180223 |
|
| RJ01 | Rejection of invention patent application after publication |