CN106845343A - A kind of remote sensing image offshore platform automatic testing method - Google Patents
A kind of remote sensing image offshore platform automatic testing method Download PDFInfo
- Publication number
- CN106845343A CN106845343A CN201611161835.0A CN201611161835A CN106845343A CN 106845343 A CN106845343 A CN 106845343A CN 201611161835 A CN201611161835 A CN 201611161835A CN 106845343 A CN106845343 A CN 106845343A
- Authority
- CN
- China
- Prior art keywords
- remote sensing
- image
- random field
- convolutional neural
- neural networks
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of remote sensing image offshore platform automatic testing method, the remotely sensed image characteristic to offshore oil and gas platform multiresolution, multiple views first is analyzed, and optimal remote sensing images are fitted by proper subspace and principal component component.Then by spatial-temporal Markov random field model, the signal to noise ratio and signal to noise ratio of offshore platform detection are improved, exports target Saliency maps picture.Finally build large-scale target criteria data set, depth convolutional neural networks are carried out with the multi-node parallel framework based on hadoop distributed file storage systems to train, tuning is continued to depth convolutional neural networks, the stability characteristic (quality) of extensive target is extracted, offshore oil and gas platform fast and accurately automatic detection is realized in magnanimity remote sensing images.
Description
Technical field
The invention belongs to remote sensing image interpretation field, more particularly to a kind of remote sensing image offshore platform automatic detection side
Method.
Background technology
1.1 are based on the relatively-stationary detection method in offshore oil and gas platform geographical position
It is relatively fixed according to offshore oil and gas platform geographical position, detection is realized the characteristics of Ship Target is constantly moved.Such as
The SAR image data based on European Environment satellite such as Casadio S., detect naval target, according to target using CFAR methods
Location invariance extracts offshore oil and gas platform;Cheng L. etc. detect naval target using two-parameter CFAR, further according to fixation
The relative angular position invariance principle of target extracts offshore platform;Yongxue Liu etc. are using Landsat land imager
(OLI) multispectral data, based on context feature and position, the consistency of yardstick detect offshore oil and gas platform.
1.2 are based on offshore oil and gas platform infrared signature detection method
Equations of The Second Kind method realizes target detection by the infrared signature of offshore oil and gas platform.Most offshore oil
Gas platform can effectively detect these offshore platforms by burning waste gas of releasing flare by extracting fire point in infrared band.
Such as Elvidge defends OLS (Operational Line-scan System) sensing data of meteorological satellite by the U.S.
Carry out the natural gas fire point detection in the whole world;Casadio etc. is passed by ATSR (Along Track Scanning Radiometer)
Sensor data, radiation feature extraction is carried out to SAR image detection target, detects the offshore platform in Atlantic Ocean North Sea region;
The satellite data that Anejionu etc. passes through Landsat and MODIS, burning natural gas is detected using radiation filtering and space filtering
Offshore platform;Meng Ruolin etc. is chosen and sliding window using the multispectral data of Landsat TM sensors by optimal threshold
Method extracts offshore oil and gas platform.
Above-mentioned two classes offshore oil and gas detection of platform method is all that the image first to single phase carries out target detection, it is clear that
In this case, whole detection rate will be directly affected by Mono temporal image detection rate, and the detection mistake of a certain phase is just
Final missing inspection or false-alarm may be caused.And for offshore oil and gas platform, under Mono temporal, verification and measurement ratio is difficult to ensure, mainly because
For:1) offshore oil and gas platform yardstick is smaller, and its length and width only has generally at 100 meters or so in the remote sensing images of normal resolution
More than ten pixel size;2) offshore oil and gas platform imaging signal is weaker, it is difficult to substantially distinguished with background;3) maritime environment is complicated,
The noise jamming such as wave, cloud layer are serious;4) imaging features with naval vessel are similar, easily disturbed by Ship Target.
The content of the invention
When the technical problems to be solved by the invention are directed to the detection offshore oil and gas platform from magnanimity remote sensing image data,
The problem that target detection accuracy rate is low, workload is big for facing, it is proposed that when one kind is based on depth convolutional neural networks and three-dimensional
The offshore platform detection method of empty Markov random field.
The present invention uses following technical scheme to solve above-mentioned technical problem
A kind of remote sensing image offshore platform automatic testing method, specifically comprises the following steps:
Step 1, the remotely sensed image characteristic to offshore oil and gas platform multiresolution, multiple views is analyzed, by feature
Space and principal component component fitting remote sensing images;
Step 2, by spatial-temporal Markov random field model, improves the signal to noise ratio and signal to noise ratio of offshore platform detection, defeated
Go out target Saliency maps picture;
Step 3, builds target criteria data set, and depth convolutional neural networks are carried out to be deposited based on hadoop distributed documents
Depth convolutional neural networks are continued tuning by the multi-node parallel framework training of storage system, and the stabilization for extracting extensive target is special
Property, realizes in magnanimity remote sensing images offshore oil and gas platform fast and accurately automatic detection.
As the further preferred scheme of remote sensing image offshore platform automatic testing method of the present invention, the step 1
Specifically comprise the following steps:
Step 1.1, chooses the multidate remote sensing image collection under different resolution, multiple views;
Step 1.2, sets up ocean detection identification region priori database;
Step 1.3, is processed the input picture collection in step 1 by the priori database set up in step 2,
Non- sea area information in remotely-sensed data, the information in the deeper region in ocean that can not possibly carry out exploitation of offshore oil and gas are picked
Remove, the image set after being processed;
Step 1.4, the image set to being obtained in step 3 is manually demarcated, and obtains mark remote sensing images collection;
Step 1.5, proper subspace and principal component to marking remote sensing images collection carry out simulation analysis, obtain multidate distant
Sense image set.
As the further preferred scheme of remote sensing image offshore platform automatic testing method of the present invention, the step 2
Specifically comprise the following steps:
Step 2.1, temporal low-pass filter is carried out to the image in image set, initializes Markov random field;
Step 2.2, according to the Multitemporal Remote Sensing Images of input, sets up three-dimensional Markov random field model, and consider distant
Sense gradation of image distributed intelligence meets finite mixtures Gauss model, obtains combination condition probability;
Step 2.3, maximal possibility estimation is carried out according to the conditional probability in step 2.2, calculates Markov random field mould
Shape parameter;
Step 2.4, Markov random field is updated with the parameter calculated, and horse is obtained again through maximal possibility estimation
Er Kefu random field models parameters;
Whether step 2.5, the Markov random field model parameter error calculated twice before and after contrast meets convergence;If full
Foot, then perform step 2.6, otherwise continues executing with step 2.4;
Step 2.6, output target Saliency maps picture is calculated by three-dimensional space-time Markov random field model.
As the further preferred scheme of remote sensing image offshore platform automatic testing method of the present invention, the step 3
Specifically comprise the following steps:
Step 3.1, obtains platform Sample Storehouse, and then stores on hadoop distributed file storage systems, in k node
Upper storage Sample Storehouse subset, allows the independent training sample database of each node, initializes each parameter of convolutional neural networks;
Step 3.2, each node carry out simultaneously Sample Storehouse subset complete network training, obtain convolutional neural networks parameter to
Amount, and use the overall neural network parameter of neural network parameter fitting of each node;
Step 3.3, each node neural network parameter is updated with overall neural network parameter;
Step 3.4, man-machine interactively determines training effect, and persistence parameters tuning repeats step 3.2 to step 3.3, directly
Requirement is reached to training;
Step 3.5, adds target sample, repeat step 3.2 to step 3.4;
Step 3.6, using the convolutional neural networks trained in step 3.5 in step 2.6 export target conspicuousness
Image detected, exports object detection results.
As the further preferred scheme of remote sensing image offshore platform automatic testing method of the present invention, in step 1.2
In, longitude and latitude, ocean depth of the several databases comprising coastal waters.
The present invention uses above technical scheme compared with prior art, with following technique effect:
1) analysis method of optical remote sensing imaging characteristic of the offshore platform under multiresolution, multiple views is established, favorably
Target and false-alarm targets are distinguished in distinguishing;
2) the three-dimensional space-time Markov random field model of multidate is set up, effective output target Saliency maps picture shows
Write and suppress picture noise and noise jamming;
3) build based on the storage of hadoop distributed documents and train depth convolutional neural networks, greatly improve instruction
Practice efficiency, and reach the accuracy of offshore platform target detection higher.
Brief description of the drawings
Fig. 1 is offshore platform detection method block diagram;
Fig. 2 is three-dimensional Markov random field model;
Fig. 3 is convolutional neural networks Organization Chart;
Fig. 4 is convolutional neural networks parallel computation Organization Chart;
Fig. 5 is multi-temporal remote sensing data simulation analysis logic diagram;
Fig. 6 is three-dimensional Markov random field model modeling logic block diagram;
Fig. 7 is depth convolutional neural networks parallel training framework logic diagram.
Specific embodiment
Technical scheme is described in further detail below in conjunction with the accompanying drawings:
The technical solution adopted by the present invention is:Remotely sensed image first to offshore oil and gas platform multiresolution, multiple views is special
Property be analyzed, optimal remote sensing images are fitted by proper subspace and principal component component.Then by space-time markov with
Airport model, improves the signal to noise ratio and signal to noise ratio of offshore platform detection, exports target Saliency maps picture.Finally build large-scale
Depth convolutional neural networks are carried out the multi-node parallel based on hadoop distributed file storage systems by target criteria data set
Framework is trained, and tuning is continued to depth convolutional neural networks, extracts the stability characteristic (quality) of extensive target, realizes magnanimity remote sensing images
Middle offshore oil and gas platform fast and accurately automatic detection.
Step 1, the remotely sensed image characteristic to offshore oil and gas platform multiresolution, multiple views is analyzed, by feature
Space and principal component component fitting remote sensing images;
Step 2, by spatial-temporal Markov random field model, improves the signal to noise ratio and signal to noise ratio of offshore platform detection, defeated
Go out target Saliency maps picture;
Step 3, builds large-scale target criteria data set, and depth convolutional neural networks are carried out to be distributed based on hadoop
Depth convolutional neural networks are continued tuning by the multi-node parallel framework training of formula document storage system, extract extensive target
Stability characteristic (quality), realize in magnanimity remote sensing images offshore oil and gas platform fast and accurately automatic detection.
As the further preferred scheme of remote sensing image offshore platform automatic testing method of the present invention, the step 1
Specifically comprise the following steps:
Step 1.1, the multidate optical remote sensing figure under selected characteristic obvious a number of different resolution, multiple views
Image set A1;
Step 1.2, sets up ocean detection identification region priori database SQ on a large scale;
Step 1.3, is carried out by the priori database SQ set up in step 2 to the input picture collection A1 in step 1
Treatment, by the non-sea area information in remotely-sensed data, the information in the deeper region in ocean that can not possibly carry out exploitation of offshore oil and gas
Rejected, the image set A2 after being processed;
Step 1.4, the image set A2 to being obtained in step 3 is manually demarcated, and obtains mark remote sensing images collection A3;
Step 1.5, proper subspace and principal component to marking remote sensing images collection carry out simulation analysis, obtain multidate distant
Sense image set A4.
As the further preferred scheme of remote sensing image offshore platform automatic testing method of the present invention, the step 2
Specifically comprise the following steps:
Step 2.1, temporal low-pass filter is carried out to the image in image set A4, initializes Markov random field;
Step 2.2, according to the Multitemporal Remote Sensing Images of input, sets up three-dimensional Markov random field model, and consider distant
Sense gradation of image distributed intelligence meets finite mixtures Gauss model, obtains combination condition probability;
Step 2.3, maximal possibility estimation is carried out according to the conditional probability in step 2.2, calculates Markov random field mould
Shape parameter;
Step 2.4, Markov random field is updated with the parameter calculated, and horse is obtained again through maximal possibility estimation
Er Kefu random field models parameters;
Whether step 2.5, the Markov random field model parameter error calculated twice before and after contrast meets convergence;If full
Foot, then perform step 2.6, otherwise continues executing with step 2.4;
Step 2.6, output target Saliency maps picture is calculated by three-dimensional space-time Markov random field model.
As the further preferred scheme of remote sensing image offshore platform automatic testing method of the present invention, the step 3
Specifically comprise the following steps:
Step 3.1, obtains platform Sample Storehouse, stores on hadoop distributed file storage systems, is deposited on k node
Storage Sample Storehouse subset, allows the independent training sample database of each node, initializes each parameter of convolutional neural networks;
Step 3.2, each node carry out simultaneously Sample Storehouse subset complete network training, obtain convolutional neural networks parameter to
Amount, and use the overall neural network parameter of neural network parameter fitting of each node;
Step 3.3, each node neural network parameter is updated with overall neural network parameter;
Step 3.4, man-machine interactively determines training effect, and persistence parameters tuning repeats step 3.2 to step 3.3, directly
Requirement is reached to training;
Step 3.5, adds target sample, repeat step 3.2 to step 3.4;
Step 3.6, using the convolutional neural networks trained in step 3.5 in step 2.6 export target conspicuousness
Image detected, exports object detection results.
Fig. 1 is the agent structure block diagram of the inventive method, random based on depth convolutional neural networks and three-dimensional markov
The detection method of the remote sensing images offshore platform of field mainly includes three parts:Part I, offshore platform remotely sensed image characteristic
Analysis emulation and fitting;Part II;The three-dimensional space-time Markov random field modeling of multidate offshore platform;Part III,
Parallel architecture training and offshore platform based on depth convolutional neural networks are automatically extracted.The output of Part I is used as second
Partial input, the output of the second part as the detection of platform part in Part III input.
Fig. 2 is three-dimensional space-time Markov random field model figure, by the single frames remote sensing images of two-dimensional coordinate system according to when
Idle discharge is arranged, and constitutes 3-dimensional image model, and any pixel is present on the adjoint point and 18 time domains on 8 spatial domains in model
Adjoint point.
Fig. 3 is convolutional neural networks Organization Chart, and convolutional neural networks are mainly used to identification displacement, scaling and other forms and turn round
The X-Y scheme of bent consistency, convolutional neural networks are a neutral nets for multilayer, and every layer is made up of multiple two dimensional surfaces, and
Each plane is made up of multiple independent neurons.C1, S2, C3 and S4 are characteristic images in figure, and C1 and C3 is convolutional layer feature
Image, it is preferred that emphasis is the extraction of feature, S2 and S4 are sample level characteristic images, it is preferred that emphasis is the calculating of feature.
Fig. 4 is convolutional neural networks parallel computation Organization Chart.Substantial amounts of training sample is stored onto different memory nodes,
It is separate between sample, have the nerve net on complete convolutional neural networks, therefore each node on each memory node
Network is only responsible for training part sample set, and node processing is once updated after completing.Sample is in the training process, each in network
Interlayer is unidirectionally successively to carry out, and is separate with the characteristic pattern of layer, neuron and neuron weights, therefore there may be
The parallel computation in convolutional neural networks of multiple samples.
Fig. 5 is multi-temporal remote sensing data simulation analysis logic diagram, is first, according to input remotely-sensed data, to judge remote sensing number
According to geographical position, according to priori, set up remotely-sensed data geographic information database, reject non-in the remotely-sensed data of input
Pelagic division.Secondly the non-targeted such as artificial calibrating platform target and naval vessel, islands and reefs are carried out to the image after treatment.Then to distant
Sense image carries out proper subspace and principal component analysis, selects optimal band image.
Fig. 6 is three-dimensional space-time Markov random field model modeling logic block diagram, and the image to being input into carries out temporal low-pass
Filter to initialize Markov random field.The half-tone information of remote sensing images meets finite mixtures Gaussian Profile, obtains three-dimensional horse
The conditional probability of Er Kefu random field models, by maximal possibility estimation solving model parameter, more new model, when maximum likelihood is estimated
Collect and hold back, export target Saliency maps picture.
Fig. 7 is depth convolutional neural networks parallel training framework logic diagram, distributed storage is carried out to Sample Storehouse, each
The training of depth convolutional neural networks is carried out on memory node simultaneously, and is confirmed by man-machine interactively, model parameter is persistently adjusted
It is excellent, and add Sample Storehouse and be trained.When preferable convolutional neural networks are reached, the image to being input into quickly is detected,
Output testing result.
The following detailed description of technical scheme and the principles of science of institute's foundation.
1st, the principal component analysis principle in offshore platform remotely sensed image specificity analysis:
1. input n frame remote sensing image datas are assumed, the i-th two field picture is converted into length is the vector of L, and then sets up L × n
Two-dimensional image Matrix C={ c1,c2,…cn}。
2. in view of the data of remote sensing image are larger, it is assumed that n≤L, the average of two-dimensional matrix image C is
For ease of description, willIt is designated asIt is rightSingular value decomposition is carried out, be can obtain
Wherein, ∑ is diagonal matrix, its diagonal element σ1,…,σnIt isCharacteristic value, and element value
To arrange sequence, i.e. σ1≥σ2≥σ3≥,…,≥σn;U is an orthogonal matrix, and the column vector in U is's
σ in characteristic vector, with ∑1,…,σnThink correspondence;V is an orthogonal matrix, and the column vector in V is's
Characteristic vector.
3. for diagonal matrix sigma, when i values are more than a certain value, σ i value very littles now retain the preceding m row of U matrixes, obtain
New matrix Um, i.e. the principal component component data of remote sensing images.
2nd, remote sensing images space-time three-dimensional Markov random field model principle is as follows:
(1) Markov model principle
Markov random field looks like comprising Markov property and random field two-layer.Markov property refers to future
State is only relevant with current state, is worth according to random assign phase space one of certain distribution when to each position
Afterwards, its entirety is just called random field.
1. the finite point set that S={ (i, j) | 1≤i≤M, 1≤j≤N } represents MN positions, i.e. position in random field are set,
Λ={ 1,2,3 ..., L } represents the phase space in state space, i.e. random field, X={ xs| s ∈ S } represent be defined on's
Random field, xsRepresent on random field X, state space is the hidden state stochastic variable of Λ, i.e. xs∈Λ.In the picture, lattice point collection
S represents the position of pixel, and X is referred to as label, the usually set of pixel value, and Λ is label stochastic variable xsSet, L represents
Divide the image into the number for different zones.
2. the set of the general neighborhood system that δ={ δ (s) | s ∈ S } is defined on S is set, it meets following characteristic:
Then position r ∈ δ (s) is referred to as the adjoint point of s, and δ (s) is referred to as the adjoint point collection of s.In the present invention, according to the European of pixel
Distance definition neighborhood system:
δ(n)(s)=and r | d (s, r)≤n, r ≠ s } (3)
In formula (3), n is the order of neighborhood system, and d () is represented with Euclidean distance.ForMeet characteristicThere are different neighbour structures in S, the subset being made up of with its adjoint point single pixel or pixel on SReferred to as one son group, the collection of sub- group c shares C to represent.
3. it is the neighborhood system on S to set δ, as random field X={ xs| s ∈ S } meet following condition:
Then X is called the Markov random field with δ as neighborhood system.
(2) iterative estimate of markov random file parameter:
1. in order to determine the neighborhood local relation of label prior probability and label, set up Markov random field with
Gibbs distribution relations.Gibbs distributions meet following joint probability distribution form:
P (X=x)=(1/Z) exp [- U (x)] (5)
Wherein,Referred to as energy function, VcX () is son group only relevant with each pixel value in sub- group c
Potential function,Referred to as partition function.The condition of equivalence that Gibbs is distributed in MRF is:
2. image is split using Markov random field be exactly potential image label Correct out, with
Reach the posterior probability (MAP) of maximum:
3. the feature of remote sensing images using image half-tone information, the half-tone information of remote sensing images meets following limited mixed
Close Gauss model.
Wherein, the gray scale of image is divided into I region, ciIt is component weight, μiIt is average,It is variance, Γ () is
Gamma functions.
Finite mixtures Gaussian Profile probability function to gradation of image carries out maximal possibility estimation, and iterates to calculate, directly
Maximal possibility estimation error before and after twice is less than α (precision prescribed value), and then obtains maximum a posteriori probability, and then tries to achieve just
True image label, obtains target Saliency maps picture.
(3) three-dimensional space-time Markov random field model principle is as follows:
Do not consider elevation information, with reference to the multi-temporal data on the remote sensing image data and time-domain in spatial domain, set up
The time-space domain Markov random field model of Multitemporal Remote Sensing Images.Three-dimensional space-time Markov model is as shown in Figure 2:For two
, there is the adjoint point on the adjoint point and 18 time domains on 8 spatial domains in the point on bit image, Ma Er is described by Gibbs random fields
Can husband's field distribution.
In this method, the son group potential function for defining space-time neighborhood system is:
β is current spatial domain group parameter, βt-1And βt+1Group parameter.Set up joint probability distribution function:
P(f|xs)=Pt-1(f|xs)Pt(f|xs)Pt+1(f|xs) (10)
The method for using the maximal possibility estimation in (2), is iterated calculating, obtains maximum a posteriori probability, and then obtain
Correct images label.
3rd, the parallel architecture rapid extraction target theory of depth convolutional neural networks is as follows:
(1) convolutional neural networks (Convolutional Neural Networks, abbreviation CNN):
Convolutional neural networks are mainly used to recognize the X-Y scheme that displacement, scaling and other forms distort consistency, convolution
Neutral net is a neutral net for multilayer, and every layer is made up of multiple two dimensional surfaces, and each plane is by multiple independent nerves
Unit's composition.The general structure of convolutional neural networks is as shown in Figure 3.
The output node for rolling up basic unit is represented by:
Wherein:Current and last layer characteristic pattern is represented respectively,Represent from m-th characteristic pattern of last layer to
N-th characteristic pattern volume convolution kernel of current layer,For neuron is biased, f () is signal activation function.
The output of sub-sampling node layer is represented by:
In formula, s × s is sub-sampling template yardstick,It is template weights.
The result that convolutional neural networks CNN connects output layer entirely is represented by:
(2) the parallel architecture principle of depth convolutional neural networks:
Model training method based on deep learning, is typically necessary and expends huge time overhead, to improve training
Speed, using the multi-node parallel computing architecture stored based on hadoop distributed documents, as shown in Figure 4.Substantial amounts of training sample
This storage is separate between sample on different memory nodes, has complete convolutional Neural net on each memory node
Neutral net on network, therefore each node is only responsible for training part sample set, and node processing is once updated after completing.
Sample is unidirectionally successively to carry out in each interlayer of network, with the characteristic pattern of layer, neuron and neuron weights in the training process
It is separate, therefore there may be the parallel computation in convolutional neural networks of multiple samples.
Claims (5)
1. a kind of remote sensing image offshore platform automatic testing method, it is characterised in that:Specifically comprise the following steps:
Step 1, the remotely sensed image characteristic to offshore oil and gas platform multiresolution, multiple views is analyzed, by proper subspace
Remote sensing images are fitted with principal component component;
Step 2, by spatial-temporal Markov random field model, improves the signal to noise ratio and signal to noise ratio of offshore platform detection, exports mesh
Mark Saliency maps picture;
Step 3, builds target criteria data set, and depth convolutional neural networks are carried out to store system based on hadoop distributed documents
Depth convolutional neural networks are continued tuning by the multi-node parallel framework training of system, extract the stability characteristic (quality) of extensive target, real
Offshore oil and gas platform fast and accurately automatic detection in existing magnanimity remote sensing images.
2. remote sensing image offshore platform automatic testing method according to claim 1, it is characterised in that:The step 1
Specifically comprise the following steps:
Step 1.1, chooses the multidate remote sensing image collection under different resolution, multiple views;
Step 1.2, sets up ocean detection identification region priori database;
Step 1.3, is processed the input picture collection in step 1 by the priori database set up in step 2, will be distant
Non- sea area information in sense data, the information in the deeper region in ocean that can not possibly carry out exploitation of offshore oil and gas are rejected,
Image set after being processed;
Step 1.4, the image set to being obtained in step 3 is manually demarcated, and obtains mark remote sensing images collection;
Step 1.5, proper subspace and principal component to marking remote sensing images collection carry out simulation analysis, obtain multi-temporal remote sensing figure
Image set.
3. remote sensing image offshore platform automatic testing method according to claim 1, it is characterised in that:The step 2
Specifically comprise the following steps:
Step 2.1, temporal low-pass filter is carried out to the image in image set, initializes Markov random field;
Step 2.2, according to the Multitemporal Remote Sensing Images of input, sets up three-dimensional Markov random field model, and consider remote sensing figure
As grayscale distribution information meets finite mixtures Gauss model, combination condition probability is obtained;
Step 2.3, maximal possibility estimation is carried out according to the conditional probability in step 2.2, calculates Markov random field model ginseng
Number;
Step 2.4, Markov random field is updated with the parameter calculated, and Ma Erke is obtained again through maximal possibility estimation
Husband's random field models parameter;
Whether step 2.5, the Markov random field model parameter error calculated twice before and after contrast meets convergence;If meeting,
Step 2.6 is then performed, step 2.4 is otherwise continued executing with;
Step 2.6, output target Saliency maps picture is calculated by three-dimensional space-time Markov random field model.
4. remote sensing image offshore platform automatic testing method according to claim 1, it is characterised in that:The step 3
Specifically comprise the following steps:
Step 3.1, obtains platform Sample Storehouse, and then stores on hadoop distributed file storage systems, is deposited on k node
Storage Sample Storehouse subset, allows the independent training sample database of each node, initializes each parameter of convolutional neural networks;
Step 3.2, each node carries out the complete network training of Sample Storehouse subset simultaneously, obtains convolutional neural networks parameter vector,
And the neural network parameter using each node is fitted overall neural network parameter;
Step 3.3, each node neural network parameter is updated with overall neural network parameter;
Step 3.4, man-machine interactively determines training effect, and persistence parameters tuning repeats step 3.2 to step 3.3, Zhi Daoxun
It is experienced and worldly-wise to requiring;
Step 3.5, adds target sample, repeat step 3.2 to step 3.4;
Step 3.6, using the convolutional neural networks trained in step 3.5 in step 2.6 export target Saliency maps picture
Detected, exported object detection results.
5. remote sensing image offshore platform automatic testing method according to claim 1, it is characterised in that:In step 1.2
In, longitude and latitude, ocean depth of the several databases comprising coastal waters.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611161835.0A CN106845343B (en) | 2016-12-15 | 2016-12-15 | Automatic detection method for optical remote sensing image offshore platform |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611161835.0A CN106845343B (en) | 2016-12-15 | 2016-12-15 | Automatic detection method for optical remote sensing image offshore platform |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106845343A true CN106845343A (en) | 2017-06-13 |
CN106845343B CN106845343B (en) | 2020-05-19 |
Family
ID=59139361
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611161835.0A Active CN106845343B (en) | 2016-12-15 | 2016-12-15 | Automatic detection method for optical remote sensing image offshore platform |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106845343B (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108009469A (en) * | 2017-10-24 | 2018-05-08 | 中国科学院电子学研究所苏州研究院 | A kind of offshore oil and gas detection of platform method based on structure recurrent neural network |
CN108335222A (en) * | 2017-12-21 | 2018-07-27 | 中国石油天然气股份有限公司 | Marine site oil-gas exploration on-site investigation method and device based on multispectral remote sensing |
CN110823837A (en) * | 2019-11-19 | 2020-02-21 | 中国科学院遥感与数字地球研究所 | Method and device for simulating water leaving radiant quantity of ocean water body |
CN111028255A (en) * | 2018-10-10 | 2020-04-17 | 千寻位置网络有限公司 | Farmland area pre-screening method and device based on prior information and deep learning |
CN112070062A (en) * | 2020-09-23 | 2020-12-11 | 南京工业职业技术大学 | Hadoop-based crop waterlogging image classification detection and implementation method |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102768356A (en) * | 2012-08-03 | 2012-11-07 | 南京大学 | Method for detecting ocean stationary targets by multi-temporal radar data |
CN103544528A (en) * | 2013-11-15 | 2014-01-29 | 南京大学 | BP neural-network classification method based on Hadoop |
US8787638B2 (en) * | 2011-04-07 | 2014-07-22 | The Chinese University Of Hong Kong | Method and device for retinal image analysis |
-
2016
- 2016-12-15 CN CN201611161835.0A patent/CN106845343B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8787638B2 (en) * | 2011-04-07 | 2014-07-22 | The Chinese University Of Hong Kong | Method and device for retinal image analysis |
CN102768356A (en) * | 2012-08-03 | 2012-11-07 | 南京大学 | Method for detecting ocean stationary targets by multi-temporal radar data |
CN103544528A (en) * | 2013-11-15 | 2014-01-29 | 南京大学 | BP neural-network classification method based on Hadoop |
Non-Patent Citations (1)
Title |
---|
JIANGQIANG REN等: "Exploiting Global Priors for RGB-D Saliency Detection", 《2015 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION WORKSHOPS》 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108009469A (en) * | 2017-10-24 | 2018-05-08 | 中国科学院电子学研究所苏州研究院 | A kind of offshore oil and gas detection of platform method based on structure recurrent neural network |
CN108009469B (en) * | 2017-10-24 | 2020-11-24 | 中国科学院电子学研究所苏州研究院 | Offshore oil and gas platform detection method based on structure recurrent neural network |
CN108335222A (en) * | 2017-12-21 | 2018-07-27 | 中国石油天然气股份有限公司 | Marine site oil-gas exploration on-site investigation method and device based on multispectral remote sensing |
CN111028255A (en) * | 2018-10-10 | 2020-04-17 | 千寻位置网络有限公司 | Farmland area pre-screening method and device based on prior information and deep learning |
CN111028255B (en) * | 2018-10-10 | 2023-07-21 | 千寻位置网络有限公司 | Farmland area pre-screening method and device based on priori information and deep learning |
CN110823837A (en) * | 2019-11-19 | 2020-02-21 | 中国科学院遥感与数字地球研究所 | Method and device for simulating water leaving radiant quantity of ocean water body |
CN112070062A (en) * | 2020-09-23 | 2020-12-11 | 南京工业职业技术大学 | Hadoop-based crop waterlogging image classification detection and implementation method |
Also Published As
Publication number | Publication date |
---|---|
CN106845343B (en) | 2020-05-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
EP3614308B1 (en) | Joint deep learning for land cover and land use classification | |
Roelfsema et al. | Multi-temporal mapping of seagrass cover, species and biomass: A semi-automated object based image analysis approach | |
CN106845343A (en) | A kind of remote sensing image offshore platform automatic testing method | |
CN103413151B (en) | Hyperspectral image classification method based on figure canonical low-rank representation Dimensionality Reduction | |
Ai et al. | Convolutional neural network to retrieve water depth in marine shallow water area from remote sensing images | |
Chen et al. | CycleGAN-STF: Spatiotemporal fusion via CycleGAN-based image generation | |
Mahmon et al. | A review on classification of satellite image using Artificial Neural Network (ANN) | |
Kremezi et al. | Pansharpening PRISMA data for marine plastic litter detection using plastic indexes | |
CN103871039B (en) | Generation method for difference chart in SAR (Synthetic Aperture Radar) image change detection | |
CN107967474A (en) | A kind of sea-surface target conspicuousness detection method based on convolutional neural networks | |
Jeon et al. | Semantic segmentation of seagrass habitat from drone imagery based on deep learning: A comparative study | |
CN106683102A (en) | SAR image segmentation method based on ridgelet filters and convolution structure model | |
CN108932455B (en) | Remote sensing image scene recognition method and device | |
CN109446894A (en) | The multispectral image change detecting method clustered based on probabilistic segmentation and Gaussian Mixture | |
Chen et al. | SwinSTFM: Remote sensing spatiotemporal fusion using Swin transformer | |
Liang et al. | Maximum likelihood classification of soil remote sensing image based on deep learning | |
Bramhe et al. | Extraction of built-up areas using convolutional neural networks and transfer learning from sentinel-2 satellite images | |
CN111178304A (en) | High-resolution remote sensing image pixel level interpretation method based on full convolution neural network | |
CN116563726A (en) | Remote sensing image ship target detection method based on convolutional neural network | |
CN105469428B (en) | A kind of detection method of small target based on morphologic filtering and SVD | |
Yang et al. | Toward country scale building detection with convolutional neural network using aerial images | |
Gao et al. | Road extraction using a dual attention dilated-linknet based on satellite images and floating vehicle trajectory data | |
CN107967454A (en) | Take the two-way convolutional neural networks Classification in Remote Sensing Image method of spatial neighborhood relation into account | |
Liu et al. | Contour-oriented cropland extraction from high resolution remote sensing imagery using richer convolution features network | |
CN117422619A (en) | Training method of image reconstruction model, image reconstruction method, device and equipment |
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 |