CN109409263A - A kind of remote sensing image city feature variation detection method based on Siamese convolutional network - Google Patents
A kind of remote sensing image city feature variation detection method based on Siamese convolutional network Download PDFInfo
- Publication number
- CN109409263A CN109409263A CN201811191860.2A CN201811191860A CN109409263A CN 109409263 A CN109409263 A CN 109409263A CN 201811191860 A CN201811191860 A CN 201811191860A CN 109409263 A CN109409263 A CN 109409263A
- Authority
- CN
- China
- Prior art keywords
- variation
- scnn
- sample
- neural networks
- convolutional neural
- 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
- G06V20/35—Categorising the entire scene, e.g. birthday party or wedding scene
- G06V20/38—Outdoor scenes
- G06V20/39—Urban scenes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of remote sensing image city feature variation detection method based on Siamese convolutional network, the Siamese convolutional network is twin convolutional neural networks SCNN, it is included choosing original training set in two phase city images of registration based on data enhancing technology, twin convolutional neural networks SCNN is set, twin convolutional neural networks SCNN is trained based on original training set, original training set is expanded using data enhancing technology;Twin convolutional neural networks SCNN is trained based on the sample set after expansion, obtains trained SCNN model, realizes that the variation to city atural object detects.The present invention enhances expansion of the technology realization to sample by data, and devises a kind of Siamese convolutional neural networks, avoids the tedious steps of artificial design features in traditional change detecting method, realizes the operation of " end-to-end ";The space attribute for fully considering image improves the precision and reliability of variation detection.
Description
Technical field
The invention belongs to remote sensing image change detection techniques fields, more particularly to the change detecting method of city atural object.
Background technique
Variation detection is two obtained using same geographic area different time points or multiple remote sensing images to find ground
The process for the variation that ball surface is occurred.Variation detection is to maintain the important means of geographic information data Up-to-date state, is that remote sensing is answered
With the important research direction in one, field.Urbanization in China was constantly accelerated in recent years, and city feature changes make rapid progress.It is right
The variation detection of city atural object is for holding urban changes rule, carrying out urban map revisicn, auxiliary city planning design and political affairs
Mansion decision etc. plays a significant role.
Traditional change detecting method needs artificial design features, this is a time-consuming and laborious job, and needs stronger
Professional knowledge.And it is difficult to design a kind of generic features suitable for all types of ground objects.In recent years, depth learning technology
Development is very fast, has obtained certain application in image recognition and variation detection field.The multilayered nonlinear of deep neural network reflects
The ability for making it have fitting arbitrary function is penetrated, therefore the classifying face of higher-dimension can be constructed, completes pattern classification in high quality
Identification mission.The main research and utilization deep neural network of the present invention realizes that the variation of " end-to-end " detects to city atural object, avoids
The process of artificial design features promotes the precision of variation detection.(bibliography: Tewkesbury A P, Comber A J,
Tate N J, et al, a critical synthesis of remotely sensed optical image change
Detection techniques, Remote Sensing of Environment, 2015;Ian Goodfellow,Yoshua
Bengio and Aaron Courville, Deep Learning, MIT Press, 2016)
Currently, the common deep neural network model of high score remote sensing imagery change detection includes autoencoder network, depth confidence net
Network, convolutional neural networks etc..Autoencoder network, depth confidence network convert one-dimensional vector for bidimensional image and are input to network mould
In type, it is lost the spatial information of image.Convolutional neural networks use locally-attached thought, using a local receptor field as
The minimum unit of feature extraction has fully considered the spatial information of image.But current research method is all based on the change of pixel
Change detection, is inputted using each neighborhood of pixels range of image as training, deep neural network is trained, model is obtained, then
Detection is changed to whole image.This method for changing detection pixel-by-pixel, using the moving window of fixed size as mind
Input through network, the information content for including is limited, cannot give full play to the advantage of deep neural network study complex characteristic.And with
Pixel is analytical unit, there is largely broken pseudo- variation.The currently used variation detection based on scene is variation with scene
The analytical unit of detection needs to capture scene, and cannot be changed inspection there are more non-scene areas in image
It surveys, therefore " all standing " variation detection cannot be carried out to image.(bibliography: Xu Y, Xiang S, Huo C, et al,
Change detection based on auto-encoder model for VHR images, Proceedings of
SPIE-The International Society for Optical Engineering, 2013;Argyridis A,
Argialas D P, Building change detection through multi-scale GEOBIA approach by
Integrating deep belief networks with fuzzy ontologies, International Journal
Of Image&Data Fusion, 2016;Liu J, Gong M, Qin K, et al, A Deep Convolutional
Coupling Network for Change Detection Based on Heterogeneous Optical and
Radar Images, IEEE Transactions on Neural Networks&Learning Systems, 2016;Wu C,
Zhang L, Zhang L, A scene change detection framework for multi-temporal very
High resolution remote sensing images, Signal Processing, 2016)
Deep learning is the machine learning method of data-driven.Since the deep neural network number of plies is deeper, parameter is more, because
This needs great amount of samples to train deep neural network.And making a tape label sample set is a time-consuming and laborious job,
And need certain professional knowledge.To solve this problem, Alex Krizhevsky etc. is using " data enhancing " technology to training
Sample is expanded.The operations such as general data enhancing technology includes greyscale transformation, rotate, cut randomly, colour dither are expanded
Fill sample.This data increase technology and achieve preferable effect in image classification, identification, object detection field.But the skill
Art is applied but to be difficult to prove effective in variation detection field.The reason is that obtain two of same geographic area different time points or multiple are distant
Feel image itself due to shooting time, weather conditions, light condition, shooting angle etc. are different and there are larger difference, simple benefits
With greyscale transformation, rotate, cut randomly, the operations such as colour dither not can increase the diversity of sample.Therefore, to variation detection
When sample is expanded, this feature of having differences property in itself is fully considered between multi-temporal remote sensing image, in conjunction with variation number
According to the characteristics of, seek suitable sample extending method.(bibliography: Krizhevsky A, Sutskever I, Hinton G
E, ImageNet classification with deep convolutional neural networks,
International Conference on Neural Information Processing Systems.Curran
Associates Inc, 2012;Zhong Y, Large patch convolutional neural networks for the
Scene classification of high spatial resolution imagery, Journal of Applied
Remote Sensing, 2016.)
Summary of the invention
For the deficiencies in the prior art, the present invention proposes a kind of city atural object based on Siamese convolutional network
Change detection techniques, and special data enhancement methods are proposed for variation test problems.
The technical scheme adopted by the invention is as follows a kind of remote sensing image city feature changes based on Siamese convolutional network
Detection method, the Siamese convolutional network be twin convolutional neural networks SCNN, based on data enhancing technology included with
Lower step,
Step 1, original training set is chosen in two phase city images of registration, including selection variation and constant area
Block is divided into the segmentation block of fixed size respectively, to each segmentation block to label is assigned, obtains two phase samples pair;
Step 2, twin convolutional neural networks SCNN is set, based on step 1 gained original training set to twin convolutional Neural
Network SCNN is trained, the twin convolutional neural networks SCNN after obtaining initial training;
The twin convolutional neural networks SCNN is made of two identical branching networks and decision-making level's network,
Branching networks are located at twin convolutional neural networks SCNN low layer, and two branching networks have identical structure and parameter, point
Other to carry out feature extraction to two phase image samples, the feature after extraction is connected by feature, generates the whole of two phase samples pair
Body characteristics are input in the decision layer network of top layer;Decision-making level's network implementations similarity measurement model, to the global feature of input
Carry out similarity measurement;
Step 3, based on the twin convolutional neural networks SCNN after step 2 gained initial training, enhance technology using data
Original training set is expanded;
Step 4, twin convolutional neural networks SCNN is trained based on the sample set after expanding obtained by step 3, is obtained
Trained SCNN model;
Step 5, it is based on the trained SCNN model of step 4, a new urban area is tested, is realized to city
The variation of atural object detects.
Moreover, step 3 expands original training set using data enhancing technology, including following sub-step,
Step 3.1, based on rule --- region of variation is a small number of in two phase images, invariant region be it is most, it is right
Constant sample carries out data enhancing;
Step 3.2, based on rule --- to each variation segmentation block number evidence in phase 1, it is located at difference therewith in phase 2
The segmentation block of position can divide block composition variation sample pair with the variation, carry out data enhancing to variation sample.
Moreover, step 3.1 includes following sub-step,
Step 3.1.1, the twin convolutional neural networks SCNN after 2 gained initial training of input step;
Step 3.1.2 selects the extended area image of two phases, is divided into the segmentation block of fixed size;
Step 3.1.3 will respectively divide the twin convolutional Neural net after block is input to initial training obtained in step 3.1.2
In network SCNN, it is changed detection, is obtained as a result, determining that each segmentation block is region of variation or invariant region;
Step 3.1.4 selects invariant region from the result that step 3.1.3 is obtained, by adjusting threshold parameter T, it is ensured that
Do not include region of variation in the invariant region selected, as probability > T of constant sample, sample is constant sample, other each point
Cutting block is region of variation;
The invariant region selected is added in the constant sample of the gained original training set of step 1 and realizes by step 3.1.5
It is expanded.
Moreover, step 3.2 includes following sub-step,
Step 3.2.1, based on region of variation obtained by step 3.1.4, by all changes segmentation block duplication in 1 image of phase
Num times, and rename, the variation sample of phase 1 is extended for original (num+1) times, and num is preset numerical value;
Step 3.2.2, based on region of variation obtained by step 3.1.4, the variation segmentation of each of corresponding 1 image of phase
Block randomly selects num in 2 image of phase and is located at the segmentation block of different location therewith, and renames, with step 3.2.1
The segmentation block composition num of the corresponding phase 1 of duplication is to variation image pair;
Variation sample obtained by step 3.2.2 is added to the variation sample of step 1 gained original training set by step 3.2.3
Middle realization is expanded.
Moreover, each branching networks of twin convolutional neural networks SCNN separately include 5 convolutional layers, 3 pond layers and
One full articulamentum.
Moreover, the decision layer network of twin convolutional neural networks SCNN is made of three full articulamentums.
Moreover, having been trained when being trained to twin convolutional neural networks SCNN using transfer learning strategy, including use
Good CaffeNet model parameter, initializes branching networks parameter, what the speed and variation for improving network training detected
Precision.
The present invention passes through expansion of specific " data enhancing " the technology realization to sample, and devises a kind of Siamese volumes
Product neural network (Siamese Convolutional Neural Network, SCNN) is realized to two phase high score image cities
The variation of city region atural object detects.The present invention is based in two phase images " region of variation be it is a small number of, invariant region is most
" this rule, the expansion to constant sample is realized by the way of " expansion of repetitive exercise SCNN --- samples selection -- sample "
It fills;Based on " in two phases positioned at different location segmentation block constitute variation sample to " realize to variation sample expansion,
Deep learning is solved to the Dependence Problem of big data.The beneficial effects of the present invention are: the invention avoids traditional variation inspections
The tedious steps of artificial design features in survey method, realize the operation of " end-to-end ";It overcomes based on autoencoder network and depth
The problem of spending the spatial information of change detecting method loss image of confidence network, fully considers the space attribute of image, improves
The precision and reliability of variation detection.
Detailed description of the invention
Fig. 1 is the training stage flow chart of the embodiment of the present invention.
Fig. 2 is the test phase flow chart of the embodiment of the present invention.
Specific embodiment
Understand for the ease of those of ordinary skill in the art and implement the present invention, with reference to the accompanying drawings and embodiments to this hair
It is bright to be described in further detail, it should be understood that implementation example described herein is merely to illustrate and explain the present invention, not
For limiting the present invention.
Referring to Fig. 1 and Fig. 2, a kind of remote sensing image city based on Siamese convolutional network provided in an embodiment of the present invention
Feature changes detection technique, comprising the following steps:
Step 1: choosing original training set in two phase city images of registration;
It includes following sub-step that the original training set of the embodiment of the present invention, which chooses specific implementation:
Step 1.1: inputting the image of two phases, be to be studied for city image in the present invention, certain city may be selected
The image of two phase of region, is denoted as prime area image, when it is implemented, a variation detection samples selection work can be provided
The image of two phases is carried out two window linkage displays by tool;
Step 1.2: the image selection variation of two phases of comparison and constant block, when specific implementation, can be selected by subscriber frame
Variation and constant block, or predefine variation and constant block;
Step 1.3: will change in the image of two phases and constant block is divided into the segmentation block of fixed size (in fact respectively
Example is applied using 64*64 pixel), segmentation block pair is formed, and obtain two phase samples pair to label is assigned to each segmentation block, it is real
Now construct sample set, wherein embodiment setting label (1,0) represents region of variation, and label (0,1) represents invariant region.
Step 2: design twin convolutional neural networks (Siamese Convolutional Neural Network, SCNN,
That is Siamese convolutional network) structure, twin convolutional neural networks SCNN is trained based on step 1 gained original training set,
SCNN model after obtaining initial training;
SCNN is made of two identical branching networks and decision-making level's network, and branching networks are located at the low of SCNN
Layer, two branching networks have identical structure and parameter, carry out feature extraction to two phase image samples respectively.It extracts
Feature afterwards is connected by feature, is generated the global feature of two phase samples pair, is input in the decision layer network of top layer.Decision
Layer network is substantially a similarity measurement model, carries out similarity measurement to the global feature of input.
The SCNN structure design of the embodiment of the present invention is specific as follows:
Two branching networks of SCNN low layer are complete using the convolutional layer of convolutional neural networks frame (CaffeNet) and first
Articulamentum, two branching networks have identical structure and parameter, separately include 5 convolutional layers, 3 pond layers and one
Full articulamentum;
The two phase image sample characteristics that two branching networks are extracted carry out feature connection, generate two phase samples pair
Global feature;
The decision layer network of SCNN high level is made of three full articulamentums, and decision layer network is substantially that a similarity is surveyed
Model is measured, similarity measurement is carried out to the global feature of input, the output of SCNN is (prob1, prob2), indicates that the sample is
The probability for changing sample is prob1, be the probability of constant sample is prob2 (prob1+prob2=1).
When being trained based on step 1 gained original training set to twin convolutional neural networks SCNN, training process is as follows,
Firstly, by the sample set after expansion be divided into training sample set (80%) and verifying sample set (20%), can be used with
The mode of machine is divided;(training sample set can be input into SCNN later, be carried out based on gradient decline and back-propagation algorithm
Repetitive exercise;As the number of iterations=T1When, sample set is verified to verify the model for being trained to collection training, is verified precision;When
The number of iterations=T2, preservation model (T2=nT1)。)
When it is implemented, can be by user preset T1And T2Value, can every wheel execute T1Secondary iteration is protected after executing n wheel
Deposit model.
Training sample set and verifying sample set all contain region of variation and invariant region respectively.
Then, training sample set is input in SCNN and is iterated training, obtain trained SCNN model;
The SCNN repetitive exercise specific implementation of the embodiment of the present invention includes following sub-step:
Step a: by the model parameter of the convolutional layer of convolutional neural networks frame (CaffeNet) and first full articulamentum
It is directed respectively into two branching networks of SCNN, realizes the initialization to two branching networks parameters;
Model parameter uses empirical value in this step, and the migration of trained CaffeNet model parameter is come, to point
Branch network parameter is initialized, and can effectively improve the speed of network training and the precision of variation detection in this way.
Step b: random initializtion is carried out to decision-making level's network parameter of SCNN;
Step c: network hyper parameter is arranged to for input to change testing result as output in corresponding sample when with two,
Empirical value can be used when specific implementation, such as: small quantities of sample number: 1000;Learning rate: 0.01;The ratio of random inactivation neuron:
0.5, it is based on stochastic gradient descent and back-propagation algorithm, training is iterated to SCNN, until judging mould by verifying precision
Type convergence, saves optimal SCNN model.When it is implemented, can rise and decline according to precision function curve and loss function curve
Situation is judged, is restrained when curve reaches steady.
Stochastic gradient descent and back-propagation algorithm are the prior art, and it will not go into details by the present invention.
Step 3: " data enhancing " technology proposed through the invention expands original training set;
It includes following sub-step that the sample set of the embodiment of the present invention, which expands specific implementation:
Step 3.1: based on " region of variation be a small number of, invariant region is most " this rule in two phase images,
Data enhancing is carried out to constant sample;
The present embodiment is to the data extending specific implementation process of constant sample:
Step 3.1.1: the SCNN model after 2 gained initial training of input step is based on original training set in step 2
The result that twin convolutional neural networks SCNN is trained;
Step 3.1.2: the image in a larger area region is selected in city, is by the large area region Image Segmentation
With the segmentation block of step 1 division size corresponding (being 64*64 pixel in embodiment);
It is consistent when with the two of step 1.1 prime area image, when another large area two in city may be selected in this step
The image of phase, as extended area image.
Step 3.1.3: it will respectively divide block image obtained in step 3.1.2 and be input to the SCNN mould that step 3.1.1 is obtained
It in type, is changed detection, obtains result (each segmentation block is region of variation or invariant region);
Step 3.1.4: selecting invariant region from the result that step 3.1.3 is obtained, by adjusting threshold parameter T, it is ensured that
Do not include region of variation in the invariant region selected, as probability > T of constant sample, sample is constant sample, other each point
Cutting block is region of variation;
In embodiment, threshold value T=0.90 is set, when the probability of constant sample is greater than 0.90, sample is constant sample.
Step 3.1.5: the invariant region selected is added in the constant sample of the gained original training set of step 1 and is realized
It is expanded.
Step 3.2: to each variation segmentation block number evidence in phase 1, being located at the segmentation block of different location in phase 2 therewith
Block composition variation sample pair can be divided with the variation.Based on the theory, data enhancing is carried out to variation sample.
The present embodiment is to the data extending specific implementation process of variation sample:
Step 3.2.1: based on region of variation obtained by step 3.1.4, all changes segmentation block image in 1 image of phase is answered
It is num times processed, and rename, the variation sample of phase 1 is extended for original (num+1) times;
Step 3.2.2: based on region of variation obtained by step 3.1.4, the variation segmentation of each of corresponding 1 image of phase
Block randomly selects num in 2 image of phase and is located at the segmentation block of different location therewith, and renames, with step 3.2.1
The segmentation block composition num of the corresponding phase 1 of duplication is to variation image pair;
Step 3.2.3: variation sample obtained by step 3.2.2 is added to the variation sample of the gained original training set of step 1
It is realized in this and it is expanded.
All changes segmentation block in 1 image of phase is respectively adopted at the mode of step 3.2.1 and step 3.2.2
After reason, the variation sample of phase 2 is extended for original (num+1) times, is added in the variation sample of original training set.
When it is implemented, the value of num can be preset.For the image of two phase of the same area, each in phase 1 becomes
Change sample, a new variation sample pair can be formed with the variation sample positioned at the sample standard deviation of different location therewith in phase 2.Base
It in this it is assumed that choosing a variation sample in 1 image of phase, and replicates num times, num is randomly selected in 2 image of phase
It is located at the sample of different location therewith, with 1 image of phase composition umn to variation image pair.All original variation samples are executed
The operation, variation sample are extended for original (num+1) times.For example, some 1 sample of phase replicates 4, in 2 sample of phase
In randomly select 4 samples for being located at different location therewith, form 4 pairs of variation images pair with 4 samples of phase 1.In addition original
This variation image pair, totally 5 variation images pair.In this way, variation sample is extended for original 5 times.
Step 4: twin convolutional neural networks SCNN being trained based on the sample set after expanding obtained by step 3, is obtained
Trained SCNN model.
Trained mode is consistent with the mode that step 2 first time is trained again, it may be assumed that
Firstly, the sample set after expansion is divided into training sample set (80%) and verifying sample set (20%);(training sample
Collection can be input into SCNN later, be iterated training based on gradient decline and back-propagation algorithm;When the number of iterations=
T1When, sample set is verified to verify the model for being trained to collection training, is verified precision;As the number of iterations=T2When, preservation model
(T2=nT1).) at this point, T1And T2Value can be adjusted to step 2 for the first time train when it is different.
Then, training sample set is input in SCNN and is iterated training, obtain trained SCNN model;
The SCNN repetitive exercise specific implementation of the embodiment of the present invention includes following sub-step:
Step a, by the model parameter of the convolutional layer of convolutional neural networks frame (CaffeNet) and first full articulamentum
It is directed respectively into two branching networks of SCNN, realizes the initialization to two branching networks parameters;
Model parameter equally uses empirical value in this step, and trained CaffeNet model parameter migration is come,
Branching networks parameter is initialized, can effectively improve the speed of network training and the precision of variation detection in this way.Specifically
When implementation, can using and step 2 in same initial parameter, using the training sample after expansion, re -training mainly optimizes
Decision-making level's network parameter.
Step b carries out random initializtion to decision-making level's network parameter of SCNN;
Step c, suitable network is arranged to for input to change testing result as output in corresponding sample when with two
Hyper parameter, the value as in step 2 settable at this time: small quantities of sample number: 1000;Learning rate: 0.01;Random inactivation nerve
Member ratio: 0.5, be based on stochastic gradient descent and back-propagation algorithm, training is iterated to SCNN, until model restrain,
Save optimal SCNN model.
Step 5: it is based on the trained SCNN model of step 4, is tested, it can be to a new large area urban area
" segmentation --- variation detection --- precision evaluation " is carried out, realizes that the variation to city atural object detects;
The embodiment of the present invention based on trained SCNN model, atural object change is carried out to a new large area urban area
Change detection.Referring to fig. 2, specific implementation process is:
Step 5.1: one large area urban area not being overlapped with training sample data of selection, by the two phase shadow of region
As being divided into the segmentation block with step 1 division size corresponding (being 64*64 pixel in embodiment) respectively;
Consistent when with the two of step 1.1 prime area image, another big two phase of region in city may be selected in this step
Image, be not overlapped with prime area image and extended area image, be denoted as test zone image.
Step 5.2: two phase Image Segmentation blocks are input in the trained SCNN model of step 4, detection is changed,
Export result;
Step 5.3: precision evaluation and result visualization are carried out to variation testing result obtained by step 5.2.
The visualization result of variation detection indicates that white represents the segmentation block of variation, and black represents constant with two-value image
Segmentation block.Changed according to ground referring to figure, can be schemed according to reference and variation testing result statistics is by correct, error detection
Divide the quantity of block, calculates cartographic accuracy, user's precision, omission factor, false detection rate, overall accuracy, the evaluation of Kappa coefficient equally accurate
Index verifies the validity of method proposed by the present invention to examine the precision of present invention variation testing result.
When it is implemented, the automatic running that computer software technology realizes the above process can be used.
It is tested using the technical solution of the embodiment of the present invention, extracts variation testing result visualization figure:
Taking image (a) --- Wuhan City region remote sensing image in 2010 is the 17 grades of images obtained from ArcGIS platform,
Image resolution is 1.19 meters/pixel;Taking image (b) --- the same area remote sensing image in 2013 is from Microsoft's map platform
The 17 grades of images obtained, image resolution are 1.19 meters/pixel;Image (c) --- ground variation is by artificial mesh referring to figure
Depending on the two phase real change striographs interpreted and factual survey obtains, white represents region of variation in figure, and black represents constant
Region;Using the above process provided by the invention, image (d) finally can be obtained --- it is detected by above-mentioned change detecting method
Variation testing result figure out.Effectiveness of the invention can be confirmed by changing testing result figure.
It should be understood that the part that this specification does not elaborate belongs to the prior art.
It should be understood that the above-mentioned description for preferred embodiment is more detailed, can not therefore be considered to this
The limitation of invention patent protection range, those skilled in the art under the inspiration of the present invention, are not departing from power of the present invention
Benefit requires to make replacement or deformation under protected ambit, fall within the scope of protection of the present invention, this hair
It is bright range is claimed to be determined by the appended claims.
Claims (7)
1. a kind of remote sensing image city feature variation detection method based on Siamese convolutional network, it is characterised in that: described
Siamese convolutional network is twin convolutional neural networks SCNN, is included the following steps based on data enhancing technology,
Step 1, original training set is chosen in two phase city images of registration, including selection variation and constant block, divided
It is not divided into the segmentation block of fixed size, to each segmentation block to label is assigned, obtains two phase samples pair;
Step 2, twin convolutional neural networks SCNN is set, based on step 1 gained original training set to twin convolutional neural networks
SCNN is trained, the twin convolutional neural networks SCNN after obtaining initial training;
The twin convolutional neural networks SCNN is made of two identical branching networks and decision-making level's network, branch
Network is located at twin convolutional neural networks SCNN low layer, and two branching networks have identical structure and parameter, right respectively
Two phase image samples carry out feature extraction, and the feature after extraction is connected by feature, generate the whole special of two phase samples pair
Sign, is input in the decision layer network of top layer;Decision-making level's network implementations similarity measurement model carries out the global feature of input
Similarity measurement;
Step 3, based on the twin convolutional neural networks SCNN after step 2 gained initial training, using data enhancing technology to first
Beginning sample set is expanded;
Step 4, twin convolutional neural networks SCNN is trained based on the sample set after expanding obtained by step 3, is trained
Good SCNN model;
Step 5, it is based on the trained SCNN model of step 4, a new urban area is tested, is realized to city atural object
Variation detection.
2. the remote sensing image city feature variation detection method based on Siamese convolutional network according to claim 1, special
Sign is: step 3 expands original training set using data enhancing technology, including following sub-step,
Step 3.1, based on rule --- region of variation is a small number of in two phase images, invariant region be it is most, to constant
Sample carries out data enhancing;
Step 3.2, based on rule --- to each variation segmentation block number evidence in phase 1, it is located at different location therewith in phase 2
Segmentation block can with the variation divide block composition variation sample pair, to variation sample carry out data enhancing.
3. the remote sensing image city feature variation detection method based on Siamese convolutional network according to claim 2, special
Sign is: step 3.1 includes following sub-step,
Step 3.1.1, the twin convolutional neural networks SCNN after 2 gained initial training of input step;
Step 3.1.2 selects the extended area image of two phases, is divided into the segmentation block of fixed size;
Step 3.1.3 will respectively divide the twin convolutional neural networks after block is input to initial training obtained in step 3.1.2
In SCNN, it is changed detection, is obtained as a result, determining that each segmentation block is region of variation or invariant region;
Step 3.1.4 selects invariant region from the result that step 3.1.3 is obtained, by adjusting threshold parameter T, it is ensured that select
Invariant region in do not include region of variation, as probability > T of constant sample, sample be constant sample, other each segmentation blocks
For region of variation;
The invariant region selected is added in the constant sample of the gained original training set of step 1 and realizes to it by step 3.1.5
Expand.
4. the remote sensing image city feature variation detection method based on Siamese convolutional network according to claim 3, special
Sign is: step 3.2 includes following sub-step,
Step 3.2.1 is replicated all changes segmentation block in 1 image of phase num times based on region of variation obtained by step 3.1.4,
And rename, the variation sample of phase 1 is extended for original (num+1) times, and num is preset numerical value;
Step 3.2.2, based on region of variation obtained by step 3.1.4, each of corresponding 1 image of phase variation segmentation block,
Num are randomly selected in 2 image of phase and is located at the segmentation block of different location therewith, and is renamed, with step 3.2.1 duplication
The segmentation block composition num of corresponding phase 1 is to variation image pair;
Variation sample obtained by step 3.2.2 is added in the variation sample of step 1 gained original training set real by step 3.2.3
Now expand.
5. the according to claim 1 or 2 or 3 or 4 remote sensing image city feature changes detections based on Siamese convolutional network
Method, it is characterised in that: each branching networks of twin convolutional neural networks SCNN separately include 5 convolutional layers, 3 pond layers
With a full articulamentum.
6. the according to claim 1 or 2 or 3 or 4 remote sensing image city feature changes detections based on Siamese convolutional network
Method, it is characterised in that: the decision layer network of twin convolutional neural networks SCNN is made of three full articulamentums.
7. the according to claim 1 or 2 or 3 or 4 remote sensing image city feature changes detections based on Siamese convolutional network
Method, it is characterised in that: when being trained to twin convolutional neural networks SCNN, using transfer learning strategy, including using
Trained CaffeNet model parameter, initializes branching networks parameter, improves the speed and variation inspection of network training
The precision of survey.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811191860.2A CN109409263B (en) | 2018-10-12 | 2018-10-12 | Method for detecting urban ground feature change of remote sensing image based on Siamese convolutional network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811191860.2A CN109409263B (en) | 2018-10-12 | 2018-10-12 | Method for detecting urban ground feature change of remote sensing image based on Siamese convolutional network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109409263A true CN109409263A (en) | 2019-03-01 |
CN109409263B CN109409263B (en) | 2021-05-04 |
Family
ID=65467086
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811191860.2A Active CN109409263B (en) | 2018-10-12 | 2018-10-12 | Method for detecting urban ground feature change of remote sensing image based on Siamese convolutional network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109409263B (en) |
Cited By (22)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109934166A (en) * | 2019-03-12 | 2019-06-25 | 中山大学 | Unmanned plane image change detection method based on semantic segmentation and twin neural network |
CN110059658A (en) * | 2019-04-26 | 2019-07-26 | 北京理工大学 | A kind of satellite-remote-sensing image multidate change detecting method based on Three dimensional convolution neural network |
CN110070140A (en) * | 2019-04-28 | 2019-07-30 | 清华大学 | Method and device is determined based on user's similitude of multi-class information |
CN110378224A (en) * | 2019-06-14 | 2019-10-25 | 香港理工大学深圳研究院 | A kind of detection method of feature changes, detection system and terminal |
CN110570397A (en) * | 2019-08-13 | 2019-12-13 | 创新奇智(重庆)科技有限公司 | Method for detecting ready-made clothes printing defects based on deep learning template matching algorithm |
CN110610191A (en) * | 2019-08-05 | 2019-12-24 | 深圳优地科技有限公司 | Elevator floor identification method and device and terminal equipment |
CN110897630A (en) * | 2019-06-27 | 2020-03-24 | 上海数创医疗科技有限公司 | Ventricular tachycardia and heart rhythm identification method and device based on transfer learning |
CN110969088A (en) * | 2019-11-01 | 2020-04-07 | 华东师范大学 | Remote sensing image change detection method based on significance detection and depth twin neural network |
CN111079671A (en) * | 2019-12-20 | 2020-04-28 | 深圳集智数字科技有限公司 | Method and device for detecting abnormal articles in scene |
CN111127532A (en) * | 2019-12-31 | 2020-05-08 | 成都信息工程大学 | Medical image deformation registration method and system based on deep learning characteristic optical flow |
CN111161218A (en) * | 2019-12-10 | 2020-05-15 | 核工业北京地质研究院 | High-resolution remote sensing image change detection method based on twin convolutional neural network |
CN111325134A (en) * | 2020-02-17 | 2020-06-23 | 武汉大学 | Remote sensing image change detection method based on cross-layer connection convolutional neural network |
CN111640159A (en) * | 2020-05-11 | 2020-09-08 | 武汉大学 | Remote sensing image change detection method based on twin convolutional neural network |
CN111738972A (en) * | 2019-03-19 | 2020-10-02 | 阿里巴巴集团控股有限公司 | Building detection system, method and device |
CN112016436A (en) * | 2020-08-28 | 2020-12-01 | 北京国遥新天地信息技术有限公司 | Remote sensing image change detection method based on deep learning |
CN112215085A (en) * | 2020-09-17 | 2021-01-12 | 云南电网有限责任公司昆明供电局 | Power transmission corridor foreign matter detection method and system based on twin network |
CN112396594A (en) * | 2020-11-27 | 2021-02-23 | 广东电网有限责任公司肇庆供电局 | Change detection model acquisition method and device, change detection method, computer device and readable storage medium |
CN112784929A (en) * | 2021-03-14 | 2021-05-11 | 西北工业大学 | Small sample image classification method and device based on double-element group expansion |
CN113158855A (en) * | 2021-04-08 | 2021-07-23 | 成都国星宇航科技有限公司 | Remote sensing image auxiliary processing method and device based on online learning |
CN113326724A (en) * | 2021-02-07 | 2021-08-31 | 海南长光卫星信息技术有限公司 | Method, device and equipment for detecting change of remote sensing image and readable storage medium |
CN113609889A (en) * | 2021-05-18 | 2021-11-05 | 武汉大学 | High-resolution remote sensing image vegetation extraction method based on sensitive feature focusing perception |
CN115019186A (en) * | 2022-08-08 | 2022-09-06 | 中科星图测控技术(合肥)有限公司 | Algorithm and system for remote sensing change detection |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105740894A (en) * | 2016-01-28 | 2016-07-06 | 北京航空航天大学 | Semantic annotation method for hyperspectral remote sensing image |
US20170140248A1 (en) * | 2015-11-13 | 2017-05-18 | Adobe Systems Incorporated | Learning image representation by distilling from multi-task networks |
CN106897681A (en) * | 2017-02-15 | 2017-06-27 | 武汉喜恩卓科技有限责任公司 | A kind of remote sensing images comparative analysis method and system |
CN108573276A (en) * | 2018-03-12 | 2018-09-25 | 浙江大学 | A kind of change detecting method based on high-resolution remote sensing image |
-
2018
- 2018-10-12 CN CN201811191860.2A patent/CN109409263B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20170140248A1 (en) * | 2015-11-13 | 2017-05-18 | Adobe Systems Incorporated | Learning image representation by distilling from multi-task networks |
CN105740894A (en) * | 2016-01-28 | 2016-07-06 | 北京航空航天大学 | Semantic annotation method for hyperspectral remote sensing image |
CN106897681A (en) * | 2017-02-15 | 2017-06-27 | 武汉喜恩卓科技有限责任公司 | A kind of remote sensing images comparative analysis method and system |
CN108573276A (en) * | 2018-03-12 | 2018-09-25 | 浙江大学 | A kind of change detecting method based on high-resolution remote sensing image |
Cited By (32)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109934166A (en) * | 2019-03-12 | 2019-06-25 | 中山大学 | Unmanned plane image change detection method based on semantic segmentation and twin neural network |
CN111738972B (en) * | 2019-03-19 | 2024-05-28 | 阿里巴巴集团控股有限公司 | Building detection system, method and device |
CN111738972A (en) * | 2019-03-19 | 2020-10-02 | 阿里巴巴集团控股有限公司 | Building detection system, method and device |
CN110059658A (en) * | 2019-04-26 | 2019-07-26 | 北京理工大学 | A kind of satellite-remote-sensing image multidate change detecting method based on Three dimensional convolution neural network |
CN110059658B (en) * | 2019-04-26 | 2020-11-24 | 北京理工大学 | Remote sensing satellite image multi-temporal change detection method based on three-dimensional convolutional neural network |
CN110070140A (en) * | 2019-04-28 | 2019-07-30 | 清华大学 | Method and device is determined based on user's similitude of multi-class information |
CN110070140B (en) * | 2019-04-28 | 2021-03-23 | 清华大学 | User similarity determination method and device based on multi-category information |
CN110378224A (en) * | 2019-06-14 | 2019-10-25 | 香港理工大学深圳研究院 | A kind of detection method of feature changes, detection system and terminal |
CN110378224B (en) * | 2019-06-14 | 2021-01-05 | 香港理工大学深圳研究院 | Detection method and detection system for ground feature change and terminal |
CN110897630A (en) * | 2019-06-27 | 2020-03-24 | 上海数创医疗科技有限公司 | Ventricular tachycardia and heart rhythm identification method and device based on transfer learning |
CN110610191A (en) * | 2019-08-05 | 2019-12-24 | 深圳优地科技有限公司 | Elevator floor identification method and device and terminal equipment |
CN110570397A (en) * | 2019-08-13 | 2019-12-13 | 创新奇智(重庆)科技有限公司 | Method for detecting ready-made clothes printing defects based on deep learning template matching algorithm |
CN110969088A (en) * | 2019-11-01 | 2020-04-07 | 华东师范大学 | Remote sensing image change detection method based on significance detection and depth twin neural network |
CN111161218A (en) * | 2019-12-10 | 2020-05-15 | 核工业北京地质研究院 | High-resolution remote sensing image change detection method based on twin convolutional neural network |
CN111079671A (en) * | 2019-12-20 | 2020-04-28 | 深圳集智数字科技有限公司 | Method and device for detecting abnormal articles in scene |
CN111127532B (en) * | 2019-12-31 | 2020-12-22 | 成都信息工程大学 | Medical image deformation registration method and system based on deep learning characteristic optical flow |
CN111127532A (en) * | 2019-12-31 | 2020-05-08 | 成都信息工程大学 | Medical image deformation registration method and system based on deep learning characteristic optical flow |
CN111325134A (en) * | 2020-02-17 | 2020-06-23 | 武汉大学 | Remote sensing image change detection method based on cross-layer connection convolutional neural network |
CN111640159B (en) * | 2020-05-11 | 2022-04-01 | 武汉大学 | Remote sensing image change detection method based on twin convolutional neural network |
CN111640159A (en) * | 2020-05-11 | 2020-09-08 | 武汉大学 | Remote sensing image change detection method based on twin convolutional neural network |
CN112016436A (en) * | 2020-08-28 | 2020-12-01 | 北京国遥新天地信息技术有限公司 | Remote sensing image change detection method based on deep learning |
CN112215085A (en) * | 2020-09-17 | 2021-01-12 | 云南电网有限责任公司昆明供电局 | Power transmission corridor foreign matter detection method and system based on twin network |
CN112396594A (en) * | 2020-11-27 | 2021-02-23 | 广东电网有限责任公司肇庆供电局 | Change detection model acquisition method and device, change detection method, computer device and readable storage medium |
CN112396594B (en) * | 2020-11-27 | 2024-03-29 | 广东电网有限责任公司肇庆供电局 | Method and device for acquiring change detection model, change detection method, computer equipment and readable storage medium |
CN113326724B (en) * | 2021-02-07 | 2024-02-02 | 海南长光卫星信息技术有限公司 | Remote sensing image change detection method, device, equipment and readable storage medium |
CN113326724A (en) * | 2021-02-07 | 2021-08-31 | 海南长光卫星信息技术有限公司 | Method, device and equipment for detecting change of remote sensing image and readable storage medium |
CN112784929B (en) * | 2021-03-14 | 2023-03-28 | 西北工业大学 | Small sample image classification method and device based on double-element group expansion |
CN112784929A (en) * | 2021-03-14 | 2021-05-11 | 西北工业大学 | Small sample image classification method and device based on double-element group expansion |
CN113158855A (en) * | 2021-04-08 | 2021-07-23 | 成都国星宇航科技有限公司 | Remote sensing image auxiliary processing method and device based on online learning |
CN113609889A (en) * | 2021-05-18 | 2021-11-05 | 武汉大学 | High-resolution remote sensing image vegetation extraction method based on sensitive feature focusing perception |
CN113609889B (en) * | 2021-05-18 | 2024-02-13 | 武汉大学 | High-resolution remote sensing image vegetation extraction method based on sensitive characteristic focusing perception |
CN115019186A (en) * | 2022-08-08 | 2022-09-06 | 中科星图测控技术(合肥)有限公司 | Algorithm and system for remote sensing change detection |
Also Published As
Publication number | Publication date |
---|---|
CN109409263B (en) | 2021-05-04 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109409263A (en) | A kind of remote sensing image city feature variation detection method based on Siamese convolutional network | |
Henry et al. | Road segmentation in SAR satellite images with deep fully convolutional neural networks | |
CN111986099B (en) | Tillage monitoring method and system based on convolutional neural network with residual error correction fused | |
Zhang et al. | Joint Deep Learning for land cover and land use classification | |
CN108596101B (en) | Remote sensing image multi-target detection method based on convolutional neural network | |
Sun et al. | Leveraging crowdsourced GPS data for road extraction from aerial imagery | |
EP3614308B1 (en) | Joint deep learning for land cover and land use classification | |
CN108596055B (en) | Airport target detection method of high-resolution remote sensing image under complex background | |
CN110009010B (en) | Wide-width optical remote sensing target detection method based on interest area redetection | |
CN107690840B (en) | Unmanned plane vision auxiliary navigation method and system | |
Iqbal et al. | Weakly-supervised domain adaptation for built-up region segmentation in aerial and satellite imagery | |
CN110163213B (en) | Remote sensing image segmentation method based on disparity map and multi-scale depth network model | |
CN107507170A (en) | A kind of airfield runway crack detection method based on multi-scale image information fusion | |
Tan et al. | Vehicle detection in high resolution satellite remote sensing images based on deep learning | |
CN108846328A (en) | Lane detection method based on geometry regularization constraint | |
CN107767400A (en) | Remote sensing images sequence moving target detection method based on stratification significance analysis | |
CN113468968B (en) | Remote sensing image rotating target detection method based on non-anchor frame | |
CN109446894A (en) | The multispectral image change detecting method clustered based on probabilistic segmentation and Gaussian Mixture | |
Amo-Boateng et al. | Instance segmentation scheme for roofs in rural areas based on Mask R-CNN | |
de Gélis et al. | Siamese KPConv: 3D multiple change detection from raw point clouds using deep learning | |
CN109635726A (en) | A kind of landslide identification method based on the symmetrical multiple dimensioned pond of depth network integration | |
CN110084107A (en) | A kind of high-resolution remote sensing image method for extracting roads and device based on improvement MRF | |
Li et al. | Framework for unknown airport detection in broad areas supported by deep learning and geographic analysis | |
Gao et al. | Road extraction using a dual attention dilated-linknet based on satellite images and floating vehicle trajectory data | |
CN105631849B (en) | The change detecting method and device of target polygon |
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 |