CN110415280A - Remote sensing image and building vector method for registering and system under multitask CNN model - Google Patents
Remote sensing image and building vector method for registering and system under multitask CNN model Download PDFInfo
- Publication number
- CN110415280A CN110415280A CN201910371472.0A CN201910371472A CN110415280A CN 110415280 A CN110415280 A CN 110415280A CN 201910371472 A CN201910371472 A CN 201910371472A CN 110415280 A CN110415280 A CN 110415280A
- Authority
- CN
- China
- Prior art keywords
- remote sensing
- vector
- sensing image
- building
- building vector
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
- G06T7/33—Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods
-
- 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
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
-
- 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
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of remote sensing images based on multitask CNN model and building vector method for registering and system, it first include by registration building vector as reference sample, to train full convolutional network model, by the model, the feature figure layer for being adapted to building recognition is further generated from high score remote sensing image;Its secondary design multitask CNN model, the building vector after feature figure layer and rasterizing is overlapped, is input in model, via convolution several times, pond and full attended operation, exports the misinformation probability and geometric correction parameter of current building vector;Building vector finally based on correction front and back calculates the reference value of multitask CNN model output result, completes model training;By the resulting multitask CNN model of training, the auto-registration of building vector and remote sensing image is realized.The present invention improves the effective precision of initial data on the basis of preserved building object vector effective information in the automatic screening and timing for carrying out building vector.
Description
Technical field
The present invention relates to Surveying Science and Technology fields, utilize multitask CNN model more specifically to a kind of, into
The method and system of row remote sensing image and building vector auto-registration.
Background technique
Accurate building vector outline information is obtained from high-resolution remote sensing image, can be urban planning, soil tune
It looks into, architecture against regulations detection and many application fields such as military surveillance provide important evidences.Due to the building in historical summary
Vector data was usually subjected to manually check, and had reliable vector structure and edge detail information, therefore compared to directly progress
Remote sensing image building extracts, and is a kind of more economical and reliable by existing building vector data and remote sensing image matching
The higher information acquiring pattern of property.But due to the removal of building, damage and asynchronous mapping data resolution, imaging
The accuracy registration of the reasons such as angle or positioning accuracy are inconsistent, Yao Shixian remote sensing image and heterologous building vector data is not only
Wrong report vector need to be deleted, also needs to solve offset and deformation of the vector relative to image degree unevenness.High-precision remote sensing image with
Building vector autoregistration technology has important meaning for the precision, quality and the application value that promote history vector data
Justice.
Existing remote sensing image and building vector registration technique can be divided mainly into rule-based refined processing method and
Vector optimization method based on active contour model.The former is general to extract image line feature first, and designs dependency rule to line
Feature is screened, is organized into groups, and then original vector is edited and replaced automatically, but be difficult in image weak feature,
Pseudo-characteristic and excessively stringent supposed premise all limit the usage range of this technology;The latter then passes through optimization comprising more
The energy function parameter of kind constraint, drives original vector to be restrained while keeping continuously smooth to building edge, but this kind of
Method does not take the structure and deformation behavior of existing building vector data into account, directly may cause using such technology to original
The over-correction of beginning building vector.In addition, establishing the low-level image spy in application engineer on the two classes technological essence
On the basis of sign, it is difficult to adapt to the complicated and diversified fabric structure in different regions, illumination in texture and different data, point
The variation of resolution and image quality, therefore its application scenarios and generalization ability are largely limited.
Summary of the invention
The technical problem to be solved in the present invention is that the larger journey of application scenarios and generalization ability for the prior art
Limited defect is spent, remote sensing image and building vector method for registering and system under a kind of multitask CNN model are provided.
The technical solution adopted by the present invention to solve the technical problems is: constructing the remote sensing under a kind of multitask CNN model
Image and building vector method for registering, specifically includes the following steps:
S1, data preparation is carried out, constructs training dataset;It includes several high score remote sensing shadows that the training data, which is concentrated,
Picture;Wherein, for every high score remote sensing image, the training dataset further includes not original with high score remote sensing image matching
Building vector, and original building vector is registrated building vector with after high score remote sensing image progress geometric precision correction;
S2, each registration building vector is got using from training data concentration, training is used for remote sensing image building
The full convolutional network model of detection;
S3, each the original building vector concentrated to training data traverse, and during traversal, use
Back-propagating and stochastic gradient descent algorithm, and the feature figure layer to be generated using full convolutional network is as auxiliary information, to institute
Multitask CNN model is stated to be trained;
In the multitask CNN model that S4, input remote sensing image data to step S3 are trained, via several times in the model
After convolution, pond and full attended operation, the autoregistration of remote sensing image and building vector is carried out.
Further, the input item of the multitask CNN model includes: that will be formed after original building vector to raster conversion
Bianry image, and by with the bianry image have identical image range high score remote sensing image, be input to step S2 instruction
After practicing resulting full convolutional network model, what is obtained has the feature figure layer of identical geographic range with high score remote sensing image, respectively
Input item as multitask CNN model;
The output item of the multitask CNN model include: building vector misinformation probability and geometric correction parameter two
Branch;
The penalty values Loss of the multitask CNN model, specific mathematical formulae are calculated using loss function are as follows:
Wherein, i indicates sample serial number, and N indicates total sample number in training batch, piIndicate current sample misinformation probability
Predicted value,Indicate the reference value of current sample misinformation probability;miIndicate that the geometric correction for original building vector is joined
Number predicted value;Indicate it is non-wrong report situation under original building vector to be registrated building vector geometric correction reference value,
Specifically, in step S3, it is when traversing the original building vector that training data is concentrated, original building vector sum is corresponding
Registration building vector compareed obtained by;Indicate the intersection entropy loss of current sample misinformation probability,The penalty values for indicating geometric correction parameter are calculated using mean square error loss function.
Further, to loss function application gradient descent algorithm, when penalty values Loss levels off to X, the multitask
CNN model training finishes, and is applied in subsequent execution step;Wherein, X >=0.
Further, remote sensing image and original building vector are matched automatically by the multitask CNN model
Quasi- process, as to each the corrected process of building vector to be registered being input in multitask CNN model,
When the misinformation probability of multitask CNN model output is greater than preset first threshold, then building vector currently entered is deleted
It removes, the geometric correction parameter otherwise exported according to multitask CNN model is corrected the building vector of input, can
It is registrated with corresponding high score remote sensing image.
Further, for when carrying out multitask CNN model training, for being input in multitask CNN model, and
The building vector being deleted in correction course, the reference value of corresponding model output item are set as: misinformation probability 1, several
What correction parameter is sky;For being input in multitask CNN model, and the building vector being retained in correction course,
The reference value of corresponding model output item is set as: misinformation probability 0, and geometric correction parameter is by taking correction front and back building arrow
The identical point coordinates of amount carry out least-squares estimation and acquire.
Further, to avoid the raw information for causing to destroy building vector to building vector exaggerated correction,
In correction course, calculating corrects the friendship between the building vector of front and back and than index, if described hand over and be less than in advance than index
When fixed second threshold, then correction result is not adopted.
Remote sensing image and building vector registration arrangement under a kind of multitask CNN model proposed by the present invention, it is specific to wrap
It includes with lower module:
Data construct module, for carrying out data preparation, construct training dataset;Wherein, the training dataset wraps
Include several high score remote sensing images;For every high score remote sensing image, the training dataset further include not with high score remote sensing shadow
As registration original building vector, and by original building vector and high score remote sensing image progress geometric precision correction after matching
Quasi- building vector;
Full convolutional network training pattern, for getting each registration building vector, instruction using from training data concentration
Practice the full convolutional network model for the detection of high score remote sensing image building;
Multitask CNN model training module, each original building vector progress time for being concentrated to training data
It goes through, during traversal, multitask CNN model is trained using back-propagating and stochastic gradient descent algorithm;Its
In, the bianry image that will be formed after original building vector to raster conversion, and will have identical image model with the bianry image
The high score remote sensing image enclosed, after being input to the resulting full convolutional network model of training, what is obtained has phase with high score remote sensing image
With the feature figure layer of geographic range, respectively as the input item of multitask CNN model;
Remote sensing image matching module should for inputting remote sensing image data into the resulting multitask CNN model of training
In model via convolution several times, pond and full attended operation after, the autoregistration of progress remote sensing image and building vector.
Further, remote sensing image proposed by the present invention and building vector registration arrangement, it is distant using such as any of the above-described
Image and building vector method for registering are felt, to the carry out autoregistration of remote sensing image and building vector.
Remote sensing image under a kind of multitask CNN model of the present invention and building vector method for registering and it is
In system, adopted using the feature figure layer that full convolutional network generates as auxiliary information by constructing multitask CNN model learning frame
Manually the building vector of correction front and back is trained multitask CNN model as learning sample, after completing training
Model is input, the misinformation probability and geometric correction parameter of output vector, completion pair with feature figure layer and original building vector
The screening and correction of original vector, to achieve the purpose that remote sensing image and building vector auto-registration.
Implement the remote sensing image and building vector method for registering and system under a kind of multitask CNN model of the invention,
It has the advantages that
1, high dimensional feature is adaptively generated from remote sensing image using CNN model for estimating the several of building vector
What correction parameter has stronger data adaptability and general compared to the conventional method based on engineer's rule and feature
Change ability;
2, the present invention can be by being pointedly arranged geometric correction model to the original building vector of different distortion degrees
Registration process is carried out, is beneficial on the basis of retaining original vector resulting structure information, promotes its precision level and using valence
Value.
Detailed description of the invention
Present invention will be further explained below with reference to the attached drawings and examples, in attached drawing:
Fig. 1 is the method flow diagram that high score remote sensing image building detects in the embodiment of the present invention 1;
Fig. 2 is the system construction drawing that high score remote sensing image building detects in the embodiment of the present invention 2;
Fig. 3 is that the full convolutional network structure in the embodiment of the present invention 1 or 2 for the detection of high score remote sensing image building is shown
It is intended to;
Fig. 4 is the multitask CNN model knot being registrated for remote sensing image with building vector in the embodiment of the present invention 1 or 2
Structure schematic diagram.
Specific embodiment
For a clearer understanding of the technical characteristics, objects and effects of the present invention, now control attached drawing is described in detail
A specific embodiment of the invention.
Embodiment 1:
Referring to FIG. 1, its method flow diagram for high score remote sensing image building detection in the embodiment of the present invention 1, this hair
Bright disclosed one kind is used for the step of high score remote sensing image building detects and includes:
S1, data preparation is carried out, constructs training dataset;The training dataset includes several high score remote sensing shadows
Picture;Wherein, for every high score remote sensing image, the training dataset further includes not original with high score remote sensing image matching
Building vector, and original building vector is registrated building vector with after high score remote sensing image progress geometric precision correction;
Specifically, in the present embodiment, carrying out artificial geometric precision correction to original building vector includes to the whole of vector outline
Body position is moved, and is increased vector nodes, deleted and moving operation, the registration building vector after making registration
With the accurate fitting of contour of building in remote sensing image;Wherein, training dataset accounts for the ratio of all data according to number to be processed
According to total amount depending on, the generally 10%-30% of total amount of data;Being registrated building vector can be with original building vector and height
Remote sensing image is divided to be deposited into sample database together, the data in sample database can be used as training data, and be used for new registration task.
S2, each registration building vector is got using from training data concentration, training is built for high score remote sensing image
Build the full convolutional network model of analyte detection;Specifically, each registration building Vector Grid is turned to bianry image, and with institute
Bianry image be used as and refer to true value, the full convolutional network model of training, specific training method is using back-propagating and random
Gradient descent algorithm carries out, and the estimated amount of damage of the full convolution model is further realized by establishing intersection entropy function.
In the present embodiment, the full convolutional network model for the detection of high score remote sensing image building of use has symmetrical
Double pyramid structures of formula, in the model, input image is first passed through to be tied comprising the classical CNN of several convolutional layers and pond layer
Structure obtains the lower characteristic pattern of resolution ratio;Then by a series of up-samplings operate, generation with it is right step by step in aforementioned CNN structure
The characteristic pattern answered finally obtains End features figure layer by once up-sampling operation again, and then exports segmentation result.To low
Dimensionality reduction can be carried out to characteristic pattern using 1 × 1 convolution kernel while resolution characteristics figure is up-sampled, and it is next in progress
Before secondary up-sampling, these dimensionality reduction features will corresponding element be added by way of with CNN structure in corresponding level feature
Figure is merged.The training of the model mainly uses back-propagating and small batch stochastic gradient descent algorithm to optimize following intersect
Entropy loss function is realized:
Wherein, yiFor network output i-th of sample (pixel) prediction probability value,(positive sample is for corresponding its true value
1, negative sample 0);M is training sample total number in batch.
Referring to FIG. 3, Fig. 3 is full convolution model structural schematic diagram provided in an embodiment of the present invention, it is complete to roll up as shown in Fig. 3
Product module type includes the convolution sum pondization operation in 3 stages, and symmetrical 3 up-samplings operation therewith, the feature of the left and right sides
Figure is connected by side and carries out Fusion Features, and the feature of fusion is finally carried out primary up-sampling operation and obtains final spy
Figure layer is levied, this feature figure layer is for generating segmentation result and establishing cross entropy loss function with reference to true value.
S3, original building vector is taken from training data concentration, and its grid is turned into bianry image, the binary map
The image range of picture is determined by the boundary rectangle of the corresponding registration building vector of original building vector sum;Specific real
During applying, to the original building vector of Mr. Yu's building, corresponding registration is found from training data concentration first and is built
Object vector is built, the boundary rectangle of above-mentioned two vector is then taken, and is subject to lesser buffer area (10-30 pixel) and determines grid
Change image size.If the original building vector is manually determined as reporting vector deletion by mistake, only with the external square of its own
Shape, and rasterized images size is determined using identical buffer area.
Each the original building vector taken out is concentrated to traverse to from training data, wherein in the process of traversal
In, multitask CNN model is trained using back-propagating and stochastic gradient descent algorithm;
Specifically, the input of the multitask CNN model includes above-mentioned original building vector in current training process
The bianry image generated after rasterizing, and bianry image is input to the resulting feature figure layer of full convolutional network model;The mould
The output of type includes Liang Ge branch, the i.e. misinformation probability of vector and geometric correction parameter;The loss of the model passes through above-mentioned two
A output valve, and by original building vector be registrated after building vector compareed, the misinformation probability true value of generation with
And geometric correction parameter reference values are compared and are calculated.
Specifically, referring to FIG. 4, it is registrated for remote sensing image with building vector to be provided in an embodiment of the present invention
Multitask CNN model structure schematic diagram;As shown, the present example, still by taking affine transformation as an example, input data (is wrapped
Include the original building vector after feature figure layer and rasterizing) via convolution sum pondization several times handle after, by one
Full articulamentum generates multi-C vector, and for the multi-C vector again after two full attended operations in parallel, synchronous generation wrong report is general
Rate (being calculated by Sigmoid activation primitive) and affine parameter correspond to two output branchs of multitask CNN model.In conjunction with accidentally
The reference value for reporting probability and geometric correction parameter, can construct as the following formula loss function:
Wherein, i indicates sample serial number, and N indicates total sample number in training batch, piIndicate current sample misinformation probability
Predicted value,Indicate that (wrong report vector is 1, otherwise for 0) for the reference value of current sample misinformation probability;miIt indicates to build for original
Build the geometric correction parameter prediction value of object vector;Indicate that original building vector is sweared to registration building under non-wrong report situation
The geometric correction reference value of amount;Indicate the intersection entropy loss of current sample misinformation probability,
The penalty values for indicating geometric correction parameter are calculated using mean square error loss function.Wherein, the Optimization Solution of loss function can
It is realized by backpropagation and small batch stochastic gradient descent algorithm.
In S4, the present embodiment, in order to further determine the registration precision of trained multitask CNN model, counted
Test data set is constructed according to while preparation;The test data set includes several high score remote sensing images;Wherein, for every
Zhang Gaofen remote sensing image, the test data set further include the not test architecture object vector with high score remote sensing image matching;
In input remote sensing image into multitask CNN model, the autoregistration of remote sensing image and building vector is carried out
Before, concentrate each test architecture object vector to traverse test data;Wherein, a test architecture object vector is often traversed
When, the test architecture object Vector Grid is turned to bianry image by setting buffer area, while being generated using full convolutional network model
The characteristics of remote sensing image figure layer of corresponding position will be input in multitask CNN model after the two superposition, export misinformation probability
With the predicted value of geometric correction parameter;It repeats to traverse, until test data concentrates all test architecture object vectors to handle
Finish, immediately according to the predicted value of the misinformation probability of output and geometric correction parameter, to the network parameter of multitask CNN model into
Row adjustment, enables remote sensing image accurately to realize the Auto-matching with building vector.
Specifically, in implementation process, rasterizing buffer size can in training dataset original building vector with
Increase 10-30 pixel on the basis of the maximum offset of registration building vector, remote sensing image is then cut out according to rasterizing bianry image
Cut phase co-extensive, and then generate feature figure layer.It is given film size that bianry image and feature figure layer, which are required to further resampling,
Size, to meet the input requirements of multitask CNN model.
In the process by the multitask CNN model to remote sensing image and the autoregistration of building vector, as pair
Each is input to the corrected process of building vector in multitask CNN model, when the mistake of multitask CNN model output
When probability being reported to be greater than preset threshold value, then building vector currently entered is deleted, it is otherwise defeated according to multitask CNN model
Geometric correction parameter out, is corrected building vector, can be registrated with corresponding high score remote sensing image;
Specifically, area's threshold value can be arranged between 0.5-0.8 according to concrete application demand in the process of implementation, for
Node coordinate each in vector is carried out geometry according to correction parameter after obtaining its geometric correction parameter by non-wrong report vector
Transformation calculations, the vector result after correction can be obtained.
As a preferred embodiment, to avoid causing to destroy building vector to building vector exaggerated correction
Raw information calculate the friendship between the building vector of correction front and back and than index in correction course, if the friendship is simultaneously
It is less than preset metrics-thresholds than index, then does not adopt correction result;Wherein, the metrics-thresholds can be by training data
Collection carries out statistics setting: friendship and ratio that training data concentrates all original building vector sums registration building vectors are calculated,
Further calculate the friendship of all samples and than average value and middle error.Specifically, in the present embodiment, in test set
When building vector is handled, if original building vector sum is registrated the friendship of building vector and than being less than training data
The friendship of concentration and than error in average value and 3 times, then do not adopt correction result.
The present invention utilizes CNN model adaptively to generate high dimensional feature from remote sensing image for estimating building vector
Geometric correction parameter, have stronger data adaptability compared to the conventional method based on engineer's rule and feature
And generalization ability;Secondly, the present invention can be by being pointedly arranged geometric correction model to the primitive architecture of different distortion degrees
Object vector carry out registration process, be beneficial on the basis of retaining original vector resulting structure information, promoted its precision level with
Application value.
Embodiment 2:
Referring to FIG. 2, it is the system construction drawing of high score remote sensing image building detection, one kind disclosed by the invention is more
Remote sensing image and building vector registration arrangement under task CNN model, including data building module L1, full convolutional network instruction
Practice model L2, multitask CNN model training module L3 and remote sensing image matching module L4, in which:
Data building module L1 constructs training dataset for carrying out data preparation;Wherein, the training dataset is equal
Including several high score remote sensing images;For every high score remote sensing image, the training dataset further include not with high score remote sensing
The original building vector of Image registration, and will be after original building vector and high score remote sensing image progress geometric precision correction
It is registrated building vector;
Full convolutional network training pattern L2 is used to get each registration building vector using from training data concentration,
Full convolutional network model of the training for the detection of high score remote sensing image building;
Each original building vector progress time that multitask CNN model training module L3 is used to concentrate training data
It goes through, during traversal, multitask CNN model is trained using back-propagating and stochastic gradient descent algorithm;Its
In, the bianry image that will be formed after original building vector to raster conversion, and will have identical image model with the bianry image
The high score remote sensing image enclosed, after being input to the resulting full convolutional network model of training, what is obtained has phase with high score remote sensing image
With the feature figure layer of geographic range, respectively as the input item of multitask CNN model;
Remote sensing image matching module L4, should for inputting remote sensing image data into the resulting multitask CNN model of training
In model via convolution several times, pond and full attended operation after, the autoregistration of progress remote sensing image and building vector.
The embodiment of the present invention is described with above attached drawing, but the invention is not limited to above-mentioned tools
Body embodiment, the above mentioned embodiment is only schematical, rather than restrictive, the ordinary skill of this field
Personnel under the inspiration of the present invention, without breaking away from the scope protected by the purposes and claims of the present invention, can also make
Many forms, all of these belong to the protection of the present invention.
Claims (8)
1. remote sensing image and building vector method for registering under a kind of multitask CNN model, specifically includes the following steps:
S1, data preparation is carried out, constructs training dataset;It includes several high score remote sensing images that the training data, which is concentrated,;Its
In, for every high score remote sensing image, the training dataset further includes the not original building with high score remote sensing image matching
Vector, and original building vector is registrated building vector with after high score remote sensing image progress geometric precision correction;
S2, each registration building vector is got using from training data concentration, training is detected for remote sensing image building
Full convolutional network model;
S3, each the original building vector concentrated to training data traverse, and during traversal, to biography after
It broadcasts and stochastic gradient descent algorithm, and the feature figure layer to be generated using full convolutional network is as auxiliary information, to described more
Business CNN model is trained;
S4, input remote sensing image data to step S3 training multitask CNN model in, in the model via convolution several times,
After pondization and full attended operation, the autoregistration of remote sensing image and building vector is carried out.
2. remote sensing image according to claim 1 and building vector method for registering, which is characterized in that the multitask
The input item of CNN model includes: the bianry image that will be formed after original building vector to raster conversion, and will be with the binary map
As having the high score remote sensing image of identical image range, after being input to the resulting full convolutional network model of step S2 training, obtain
With high score remote sensing image have identical geographic range feature figure layer;
The output item of the multitask CNN model includes misinformation probability and the geometric correction parameter Liang Ge branch of building vector;
The penalty values Loss of the multitask CNN model, specific mathematical formulae are calculated using loss function are as follows:
Wherein, i indicates sample serial number, and N indicates total sample number in training batch, piIndicate the predicted value of current sample misinformation probability,Indicate the reference value of current sample misinformation probability;miIndicate the geometric correction parameter prediction for being directed to original building vector
Value;Indicate that original building vector is to the geometric correction reference value for being registrated building vector under non-wrong report situation, specifically,
In step S3, when traversing the original building vector that training data is concentrated, the corresponding registration of original building vector sum is built
It builds obtained by object vector compareed;Indicate the intersection entropy loss of current sample misinformation probability,
The penalty values for indicating geometric correction parameter are calculated using mean square error loss function.
3. remote sensing image according to claim 2 and building vector method for registering, which is characterized in that answer loss function
With gradient descent algorithm, when penalty values Loss levels off to X, the multitask CNN model training is finished, and after being applied to
In continuous execution step;Wherein, X >=0.
4. remote sensing image according to claim 2 and building vector method for registering, which is characterized in that pass through described more
CNN model be engaged in remote sensing image and original building vector, carries out the process of autoregistration, as to being input to multitask CNN
Each corrected process of building vector to be registered in model, when the misinformation probability of multitask CNN model output is big
When preset first threshold, then building vector currently entered is deleted, otherwise according to the several of multitask CNN model output
What correction parameter, is corrected the building vector of input, can be registrated with corresponding high score remote sensing image.
5. remote sensing image according to claim 4 and building vector method for registering, which is characterized in that for more in progress
When task CNN model training, for being input in multitask CNN model, and the building vector being deleted in correction course,
The reference value of its corresponding model output item is set as: misinformation probability 1, and geometric correction parameter is sky;For being input to multitask
In CNN model, and the building vector being retained in correction course, the reference value of corresponding model output item are set as: accidentally
Reporting probability is 0, and geometric correction parameter passes through the identical point coordinates for taking correction front and back building vector, carries out least-squares estimation and asks
.
6. remote sensing image according to claim 5 and building vector method for registering, which is characterized in that avoid to building
Object vector exaggerated correction leads to the raw information for destroying building vector, in correction course, calculates the building of correction front and back
Friendship between object vector and than index does not adopt correction result if the friendship and when being less than scheduled second threshold than index.
7. remote sensing image and building vector registration arrangement under a kind of multitask CNN model, specifically include with lower module:
Data construct module, for carrying out data preparation, construct training dataset;Wherein, if the training dataset includes
Dry Zhang Gaofen remote sensing image;For every high score remote sensing image, the training dataset further includes not matching with high score remote sensing image
Quasi- original building vector, and original building vector is registrated building with after high score remote sensing image progress geometric precision correction
Object vector;
Full convolutional network training pattern, for getting each registration building vector using from training data concentration, training is used
In the full convolutional network model of high score remote sensing image building detection;
Multitask CNN model training module, each original building vector for concentrating to training data traverse, time
During going through, multitask CNN model is trained using back-propagating and stochastic gradient descent algorithm;It wherein, will be original
The bianry image formed after building vector to raster conversion, and will have the high score of identical image range distant with the bianry image
Feel image, after being input to the resulting full convolutional network model of training, what is obtained has identical geographic range with high score remote sensing image
Feature figure layer, respectively as the input item of multitask CNN model;
Remote sensing image matching module, for inputting remote sensing image data into the resulting multitask CNN model of training, in the model
After convolution several times, pond and full attended operation, the autoregistration of remote sensing image and building vector is carried out.
8. remote sensing image according to claim 7 and building vector registration arrangement, which is characterized in that wanted using such as right
Any one of 1-6 remote sensing image and building vector method for registering are asked, the progress of remote sensing image and building vector is matched automatically
It is quasi-.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910371472.0A CN110415280B (en) | 2019-05-06 | 2019-05-06 | Remote sensing image and building vector registration method and system under multitask CNN model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910371472.0A CN110415280B (en) | 2019-05-06 | 2019-05-06 | Remote sensing image and building vector registration method and system under multitask CNN model |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110415280A true CN110415280A (en) | 2019-11-05 |
CN110415280B CN110415280B (en) | 2021-07-13 |
Family
ID=68357762
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910371472.0A Active CN110415280B (en) | 2019-05-06 | 2019-05-06 | Remote sensing image and building vector registration method and system under multitask CNN model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110415280B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113112589A (en) * | 2021-04-13 | 2021-07-13 | 哈尔滨工程大学 | Three-dimensional reconstruction method of incremental remote sensing image based on space occupation probability fusion |
CN113139453A (en) * | 2021-04-19 | 2021-07-20 | 中国地质大学(武汉) | Orthoimage high-rise building base vector extraction method based on deep learning |
CN113326759A (en) * | 2021-05-26 | 2021-08-31 | 中国地质大学(武汉) | Uncertainty estimation method for remote sensing image building identification model |
CN113420103A (en) * | 2021-06-27 | 2021-09-21 | 王彬 | Vector global database accurate survey and high-speed mapping system |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102930540A (en) * | 2012-10-26 | 2013-02-13 | 中国地质大学(武汉) | Method and system for detecting contour of urban building |
US20140064554A1 (en) * | 2011-11-14 | 2014-03-06 | San Diego State University Research Foundation | Image station matching, preprocessing, spatial registration and change detection with multi-temporal remotely-sensed imagery |
CN105809693A (en) * | 2016-03-10 | 2016-07-27 | 西安电子科技大学 | SAR image registration method based on deep neural networks |
CN109035315A (en) * | 2018-08-28 | 2018-12-18 | 武汉大学 | Merge the remote sensing image registration method and system of SIFT feature and CNN feature |
US20190043217A1 (en) * | 2015-08-26 | 2019-02-07 | Digitalglobe, Inc. | Broad area geospatial object detection using autogenerated deep learning models |
CN109345575A (en) * | 2018-09-17 | 2019-02-15 | 中国科学院深圳先进技术研究院 | A kind of method for registering images and device based on deep learning |
-
2019
- 2019-05-06 CN CN201910371472.0A patent/CN110415280B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140064554A1 (en) * | 2011-11-14 | 2014-03-06 | San Diego State University Research Foundation | Image station matching, preprocessing, spatial registration and change detection with multi-temporal remotely-sensed imagery |
CN102930540A (en) * | 2012-10-26 | 2013-02-13 | 中国地质大学(武汉) | Method and system for detecting contour of urban building |
US20190043217A1 (en) * | 2015-08-26 | 2019-02-07 | Digitalglobe, Inc. | Broad area geospatial object detection using autogenerated deep learning models |
CN105809693A (en) * | 2016-03-10 | 2016-07-27 | 西安电子科技大学 | SAR image registration method based on deep neural networks |
CN109035315A (en) * | 2018-08-28 | 2018-12-18 | 武汉大学 | Merge the remote sensing image registration method and system of SIFT feature and CNN feature |
CN109345575A (en) * | 2018-09-17 | 2019-02-15 | 中国科学院深圳先进技术研究院 | A kind of method for registering images and device based on deep learning |
Non-Patent Citations (3)
Title |
---|
GUANGMING WU ET AL.: "A Boundary Regulated Network for Accurate Roof Segmentation and Outline Extraction", 《REMOTE SENSING 2018》 * |
伍广明 等: "基于U型卷积神经网络的航空影像建筑物检测", 《测绘学报》 * |
孙金彦 等: "高分辨率遥感影像中建筑物轮廓信息矢量化", 《遥感学报》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113112589A (en) * | 2021-04-13 | 2021-07-13 | 哈尔滨工程大学 | Three-dimensional reconstruction method of incremental remote sensing image based on space occupation probability fusion |
CN113139453A (en) * | 2021-04-19 | 2021-07-20 | 中国地质大学(武汉) | Orthoimage high-rise building base vector extraction method based on deep learning |
CN113326759A (en) * | 2021-05-26 | 2021-08-31 | 中国地质大学(武汉) | Uncertainty estimation method for remote sensing image building identification model |
CN113420103A (en) * | 2021-06-27 | 2021-09-21 | 王彬 | Vector global database accurate survey and high-speed mapping system |
Also Published As
Publication number | Publication date |
---|---|
CN110415280B (en) | 2021-07-13 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110415280A (en) | Remote sensing image and building vector method for registering and system under multitask CNN model | |
CN110287846B (en) | Attention mechanism-based face key point detection method | |
CN110443144A (en) | A kind of human body image key point Attitude estimation method | |
CN108510451B (en) | Method for reconstructing license plate based on double-layer convolutional neural network | |
CN109583483A (en) | A kind of object detection method and system based on convolutional neural networks | |
CN111476188B (en) | Crowd counting method, system, medium and electronic equipment based on feature pyramid | |
CN104657986B (en) | A kind of quasi- dense matching extended method merged based on subspace with consistency constraint | |
CN110163207A (en) | One kind is based on Mask-RCNN ship target localization method and storage equipment | |
CN109145836A (en) | Ship target video detection method based on deep learning network and Kalman filtering | |
CN110263768A (en) | A kind of face identification method based on depth residual error network | |
CN102682428A (en) | Fingerprint image computer automatic mending method based on direction fields | |
CN113362247A (en) | Semantic live-action three-dimensional reconstruction method and system of laser fusion multi-view camera | |
CN113570508A (en) | Image restoration method and device, storage medium and terminal | |
CN113298736B (en) | Face image restoration method based on face pattern | |
CN113160062A (en) | Infrared image target detection method, device, equipment and storage medium | |
CN109784171A (en) | Car damage identification method for screening images, device, readable storage medium storing program for executing and server | |
CN106372597A (en) | CNN traffic detection method based on adaptive context information | |
CN115147648A (en) | Tea shoot identification method based on improved YOLOv5 target detection | |
CN106611410A (en) | Pansharpen fusion optimization method based on pyramid model | |
CN109872326A (en) | Profile testing method based on the connection of deeply network hop | |
EP3671635B1 (en) | Curvilinear object segmentation with noise priors | |
CN115019274A (en) | Pavement disease identification method integrating tracking and retrieval algorithm | |
CN107451954A (en) | Iterated pixel interpolation method based on image low-rank property | |
CN109949334A (en) | Profile testing method based on the connection of deeply network residual error | |
CN112396036A (en) | Method for re-identifying blocked pedestrians by combining space transformation network and multi-scale feature extraction |
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 |