CN108764039A - Building extracting method, medium and the computing device of neural network, remote sensing image - Google Patents
Building extracting method, medium and the computing device of neural network, remote sensing image Download PDFInfo
- Publication number
- CN108764039A CN108764039A CN201810373725.3A CN201810373725A CN108764039A CN 108764039 A CN108764039 A CN 108764039A CN 201810373725 A CN201810373725 A CN 201810373725A CN 108764039 A CN108764039 A CN 108764039A
- Authority
- CN
- China
- Prior art keywords
- layer
- building
- remote sensing
- network system
- sensing image
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/176—Urban or other man-made structures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (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)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of neural network, building extracting method, medium and the computing devices of remote sensing image.Disclosed neural network, the building for remote sensing image extract, including：Input layer, the first to the 5th convolutional layer, first to fourth pond layer in VGG networks；The input terminal of first single scale fused layer, the first single scale fused layer is connected to the output end of the first convolutional layer；The input terminal of second to the 5th single scale fused layer, the second to the 5th single scale fused layer is respectively connected to the output end of the second to the 5th convolutional layer；The input terminal of first to fourth up-sampling layer, first to fourth up-sampling layer is respectively connected to the output end of the second to the 5th single scale fused layer；Multiple dimensioned splicing fused layer, the input terminal of multiple dimensioned splicing fused layer are connected to the output end of the first single scale fused layer, first to fourth up-sampling layer；Output layer.Densely distributed and various scale building can be effectively treated in disclosed neural network, improve the precision of building automation extraction.
Description
Technical field
The present invention relates to neural network and image processing field more particularly to a kind of neural network, the buildings of remote sensing image
Object extracting method, medium and computing device.
Background technology
With the rapid development of sensor technology, the spatial resolution of remote sensing image is continuously improved.By computer vision
The inspiration of field deep learning algorithm, at present scholar mostly use convolutional neural networks realize remote sensing images semantic segmentation appoint
Business.Although the method for some forefronts achieves good effect in remote sensing image semantic segmentation task, all do not have
There are the Some features for considering remote sensing image itself.First, in conventional computer vision semantic segmentation task, one in image to be detected
As only a few be distributed more loose between target to tens targets, see Fig. 1 (a).However in remote sensing image, build
It is general more intensive to build distribution, especially in settlement place region, sees Fig. 1 (b).Secondly, conventional computer vision semantic segmentation is appointed
In business, target size to be detected is generally large, and length and width are generally in tens to hundreds of pixels, and building size one in remote sensing image
As it is much smaller, scale (pixel number corresponding to the image of different buildings itself) variation it is also larger, see Fig. 1 (c).
In order to ensure the accuracy of semantic segmentation, first have to ensure the accuracy that building (feature) extracts.Although existing
Some are had existed in technology in conjunction with convolutional neural networks to extract the technical solution of certain specific objectives in remote sensing image.Example
Such as, a kind of remote sensing images road based on full convolutional neural networks is disclosed in the patent application of Publication No. CN107025440A
Road extracting method, disclosed technical solution realize structural output using full convolutional neural networks, can fully excavate remote sensing
The two dimensional geometry correlation of road in image.However, there are no can make full use of convolutional neural networks in the prior art
To extract the effective ways of the characteristic information of the building of different scale in remote sensing image.
It is, therefore, desirable to provide new technical solution come combine convolutional neural networks under different scale image feature carry out
Fusion, the precision that the automation to effectively improve different scale building is extracted.
Invention content
Nerve network system according to the present invention, the building for remote sensing image extract, including：
Input layer, the first to the 5th convolutional layer, first to fourth pond layer in VGG networks；
The input terminal of first single scale fused layer, the first single scale fused layer is connected to the output end of the first convolutional layer, uses
Single channel is merged in merging the first scale multi-channel feature figure for being exported of the first convolutional layer and exporting the first scale after fusion
Characteristic pattern；
Second to the 5th single scale fused layer, the input terminal of the second to the 5th single scale fused layer be respectively connected to second to
The output end of 5th convolutional layer, it is special for merging the second to the 5th scale multichannel that the second to the 5th convolutional layer is exported respectively
Sign is schemed and the second to the 5th scale after output fusion merges single channel characteristic pattern respectively；
The input terminal of first to fourth up-sampling layer, first to fourth up-sampling layer is respectively connected to second to the 5th single ruler
Spend the output end of fused layer；
The input terminal of multiple dimensioned splicing fused layer, multiple dimensioned splicing fused layer is connected to the first single scale fused layer, first
To the output end of the 4th up-sampling layer, the spy for merging the first single scale fused layer, first to fourth up-sampling layer is exported
Sign figure simultaneously exports the Multiscale Fusion single channel characteristic pattern after fusion；
Output layer, the input terminal of output layer is connected to the output end of multiple dimensioned splicing fused layer, for being based on multiple dimensioned melt
It closes single channel characteristic pattern and exports building feature figure,
Wherein, the first single scale fused layer, the output end of first to fourth up-sampling layer and multiple dimensioned splicing fused layer
What output end was exported is two-dimentional single channel characteristic pattern identical with the resolution ratio of remote sensing image.
Nerve network system according to the present invention further includes：
First to fourth cuts layer, and first to fourth cutting layer is separately positioned on first to fourth up-sampling layer and multiple dimensioned
Between splicing fused layer, for being cut to and being originally inputted image by the characteristic pattern that first to fourth up-sampling layer is exported respectively
Identical resolution ratio.
Nerve network system according to the present invention further includes following layers after the first to the 5th convolutional layer：
First to the 5th ReLU layers, the first to the 5th Normalization layers of Batch, first to the 5th Dropout layers,
For avoiding over-fitting, the generalization ability of nerve network system is improved.
Building extracting method according to the present invention for remote sensing image, including：
Build housebroken nerve network system as described above；
The building feature figure corresponding to remote sensing image is obtained using housebroken nerve network system.
Building extracting method according to the present invention for remote sensing image is building housebroken god as described above
Before the step of network system, further include：
Use the remote sensing training image comprising building and the data set pair of label image corresponding with remote sensing training image
Nerve network system is trained, to obtain housebroken nerve network system.
Building extracting method according to the present invention for remote sensing image is obtaining the building corresponding to remote sensing image
After characteristic pattern, final building distribution map is obtained using threshold method.
Building extracting method according to the present invention for remote sensing image, when being trained to nerve network system,
Using Sigmoid Cross Entropy Loss loss functions and use stochastic gradient descent algorithm.
Computer readable storage medium according to the present invention is stored with computer program on the storage medium, and program is located
The step of reason device realizes method as described above when executing.
Computing device according to the present invention, including memory, processor and storage are on a memory and can be on a processor
The step of computer program of operation, processor realizes method as described above when executing program.
Above-mentioned technical proposal according to the present invention directly utilizes the multi-scale information in deep layer convolutional neural networks, can
Densely distributed and various scale building is effectively treated, improves the precision of building automation extraction.In addition, according to the present invention
Above-mentioned technical proposal use whole picture image as input, directly output divide (that is, building extraction) result without into
Row has the slice of overlapping, and the efficiency of building extraction greatly improved.
Description of the drawings
It is incorporated into specification and the attached drawing of a part for constitution instruction shows the embodiment of the present invention, and with
Relevant verbal description principle for explaining the present invention together.In the drawings, similar reference numeral is for indicating class
As element.Drawings in the following description are some embodiments of the invention, rather than whole embodiments.It is common for this field
It, without creative efforts, can be obtain other attached drawings according to these attached drawings for technical staff.
Fig. 1 shows the schematic diagram of conventional image to be detected and present invention remote sensing image to be detected.
Fig. 2 schematically illustrates the schematic configuration diagram of nerve network system according to the present invention.
Fig. 3 schematically illustrates the exemplary flow of the building extracting method according to the present invention for remote sensing image
Figure.
Fig. 4 schematically illustrates the different images figure of each layer of output of nerve network system shown in Fig. 2.
Fig. 5 schematically illustrate a width original satellite remote sensing image, its corresponding true tag striograph and according to
The building feature figure of technical scheme of the present invention institute reality output.
Fig. 6 schematically illustrates accuracy rate-recall rate according to the technique and scheme of the present invention under different coefficient of relaxation
Curve.
Specific implementation mode
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with the embodiment of the present invention
In attached drawing, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described embodiment is
A part of the embodiment of the present invention, instead of all the embodiments.Based on the embodiments of the present invention, ordinary skill people
The every other embodiment that member is obtained without making creative work, shall fall within the protection scope of the present invention.It needs
It is noted that in the absence of conflict, the features in the embodiments and the embodiments of the present application mutually can be combined arbitrarily.
Fig. 1 shows the schematic diagram of conventional image to be detected and present invention remote sensing image to be detected.
Such as the description that background technology part combination Fig. 1 is done, since there are above-mentioned areas between remote sensing image and normal image
Not, it needs to propose new technical solution to combine convolutional neural networks to merge the image feature under different scale, to
Effectively improve the precision of the automation extraction of the building of different scale.
Fig. 2 schematically illustrates the schematic configuration diagram of nerve network system according to the present invention.
As shown in Fig. 2, nerve network system according to the present invention, the building for remote sensing image extracts, including：
Input layer (corresponding to " the input image " in Fig. 2), the first to the 5th convolutional layer in VGG networks is (right respectively
Layer collection where " Conv1_2, Conv2_2, Conv3_3, Conv4_3, Conv5_3 " that should be in Fig. 2), first to fourth pond
Change layer (corresponding respectively to " Pool1, Pool2, Pool3, Pool4 " layer in Fig. 2)；
First single scale fused layer (the 1st " Conv " corresponding to the left side of horizontal distribution in Fig. 2), the first single scale melts
The input terminal for closing layer is connected to the output end of the first convolutional layer, the first scale multichannel exported for merging the first convolutional layer
Characteristic pattern simultaneously exports the first scale fusion single channel characteristic pattern after fusion；
Second to the 5th single scale fused layer (the 2nd to the 5th " Conv " corresponding respectively to horizontal distribution in Fig. 2), the
The input terminal of two to the 5th single scale fused layers is respectively connected to the output end of the second to the 5th convolutional layer, for merging respectively
Second to the 5th after the second to the 5th scale multi-channel feature figure that two to the 5th convolutional layers are exported and respectively output fusion
Scale merges single channel characteristic pattern；
First to fourth up-sampling layer (corresponds respectively in Fig. 2 " 2 × Upsampling ", " 4 × Upsampling ", " 8
× Upsampling ", " 16 × Upsampling "), the input terminal of first to fourth up-sampling layer is respectively connected to second to the
The output end of five single scale fused layers；
Multiple dimensioned splicing fused layer (corresponding to " Concat " layer in Fig. 2), the input terminal of multiple dimensioned splicing fused layer connect
It is connected to the output end of the first single scale fused layer, first to fourth up-sampling layer, for merging the first single scale fused layer, first
To the 4th characteristic pattern that is exported of up-sampling layer and export the Multiscale Fusion single channel characteristic pattern after fusion；
Output layer (corresponds to " Conv " above " P " in Fig. 2), and the input terminal of output layer is connected to multiple dimensioned splicing and melts
The output end for closing layer, for (being corresponded in Fig. 2 based on Multiscale Fusion single channel characteristic pattern output building feature figure
" P "),
Wherein, the first single scale fused layer, the output end of first to fourth up-sampling layer and multiple dimensioned splicing fused layer
What output end was exported is two-dimentional single channel characteristic pattern identical with the resolution ratio of remote sensing image.
In the above-mentioned technical solutions, for the first to the 5th single scale fused layer, although the input corresponding to them
Port number (that is, quantity) C of characteristic pattern is different (as shown in Fig. 2, respectively 64,128,256,512,512), however, due to it
Merged respectively using the different convolution kernels that dimension is 1* (64,128,256,512,512) * 1*1 it is all under each self scale
Input feature vector figure, so can finally export 1 single channel characteristic pattern (that is, the first to the 5th scale merges single channel feature
Figure, they be resolution ratio be respectively 2562,1282,642,322,162 5 different resolutions single channel characteristic pattern).
It is similar with the fusion method of the above-mentioned first to the 5th single scale fused layer for multiple dimensioned splicing fused layer,
The port number C (that is, quantity) of its input feature vector figure (including the characteristic patterns of the first single scale fused layer output and using upper that are 5
The first to fourth 4 new feature figure being exported of up-sampling layer that the mode of sampling obtains, their resolution ratio with it is original distant
The resolution ratio for feeling image is identical).Therefore, this 5 characteristic patterns can be spliced into a 5 channel probability graphs (that is, characteristic pattern), and
Single channel prognostic chart (that is, Multiscale Fusion single channel characteristic pattern) is obtained using the convolution kernel that dimension is 1*5*1*1.
Although technical solution shown in Fig. 2 need not cut first to fourth up-sampling layer, However, alternatively,
Above-mentioned nerve network system can also include：
First to fourth cutting layer (each " Crop " that layer where corresponding respectively to " P2 " to " P5 " in Fig. 2 is concentrated),
First to fourth cutting layer is separately positioned between first to fourth up-sampling layer and multiple dimensioned splicing fused layer, for respectively will
The characteristic pattern that first to fourth up-sampling layer is exported is cut to resolution ratio identical with image is originally inputted.It is distant with automatic adaptation
The inconsistent situation of the resolution ratio of sense image resolution and first to fourth up-sampling layer output characteristic pattern.
Optionally, after the first to the 5th convolutional layer, above-mentioned nerve network system further include following layers (in fig. 2 not
It shows)：
First to the 5th ReLU layers, the first to the 5th Normalization layers of Batch, first to the 5th Dropout layers,
For avoiding over-fitting, the generalization ability of nerve network system is improved.
Network shallow-layer as shown in Figure 2 generates the characteristic pattern with fine spatial resolution but low level semantic information, such as
Deep layer shown in Fig. 2 generates the rough features figure with high-level semantics information, and the Feature Mapping of middle layer as shown in Figure 2 corresponds to
In certain intergrade features.Above-mentioned technical proposal can integrate these different characteristic patterns, and therefore, can efficiently extract has
Different appearances or the building blocked.
Fig. 3 schematically illustrates the exemplary flow of the building extracting method according to the present invention for remote sensing image
Figure.
Building extracting method according to the present invention for remote sensing image, including：
Step S302：Build housebroken nerve network system as described above；
Step S304：Remote sensing image is obtained (that is, " input in Fig. 2 using housebroken nerve network system
Image ") corresponding to building probability graph (that is, building feature figure as described above, i.e. building extract prognostic chart
“P”)。
Optionally, it is used for the building extracting method of remote sensing image, further includes before step S302：
Step S302 '：Using comprising building remote sensing training image (that is, " input image " in Fig. 2) and with it is distant
The data set of the corresponding label image (that is, " input map " in Fig. 2) of sense training image instructs nerve network system
Practice, to obtain housebroken nerve network system.
Optionally, it in step S304 and step S302 ', obtains above-mentioned building feature figure using threshold method and (finally builds
Build extraction result).
Optionally, in step S302 ', when being trained to nerve network system, Sigmoid Cross are used
Entropy Loss loss functions (that is, calculating function corresponding to " Loss " in Fig. 2) and stochastic gradient descent algorithm.
In order to make those skilled in the art more fully understand the advantageous effects of the present invention, below in conjunction with specific implementation
Example illustrates.
Fig. 4 schematically illustrates the different images figure of each layer of output of nerve network system shown in Fig. 2.
As shown in figure 4, Fig. 4 (a) is original the defending that a width (the first scale) is selected from Massachusetts remotely-sensed data collection
Star remote sensing image (that is, " input image " in Fig. 2), Fig. 4 (b) are that there is the second scale feature figure of small receptive field to pass through
Characteristic pattern (that is, " P2 " in Fig. 2) after interpolation, the edge and angle of original satellite remote sensing image can be extracted based on this feature figure
The low-level features such as point.Fig. 4 (c) be have the third scale feature figure of larger receptive field after interpolation characteristic pattern (that is, Fig. 2
In " P3 "), this feature figure can sketch out the preliminary profile of building.Fig. 4 (d) is the 4th scale for having bigger receptive field
Characteristic pattern (that is, " P4 " in Fig. 2) of the characteristic pattern after interpolation, can recognize that the non-building such as lake based on this feature figure
Region.Fig. 4 (e) is characteristic pattern of the 5th scale feature figure with maximum receptive field after interpolation (that is, in Fig. 2
" P5 "), the non-building areas domain such as lake and bare area can recognize that based on this feature figure.Finally, multi-level semantic information is integrated
A reliable prediction (Multiscale Fusion single channel characteristic pattern described above, that is, in Fig. 2 is obtained with spatial information
" P "), as shown in Fig. 4 (f).
Fig. 5 schematically illustrates original satellite remote sensing image, its corresponding true tag striograph and according to this hair
The building feature figure of bright technical solution institute reality output.
As shown in figure 5, Fig. 5 (a) is the width original satellite remote sensing image selected from Massachusetts remotely-sensed data collection,
Fig. 5 (b) is its true tag striograph, and Fig. 5 (c) is prediction label figure (that is, institute's reality output according to the technique and scheme of the present invention
Building feature figure).In terms of improvement of visual effect, building distribution situation can be predicted well according to the technique and scheme of the present invention,
And building boundary accurate.
Fig. 6 schematically illustrates accuracy rate-recall rate according to the technique and scheme of the present invention under different coefficient of relaxation
Curve.
Accuracy rate is defined as ratio of the pixel that detected within the adjacent ρ pixel coverage of real pixel, it will
Recall rate is defined as ratio of the real pixel within the adjacent ρ pixel coverage for the pixel that detected.According to Fig. 6 (a)
Accuracy rate-recall rate curve of the technical scheme of the present invention in ρ=3, corresponding model accuracy is about 0.9668 (corresponding diagram
6 (a) middle symbol × expression, accuracy rate and the equal point breakeven of recall rate).Fig. 6 (b) is the skill according to the present invention
Accuracy rate-recall rate curve of the art scheme in ρ=0, corresponding model accuracy are about 0.8424 (the middle symbol of corresponding diagram 6 (b)
Number × indicate, accuracy rate and the equal point breakeven of recall rate).
Table 1 gives including Mnih.V in its doctoral thesis " Machine learning for aerial image
Mnih-CNN schemes and Mnih-CNN+CRF schemes and Saito disclosed in labeling, Doctoral (2013) " exist
“Multiple object extraction from aerial imagery with convolutional neural
The skill of Saito-multi-MA schemes and Saito-multi-MA&CIS schemes and the present invention disclosed in networks "
Building extraction performance comparison between different schemes including art scheme.
Performance comparison between 1 different technologies scheme of table
| Model | Breakeven (ρ=3) | Breakeven (ρ=0) | Predicted time (s) |
| Mnih-CNN | 0.9271 | 0.7661 | 8.7 |
| Mnih-CNN+CRF | 0.9282 | 0.7638 | 26.6 |
| Saito-multi-MA | 0.9503 | 0.7873 | 67.72 |
| Saito-multi-MA&CIS | 0.9509 | 0.7872 | 67.84 |
| Technical scheme of the present invention | 0.9668 | 0.8424 | 2.05 |
Note：Predicted time is average time, the video card model used needed for the prediction of single width 1500*1500 size testing images
For NVIDIA TITAN X.
From the results shown in Table 1, in terms of the model accuracy either under different coefficient of relaxation (ρ=3 and ρ=0),
Or in terms of predicted time, above-mentioned technical proposal according to the present invention all has superior technique effect.It can not only be significant
Extraction accuracy is improved, operation time can also be reduced.
In conjunction with above-mentioned technical proposal according to the present invention, it is also proposed that a kind of computer readable storage medium, storage medium
On the step of being stored with computer program, method as shown in Figure 3 is realized when program is executed by processor.
In conjunction with above-mentioned technical proposal according to the present invention, it is also proposed that a kind of computing device, including memory, processor and
The computer program that can be run on a memory and on a processor is stored, processor realizes side as shown in Figure 3 when executing program
The step of method.
Above-mentioned technical proposal according to the present invention extracts each resolution ratio in network using VGG networks as basic structure
Last layer in characteristic pattern is fused to single pass characteristic pattern using convolution operation.Spliced by resampling and characteristic pattern, is obtained
To final prediction result.
Above-mentioned technical proposal according to the present invention directly utilizes the multi-scale information in deep layer convolutional neural networks, can
Densely distributed and various scale building is effectively treated, improves the precision of building automation extraction.In addition, according to the present invention
Above-mentioned technical proposal use whole picture image as input, directly output divide (that is, building extraction) result without into
Row has the slice of overlapping, and the efficiency of building extraction greatly improved.
Above-mentioned technical proposal according to the present invention, has further the advantage that：1) each resolution characteristics figure can be merged, from
And input image multi-scale information is extracted, realize the accurate extraction of building；2) due to need not be at prediction (that is, extraction)
Row model integrated does not need post-processing operation yet, thus building extraction efficiency greatly improved；3) due to the use of full convolution net
Network, thus the input image of arbitrary dimension can be received in the case where video memory allows.
In addition, above-mentioned technical proposal according to the present invention, directly uses whole picture image as input, before primary network
It is obtained with segmentation (that is, building extraction) to propagation as a result, without carrying out mould by way of thering is overlapping to be sliced
Type is integrated, need not also carry out post-processing operation, building extraction efficiency greatly improved.By being based on Massachusetts
Remotely-sensed data collection carry out above-mentioned contrast test result it can be shown that above-mentioned technical proposal according to the present invention either in precision
Or all it is significantly better than other methods in efficiency.
Descriptions above can combine implementation individually or in various ways, and these variants all exist
Within protection scope of the present invention.
It will appreciated by the skilled person that whole or certain steps in method disclosed hereinabove, system, dress
Function module/unit in setting may be implemented as software, firmware, hardware and its combination appropriate.In hardware embodiment,
Division between the function module/unit referred in the above description not necessarily corresponds to the division of physical assemblies；For example, one
Physical assemblies can have multiple functions or a function or step that can be executed by several physical assemblies cooperations.Certain groups
Part or all components may be implemented as by processor, such as the software that digital signal processor or microprocessor execute, or by
It is embodied as hardware, or is implemented as integrated circuit, such as application-specific integrated circuit.Such software can be distributed in computer-readable
On medium, computer-readable medium may include computer storage media (or non-transitory medium) and communication media (or temporarily
Property medium).As known to a person of ordinary skill in the art, term computer storage medium is included in for storing information (such as
Computer-readable instruction, data structure, program module or other data) any method or technique in the volatibility implemented and non-
Volatibility, removable and nonremovable medium.Computer storage media include but not limited to RAM, ROM, EEPROM, flash memory or its
His memory technology, CD-ROM, digital versatile disc (DVD) or other optical disc storages, magnetic holder, tape, disk storage or other
Magnetic memory apparatus or any other medium that can be used for storing desired information and can be accessed by a computer.This
Outside, known to a person of ordinary skill in the art to be, communication media generally comprises computer-readable instruction, data structure, program mould
Other data in the modulated data signal of block or such as carrier wave or other transmission mechanisms etc, and may include any information
Delivery media.
Finally it should be noted that：The above embodiments are merely illustrative of the technical solutions of the present invention, rather than its limitations.Although
Present invention has been described in detail with reference to the aforementioned embodiments, it will be understood by those of ordinary skill in the art that：It still may be used
With technical scheme described in the above embodiments is modified or equivalent replacement of some of the technical features；
And these modifications or replacements, the spirit of the technical solution for various embodiments of the present invention that it does not separate the essence of the corresponding technical solution
And range.
Claims (9)
1. a kind of nerve network system, which is characterized in that the building for remote sensing image extracts, including：
Input layer, the first to the 5th convolutional layer, first to fourth pond layer in VGG networks；
First single scale fused layer, the input terminal of the first single scale fused layer are connected to the output of first convolutional layer
End, for merging the first scale multi-channel feature figure that first convolutional layer is exported and the first scale exported after fusion melts
Close single channel characteristic pattern；
The input terminal of second to the 5th single scale fused layer, the described second to the 5th single scale fused layer is respectively connected to described
The output end of two to the 5th convolutional layers, the second to the 5th scale exported for merging the described second to the 5th convolutional layer respectively
The second to the 5th scale after multi-channel feature figure and respectively output fusion merges single channel characteristic pattern；
The input terminal of first to fourth up-sampling layer, the first to fourth up-sampling layer is respectively connected to described second to the 5th
The output end of single scale fused layer；
Multiple dimensioned splicing fused layer, it is described it is multiple dimensioned splicing fused layer input terminal be connected to the first single scale fused layer,
The output end of the first to fourth up-sampling layer, for merging the first single scale fused layer, described first to fourth
Characteristic pattern that sample level is exported simultaneously exports the Multiscale Fusion single channel characteristic pattern after fusion；
Output layer, the input terminal of the output layer is connected to the output end of the multiple dimensioned splicing fused layer, for based on described
Multiscale Fusion single channel characteristic pattern exports building feature figure,
Wherein, the first single scale fused layer, the output end of the first to fourth up-sampling layer and the multiple dimensioned splicing
What the output end of fused layer was exported is two-dimentional single channel characteristic pattern identical with the resolution ratio of the remote sensing image.
2. nerve network system as described in claim 1, which is characterized in that further include：
First to fourth cuts layer, and the first to fourth cutting layer is separately positioned on the first to fourth up-sampling layer and institute
Between stating multiple dimensioned splicing fused layer, for respectively by described first to fourth up-sample the characteristic pattern that is exported of layer be cut to
It is originally inputted the identical resolution ratio of image.
3. nerve network system as claimed in claim 1 or 2, which is characterized in that after the described first to the 5th convolutional layer
Further include following layers：
First to the 5th ReLU layers, the first to the 5th Normalization layers of Batch, first to the 5th Dropout layers, be used for
Over-fitting is avoided, the generalization ability of the nerve network system is improved.
4. a kind of building extracting method for remote sensing image, which is characterized in that including：
Build the housebroken nerve network system as described in any one of claim 1-3；
The building feature figure corresponding to remote sensing image is obtained using housebroken nerve network system.
5. being used for the building extracting method of remote sensing image as claimed in claim 4, which is characterized in that in the structure through instruction
Before the step of experienced nerve network system as described in any one of claim 1-3, further include：
Use the remote sensing training image comprising building and the data set pair of label image corresponding with remote sensing training image
The nerve network system is trained, to obtain the housebroken nerve network system.
6. being used for the building extracting method of remote sensing image as described in claim 4 or 5, which is characterized in that described in acquisition
After the building feature figure corresponding to remote sensing image, final building distribution map is obtained using threshold method.
7. being used for the building extracting method of remote sensing image as claimed in claim 5, which is characterized in that the nerve net
When network system is trained, calculated using Sigmoid Cross Entropy Loss loss functions and using stochastic gradient descent
Method.
8. a kind of computer readable storage medium, which is characterized in that be stored with computer program, the journey on the storage medium
The step of any one of claim 4 to 7 the method is realized when sequence is executed by processor.
9. a kind of computing device, which is characterized in that including memory, processor and be stored on the memory and can be described
The computer program run on processor, the processor realize any one of claim 4 to 7 institute when executing described program
The step of stating method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810373725.3A CN108764039B (en) | 2018-04-24 | 2018-04-24 | Neural network, building extraction method of remote sensing image, medium and computing equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810373725.3A CN108764039B (en) | 2018-04-24 | 2018-04-24 | Neural network, building extraction method of remote sensing image, medium and computing equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108764039A true CN108764039A (en) | 2018-11-06 |
| CN108764039B CN108764039B (en) | 2020-12-01 |
Family
ID=64011327
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810373725.3A Active CN108764039B (en) | 2018-04-24 | 2018-04-24 | Neural network, building extraction method of remote sensing image, medium and computing equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108764039B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109753928A (en) * | 2019-01-03 | 2019-05-14 | 北京百度网讯科技有限公司 | The recognition methods of architecture against regulations object and device |
| CN109859167A (en) * | 2018-12-28 | 2019-06-07 | 中国农业大学 | The appraisal procedure and device of cucumber downy mildew severity |
| CN109871798A (en) * | 2019-02-01 | 2019-06-11 | 浙江大学 | A kind of remote sensing image building extracting method based on convolutional neural networks |
| CN109934110A (en) * | 2019-02-02 | 2019-06-25 | 广州中科云图智能科技有限公司 | A kind of river squatter building house recognition methods nearby |
| CN110163207A (en) * | 2019-05-20 | 2019-08-23 | 福建船政交通职业学院 | One kind is based on Mask-RCNN ship target localization method and storage equipment |
| CN110263797A (en) * | 2019-06-21 | 2019-09-20 | 北京字节跳动网络技术有限公司 | Crucial the point estimation method, device, equipment and the readable storage medium storing program for executing of skeleton |
| CN110991252A (en) * | 2019-11-07 | 2020-04-10 | 郑州大学 | Detection method for crowd distribution and counting in unbalanced scene |
| CN113486840A (en) * | 2021-07-21 | 2021-10-08 | 武昌理工学院 | Building rapid extraction method based on composite network correction |
| CN116052019A (en) * | 2023-03-31 | 2023-05-02 | 深圳市规划和自然资源数据管理中心(深圳市空间地理信息中心) | High-quality detection method suitable for built-up area of large-area high-resolution satellite image |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106056628A (en) * | 2016-05-30 | 2016-10-26 | 中国科学院计算技术研究所 | Target tracking method and system based on deep convolution nerve network feature fusion |
| US20160342888A1 (en) * | 2015-05-20 | 2016-11-24 | Nec Laboratories America, Inc. | Memory efficiency for convolutional neural networks operating on graphics processing units |
| CN106886977A (en) * | 2017-02-08 | 2017-06-23 | 徐州工程学院 | A kind of many figure autoregistrations and anastomosing and splicing method |
| CN107092871A (en) * | 2017-04-06 | 2017-08-25 | 重庆市地理信息中心 | Remote sensing image building detection method based on multiple dimensioned multiple features fusion |
| CN107092870A (en) * | 2017-04-05 | 2017-08-25 | 武汉大学 | A kind of high resolution image semantics information extracting method and system |
| CN107123083A (en) * | 2017-05-02 | 2017-09-01 | 中国科学技术大学 | Face edit methods |
| CN107169974A (en) * | 2017-05-26 | 2017-09-15 | 中国科学技术大学 | It is a kind of based on the image partition method for supervising full convolutional neural networks more |
| CN107220657A (en) * | 2017-05-10 | 2017-09-29 | 中国地质大学（武汉） | A kind of method of high-resolution remote sensing image scene classification towards small data set |
| US20170308753A1 (en) * | 2016-04-26 | 2017-10-26 | Disney Enterprises, Inc. | Systems and Methods for Identifying Activities and/or Events in Media Contents Based on Object Data and Scene Data |
-
2018
- 2018-04-24 CN CN201810373725.3A patent/CN108764039B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160342888A1 (en) * | 2015-05-20 | 2016-11-24 | Nec Laboratories America, Inc. | Memory efficiency for convolutional neural networks operating on graphics processing units |
| US20170308753A1 (en) * | 2016-04-26 | 2017-10-26 | Disney Enterprises, Inc. | Systems and Methods for Identifying Activities and/or Events in Media Contents Based on Object Data and Scene Data |
| CN106056628A (en) * | 2016-05-30 | 2016-10-26 | 中国科学院计算技术研究所 | Target tracking method and system based on deep convolution nerve network feature fusion |
| CN106886977A (en) * | 2017-02-08 | 2017-06-23 | 徐州工程学院 | A kind of many figure autoregistrations and anastomosing and splicing method |
| CN107092870A (en) * | 2017-04-05 | 2017-08-25 | 武汉大学 | A kind of high resolution image semantics information extracting method and system |
| CN107092871A (en) * | 2017-04-06 | 2017-08-25 | 重庆市地理信息中心 | Remote sensing image building detection method based on multiple dimensioned multiple features fusion |
| CN107123083A (en) * | 2017-05-02 | 2017-09-01 | 中国科学技术大学 | Face edit methods |
| CN107220657A (en) * | 2017-05-10 | 2017-09-29 | 中国地质大学（武汉） | A kind of method of high-resolution remote sensing image scene classification towards small data set |
| CN107169974A (en) * | 2017-05-26 | 2017-09-15 | 中国科学技术大学 | It is a kind of based on the image partition method for supervising full convolutional neural networks more |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109859167A (en) * | 2018-12-28 | 2019-06-07 | 中国农业大学 | The appraisal procedure and device of cucumber downy mildew severity |
| CN109753928A (en) * | 2019-01-03 | 2019-05-14 | 北京百度网讯科技有限公司 | The recognition methods of architecture against regulations object and device |
| CN109871798A (en) * | 2019-02-01 | 2019-06-11 | 浙江大学 | A kind of remote sensing image building extracting method based on convolutional neural networks |
| CN109934110A (en) * | 2019-02-02 | 2019-06-25 | 广州中科云图智能科技有限公司 | A kind of river squatter building house recognition methods nearby |
| CN109934110B (en) * | 2019-02-02 | 2021-01-12 | 广州中科云图智能科技有限公司 | Method for identifying illegal buildings near river channel |
| CN110163207A (en) * | 2019-05-20 | 2019-08-23 | 福建船政交通职业学院 | One kind is based on Mask-RCNN ship target localization method and storage equipment |
| CN110163207B (en) * | 2019-05-20 | 2022-03-11 | 福建船政交通职业学院 | Ship target positioning method based on Mask-RCNN and storage device |
| CN110263797A (en) * | 2019-06-21 | 2019-09-20 | 北京字节跳动网络技术有限公司 | Crucial the point estimation method, device, equipment and the readable storage medium storing program for executing of skeleton |
| CN110991252A (en) * | 2019-11-07 | 2020-04-10 | 郑州大学 | Detection method for crowd distribution and counting in unbalanced scene |
| CN110991252B (en) * | 2019-11-07 | 2023-07-21 | 郑州大学 | Detection method for people group distribution and counting in unbalanced scene |
| CN113486840A (en) * | 2021-07-21 | 2021-10-08 | 武昌理工学院 | Building rapid extraction method based on composite network correction |
| CN116052019A (en) * | 2023-03-31 | 2023-05-02 | 深圳市规划和自然资源数据管理中心(深圳市空间地理信息中心) | High-quality detection method suitable for built-up area of large-area high-resolution satellite image |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108764039B (en) | 2020-12-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108764039A (en) | Building extracting method, medium and the computing device of neural network, remote sensing image | |
| Berman et al. | Single image dehazing using haze-lines | |
| CN105574513B (en) | Character detecting method and device | |
| US20180253865A1 (en) | Image matting using deep learning | |
| CN111161311A (en) | Visual multi-target tracking method and device based on deep learning | |
| Roeder et al. | A computational image analysis glossary for biologists | |
| CN112348828B (en) | Instance segmentation method and device based on neural network and storage medium | |
| CN110675407B (en) | Image instance segmentation method and device, electronic equipment and storage medium | |
| CN112651978A (en) | Sublingual microcirculation image segmentation method and device, electronic equipment and storage medium | |
| CN110298844A (en) | X-ray contrastographic picture blood vessel segmentation and recognition methods and device | |
| CN111553362B (en) | Video processing method, electronic device and computer readable storage medium | |
| CN112598045A (en) | Method for training neural network, image recognition method and image recognition device | |
| CN109389096B (en) | Detection method and device | |
| CN109919915A (en) | Retina fundus image abnormal region detection method and device based on deep learning | |
| CN110517273B (en) | Cytology image segmentation method based on dynamic gradient threshold | |
| CN112330684A (en) | Object segmentation method and device, computer equipment and storage medium | |
| CN107547803A (en) | Video segmentation result edge optimization processing method, device and computing device | |
| Bergamasco et al. | A dual-branch deep learning architecture for multisensor and multitemporal remote sensing semantic segmentation | |
| CN111079807A (en) | Ground object classification method and device | |
| Mostafa et al. | Corresponding regions for shadow restoration in satellite high-resolution images | |
| CN111860208A (en) | Remote sensing image ground object classification method, system, device and medium based on super pixels | |
| Al Shehhi et al. | An Automatic Cognitive Graph‐Based Segmentation for Detection of Blood Vessels in Retinal Images | |
| CN112949458B (en) | Training method of target tracking segmentation model, target tracking segmentation method and device | |
| Yao et al. | End-to-end adaptive object detection with learnable Retinex for low-light city environment | |
| CN111311601A (en) | Segmentation method and device for spliced image |
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 |