CN108460739A - A kind of thin cloud in remote sensing image minimizing technology based on generation confrontation network - Google Patents
A kind of thin cloud in remote sensing image minimizing technology based on generation confrontation network Download PDFInfo
- Publication number
- CN108460739A CN108460739A CN201810174430.3A CN201810174430A CN108460739A CN 108460739 A CN108460739 A CN 108460739A CN 201810174430 A CN201810174430 A CN 201810174430A CN 108460739 A CN108460739 A CN 108460739A
- Authority
- CN
- China
- Prior art keywords
- network
- image
- thin cloud
- cloud
- remote sensing
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000005516 engineering process Methods 0.000 title claims abstract description 18
- 230000006870 function Effects 0.000 claims abstract description 50
- 238000000034 method Methods 0.000 claims abstract description 29
- 238000009826 distribution Methods 0.000 claims abstract description 17
- 238000013461 design Methods 0.000 claims abstract description 15
- 238000013528 artificial neural network Methods 0.000 claims abstract description 4
- 238000012549 training Methods 0.000 claims description 46
- 230000004069 differentiation Effects 0.000 claims description 15
- 230000004913 activation Effects 0.000 claims description 13
- 230000004044 response Effects 0.000 claims description 13
- 230000015572 biosynthetic process Effects 0.000 claims description 11
- 238000003786 synthesis reaction Methods 0.000 claims description 11
- 238000005457 optimization Methods 0.000 claims description 10
- 238000003384 imaging method Methods 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 9
- 230000005540 biological transmission Effects 0.000 claims description 8
- 238000013507 mapping Methods 0.000 claims description 8
- 230000000644 propagated effect Effects 0.000 claims description 8
- 238000000605 extraction Methods 0.000 claims description 6
- 230000005855 radiation Effects 0.000 claims description 6
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 4
- 238000004422 calculation algorithm Methods 0.000 claims description 4
- 230000000694 effects Effects 0.000 claims description 4
- VMXUWOKSQNHOCA-UKTHLTGXSA-N ranitidine Chemical compound [O-][N+](=O)\C=C(/NC)NCCSCC1=CC=C(CN(C)C)O1 VMXUWOKSQNHOCA-UKTHLTGXSA-N 0.000 claims description 4
- 238000005070 sampling Methods 0.000 claims description 4
- 238000004088 simulation Methods 0.000 claims description 4
- 238000004040 coloring Methods 0.000 claims description 3
- 238000013527 convolutional neural network Methods 0.000 claims description 3
- 238000005520 cutting process Methods 0.000 claims description 3
- 238000003475 lamination Methods 0.000 claims description 3
- 239000003643 water by type Substances 0.000 claims description 3
- 238000004458 analytical method Methods 0.000 claims description 2
- 238000001514 detection method Methods 0.000 claims description 2
- 238000004880 explosion Methods 0.000 claims description 2
- 239000000284 extract Substances 0.000 claims description 2
- 238000013178 mathematical model Methods 0.000 claims description 2
- 238000003062 neural network model Methods 0.000 claims description 2
- 238000012216 screening Methods 0.000 claims description 2
- 238000000179 transient infrared spectroscopy Methods 0.000 claims description 2
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims 1
- 230000007423 decrease Effects 0.000 claims 1
- 230000008034 disappearance Effects 0.000 claims 1
- 238000013135 deep learning Methods 0.000 description 4
- 238000004519 manufacturing process Methods 0.000 description 4
- 230000002123 temporal effect Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 239000000203 mixture Substances 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 230000009897 systematic effect Effects 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- VAYOSLLFUXYJDT-RDTXWAMCSA-N Lysergic acid diethylamide Chemical compound C1=CC(C=2[C@H](N(C)C[C@@H](C=2)C(=O)N(CC)CC)C2)=C3C2=CNC3=C1 VAYOSLLFUXYJDT-RDTXWAMCSA-N 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000006854 communication Effects 0.000 description 1
- 238000009833 condensation Methods 0.000 description 1
- 230000005494 condensation Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 239000013078 crystal Substances 0.000 description 1
- 238000013079 data visualisation Methods 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
- 230000002269 spontaneous effect Effects 0.000 description 1
- 230000002194 synthesizing effect Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention relates to a kind of based on the thin cloud in remote sensing image minimizing technology for generating confrontation network, includes the following steps：Step 1：Establish thin cloud removal system model；Step 2：Network model designs；Step 3：Recognize criterion function structure；Step 4：Bao Yun removes System Discrimination；Step 5：Thin cloud in remote sensing image removes.The present invention removes problem using confrontation type neural network is generated to model thin cloud, and carries out System Discrimination to it, it can be achieved that thin cloud removal end to end.Constructed criterion function combines error of both data distribution and reconstruction precision so that system is capable of the feature of preferably learning data, to realize the removal of thin cloud.There is method proposed by the present invention the ability for adaptively removing uneven thin cloud, the image restored to have good color and texture homogeneity.
Description
Technical field：
The present invention relates to a kind of based on the thin cloud in remote sensing image minimizing technology for generating confrontation network, belongs to remote sensing image processing
Applied technical field.
Background technology：
The tasks such as remote sensing images are geoscience, meteorology, environmental monitoring, military monitoring provide abundant information,
Through as a kind of tool more and more important in modernization observing and controlling means.However, there are a large amount of steam, ice crystal and micronic dusts in air
The nuclei of condensation, they are assembled in the form of cloud.And electromagnetic wave is highly prone to the influence of cloud in communication process, and signal is caused to be reflected
Or scattering so that the signal that sensor receives is weakened, and influences image quality, and interested area information is caused to be lost, after giving
The interpretation and interpretation of continuous remote sensing images cause very big difficulty.
One effective thin cloud minimizing technology can improve picture quality under limited image-forming condition, restore image letter
Breath contributes to subsequent interpretation and application, improves the utilization rate to remote sensing images.Currently, the thin cloud removal of many remote sensing images
Method has been suggested.These method one kind are to be based on multi-temporal remote sensing data, when by detecting cloud sector and alternative more utilization
Phase data carries out data replacement.But since sensor imaging has intervals, atmospheric radiation condition and geomorphologic conditions that can send out
Changing, and it is of a high price for the imaging of the fixed point of target location for a long time, therefore the ideal multi-temporal data of areal is not
Easily obtain.Another kind of method is based on existing Mono temporal data and is directly handled, and processing method is mostly based on simplified imaging model
And priori, but since the type of cloud is changeable, complicated to the response characteristic of wavelength, an effective model is not easy to establish, and algorithm exists
It is showed in various non-uniform thin cloud removals not good enough.In recent years, with the development of deep learning, the method based on learning model
It is suggested.Existing method is all based on a convolutional neural networks end to end, and directly study has cloud atlas picture reflecting to cloudless image
It penetrates.This discriminative model is the recurrence to available data, cannot acquire the true distribution character of data, therefore extensive energy well
Power is not strong, performs poor in complex scene and various thin clouds.
The present invention is devised and a kind of is gone based on the thin cloud for generating confrontation network for thin cloud present in single width remote sensing images
Except method.It is a kind of production model in deep learning to generate confrontation network, and network and a differentiation network are generated by one
Composition.Traditional discriminate deep learning method is input to the mapping relations of output by the study of depth convolutional network；And it generates
One binary minimax game function of Web vector graphic is fought as object function, essence has been measured Jason-Shannon and dissipated
Degree, is the measurement to distribution similarity so that network is capable of the distribution of learning data rather than simple mapping relations.The present invention
The thin cloud removal of image is regarded to the identification problem of nonparametric model system as, design condition production confrontation network establishes direct die
Type, the object function to weigh distribution similarity learn condition distribution of the cloudless data in the case where there is cloud data as criterion function,
To which the thin cloud for obtaining satisfied goes division result.The system can be restored to adaptively being removed to uneven thin cloud
Image has good color consistency and structural integrity.
Invention content：
1, purpose：
The purpose of the present invention is to provide a kind of based on the thin cloud in remote sensing image minimizing technology for generating confrontation network, is used for real
The thin cloud removal of existing Mono temporal remote sensing images, promotes picture quality.This method is by using a certain amount of remote sensing images sample training
It generates confrontation network and carries out System Discrimination, obtained network model is suitable for a variety of satellite sensor images, can be to Mono temporal
Remote sensing images realize good thin cloud removal.
2, technical solution：The invention is realized by the following technical scheme.
The present invention carries out system modelling to thin cloud removal problem first, then designs one and generates confrontation network end to end
System Discrimination is carried out as identification system.Due to by the true value figure corresponding to cloud block image, i.e., same place it is cloudless true
Value is not easy to obtain, therefore uses simulation means, obtains enough samples with true value label to generating network and differentiating that network is instructed
Practice so that generate the spontaneous extraction feature learning data distribution of network, the identification of system is carried out according to the criterion function of design.Network
After the completion of training, obtained generation model can realize the cloud removing to input picture.The invention includes 5 steps：It establishes thin
Cloud removes system total model, network model designs, identification criterion function structure, Bao Yun remove System Discrimination, remote sensing images are thin
Cloud removes.It is specific as follows：
Step 1：Establish thin cloud removal system total model
The present invention carries out thin cloud removal for the remote sensing images that the land Landsat-8 OLI imager is acquired.
Landsat-8 is the satellite that the U.S. emitted on 2 11st, 2013, carries the land OLI imager and TIRS thermal infrared sensors
Device, the land the OLI imager of Landsat-8 include 9 wave bands, wherein the 4th, 3, that 2 wave bands have corresponded to red, green, blue three is visible
Optical band.We select the 4th, 3,2 wave bands obtain True color synthesis image, carry out thin cloud removal.
Bao Yun imaging distortion models can be described as：
S (i, j)=aLr (i, j) t (i, j)+L (1-t (i, j)) (1)
Wherein, s (i, j) is the signal that sensor receives at point (i, j), and L is air light radiation, and r (i, j) is ground
True reflection namely our desired cloudless images, t (i, j) is transmission plot, and a is air light attenuation coefficient, between 0~1 model
In enclosing.
According to the imaging distortion model, when known to transmission plot, cloudless clear image can be established and had between cloud atlas picture
Linear relationship.Enable what y was denoted as condition entry to have a cloud atlas picture, z represents the noise introduced,Indicate the cloudless image restored,
G, which is represented, generates model, i.e., thin cloud goes division operation, then thin cloud removal system can be indicated with following simplification mathematical model：
System model is removed according to the thin cloud that formula (2) is established, as long as we obtain parameter by the network identification of design
The generation model g of change then can have cloud atlas picture to predict cloudless clear image by functional relation from input, with realization pair
The thin cloud of remote sensing images removes.
Step 2：Network model designs
The present invention fights neural network using one group of generation and is recognized to thin cloud removal system model.Used in identification system
Network is made of two sub- networks：It generates network G and differentiates network D.It generates network G and predicts cloudless image using input information,
Receiving has cloud atlas picture and noise, is driven and is trained with criterion function so that the cloudless image of generation, which can allow, differentiates that network D differentiates
It is true.Differentiate that the effect of network D is to differentiate that auxiliary generates network G to the cloudless image of generation and true cloudless image
Training.Two sub-networks are contended with binary minimax game loss, in the process, it is true to generate network G study
The condition of data is distributed.By the confrontation study of two sub-networks, identified parameters are obtained.After the completion of identification, network G is generated
There is the alternative model of identical input-output characteristic as model g is generated with thin cloud removal system.Thin to remote sensing images progress
When cloud goes division operation, generates network G and carry out a propagated forward, can be calculated by the parameter picked out and recover no cloud atlas
Picture is not required to again by differentiation network D.
1) design of network G is generated：
The generation network G that the present invention designs is the depth convolutional neural networks being made of several residual error study modules,
The fine fitting to nonlinear model can be realized by several convolution-standardization-nonlinear activation operation, pass through identical mapping
Constitute the study of residual error structure.The local module can be described as：
Wherein,Indicate that obtaining n-th layer i-th by (n-1) layer jth characteristic pattern opens convolution kernel used in characteristic pattern,Indicate i-th characteristic pattern of network n-th layer,Indicate the jth characteristic pattern of network (n-1) layer,It is the i-th of n-th layer
Open the bias vector of characteristic pattern.NnFor the standardized operation to n-th layer output response, ΦnFor to the non-thread of n-th layer output response
Property activation operation.The identical mapping that characteristic pattern opens to n-th layer i-th characteristic pattern is opened for (n-1) layer i-th, this structure to roll up
Residual error of the lamination mainly between input and desired output learns so that fitting is more easy, to reduce generation network
Training difficulty.Designed generation network structure is：
DownsampleBlock×3→ResBlock×6→UpsampleBlock×3
DownsampleBlock:Conv (4 × 4, stride=2) → BN → LeakyRelu
UpsampleBlock:TransposedConv (4 × 4, stride=2) → BN → LeakyRelu
Wherein, Conv represents convolutional layer, and TransposedConv is warp lamination, and numerical value is is used convolution kernel in bracket
Size and step-length, herein use 4 × 4 sizes convolution kernel, and step-length be 2；DownsampleBlock is down-sampled module,
The dimension of image is reduced, the parameter amount learnt required for network is reduced；ResBlock is shown in formula (3) with identical mapping
Convolution-standardization-nonlinear activation layer composition, also uses the convolution kernel of 4 × 4 sizes, and 6 residual errors are employed herein
Block.UpsampleBlock is up-sampling module, corresponding DownsampleBlock modules, for restoring image original size.It adopts
Training parameter amount is on the one hand reduced with smaller convolution kernel, on the other hand so that the reconstruction process of image is reinforced to local message
Utilization.With doing and standardizing to response output there are one standardization layer (BN) after each convolutional layer, make its mean value and variance one
It causes, avoid gradient explosion local in training or disappears.It is a nonlinear activation layer, nonlinear activation function after standardization layer
Using LeakyRelu functions.
2) differentiate the design of network D：
When differentiating to generation image and true picture, differentiate that network D needs simultaneously from the sense of reality for generating image
And whether thin cloud in terms of effectively removing two differentiate.It is local that this just needs differentiation network D not to be concerned only with
Textural characteristics judge the continuity at edge and texture, the sharp keen degree of texture, while also will be to the spy of image entirety
Sign, such as brightness, whether the whole visual quality of contrast and colouring information declines is judged.Therefore, the present invention devises one
A differentiation network differentiated in conjunction with Analysis On Multi-scale Features, structure are as follows：
DownsampleBlock×3→MultiscaleBlock×3
Wherein, down-sampled module is structurally and functionally identical with generation network G；Each MultiscaleBlock is
One Multi resolution feature extraction module is made of three branches, and each branch uses various sizes of convolution kernel, to obtain difference
The receptive field of size obtains the feature of image, i.e., from different scales：
MultiscaleBlock=[conv (1 × 1), conv (3 × 3), conv (5 × 5)]
Each convolutional layer use step-length for 1 convolution operation, and edge filling appropriate is carried out, so that exporting feature
Figure size is consistent.Then, it connects to the activation output that each branch obtains, obtains the feature of different scale.With life
Identical at network G, there are one output standardization layer and a nonlinear activation layers for connection after each convolutional layer.
Step 3：Recognize criterion function structure
Model g input-output characteristics having the same are generated to make generation network G and thin cloud remove system, it is also necessary to the two
Training supervisory signals of the error as network.In the present invention, this signal is by the output and supervision sample of differentiation network D to giving
Go out.Correspondingly, criterion function is also made of two parts.
First part, data distribution criterion function.This part is to differentiating that the output result of network D calculates true and false two classification
Cross entropy is as criterion function.For differentiating for network, needs to maximize criterion function, by cloudless true value figure and generate net
The cloudless image strict differences that network restores are opened；For generating for network G, need to minimize this criterion function, with " deception "
Differentiate network, differentiation network is made to restore the classification that image carries out mistake to it.Ideally, differentiate network to the true of input
It is 0.5 that image and generation image discriminating, which are genuine probability,.Therefore, this criterion function can be write：
J1=E [log D (x)]+E [log (1-D (G (y, z)))] (4)
Wherein, x is true cloudless image, and y is represented has cloud atlas picture, z to represent the noise introduced, D as condition entry
() and G () respectively represent the output for differentiating network and generating network, and E () indicates each batch sample to being inputted when training
Mathematic expectaion is sought in corresponding output.Pass through the two minimax alternative optimization so that generating network can be indirectly by differentiation net
Network extracts high frequency detail feature and global characteristics.
Second part, response error function.For input sample, structure quadratic form error function is as follows：
Generate the Euclidean distance between prognostic chart and cloudless true value figure.Subscript k indicates to carry out in small batches to network in formula
The serial number of sample in each small lot sample is taken when amount stochastic gradient descent training.xk, ykAnd zkRespectively represent kth group without
Cloud sample, the noise for having cloud sample and this group of input sample being introduced.Network is generated to obtain by minimizing this loss function
The smooth characteristics of low-frequency such as average colouring information is taken, regression forecasting is done to image.
Therefore, the criterion function of final design of the present invention is：
J=J1+J2 (6)
Step 4：Bao Yun removes System Discrimination
1) training sample obtains
The identification that thin cloud removal system generates model g is carried out using confrontation type neural network model is generated, is a supervision
The process of study needs the data of tape label to generating network and differentiating that network is trained.However remote sensing satellite is to same
The return visit of point has some cycles interval, and atmospheric radiation, geomorphic feature can great changes will take place within the time period, and for a long time
The cost for pinpointing shooting is again very high, so ideal have in pairs cloud atlas picture and its cloudless true value figure (label) is difficult to obtain
.Therefore, the means of the thin cloud minimizing technology generally use emulation based on supervised learning obtain enough training samples.Together
Sample, the present invention in clear image using the method for emulating thin cloud is generated, and using clear image as cloudless true value figure, obtains band mark
Sign data.
1. clear cloudless true value figure obtains
Cloudless clear image is chosen from remote sensing satellite image, choosing image should covering farm land, forest, plant as much as possible
The various landforms such as quilt, mountain area, waters, city, to establish feature remote sensing images tranining database complete as possible so that most
The thin cloud removal system obtained eventually can be widely used in the true color image acquired in various satellite sensors.Due to general
Remote sensing images picture is all bigger, it is also necessary to carry out cutting the operations such as screening to image.To original multi-spectral remote sensing image, we
It opens it in ENVI softwares, selection the 4th, 3,2 wave bands obtain True color synthesis image, and manually fixed greatly with 256 × 256
It is small that image is cut and preserved.
2. thin cloud image simulation synthesis
The thin cloud atlas that the present invention uses is imitative into racking using Landsat-8 satellite bands 9 (cirrus wave band) as emulation mode
Very.For the wavelength of Landsat-8 satellite OLI sensor cirrus wave bands at 1.36~1.38 μm, high-altitude steam has compared with strong reflection it
Characteristic, low latitude atural object and dry environment reflection are very weak, therefore cloud layer responds highly significant in the wave band, and translucent white is presented,
Atural object response is faint, and large stretch of black is presented in the picture, is commonly used for the tasks such as cloud detection.Using the wave band, extraction cloud layer is anti-
Gray-scale map is penetrated, as transmission plot t, is brought into the imaging distortion model of formula (1), and combines the atmosphere light of different level, you can
It obtains synthesizing thin cloud remote sensing images.
2) identification system network training
The present invention declines thought to carry out parametric solution to identification system using gradient, and specific optimization is calculated using Adam optimizations
Method.Using the pairs of sample that emulation obtains as the input of identification system and ideal output, to generating network G and differentiating network
D is iterated training, and constantly updates the parameter of two sub-networks, when criterion function loss tends towards stability, two network instructions
White silk finishes.
Obtained generation network G is to generate model g with the thin cloud in remote sensing image removal system described in formula (2) to have phase
With the alternative model of input-output characteristic, identification system model training finishes namely determines the parameter for generating network G, so far,
Bao Yun removes System Discrimination and completes.
Step 5：Thin cloud in remote sensing image removes
After the completion of the identification training process of step 4, network model is provided with identical defeated with practical thin cloud removal system
Enter output characteristics.The remote sensing images for having Bao Yun for a web are inputted and generate a network G propagated forward of progress, utilize
Trained parameter carries out operation, you can exports the cloudless image being resumed.
3, advantage and effect
The present invention devises one and removes system based on the thin cloud in remote sensing image for generating confrontation network, passes through a generation net
Network carries out Forward modeling to system, and the identification of confrontation network (differentiating a network) auxiliary system is used in combination, thin to obtain having
The network of cloud removal ability.This method will have cloud atlas picture as mode input, and according to data distribution criterion function and quadratic form
Response error criterion function error signal trains generation by backpropagation and differentiates two sub-networks so that is final
Model have the input and output response characteristic consistent with real system, completion System Discrimination.After the completion of identification, obtained model
It can be to there is cloud remote sensing images to carry out thin cloud removal.The present invention has the following advantages：
1. the present invention is simplified the artificial mathematical derivation that carries out and is modeled the complicated calculations brought with a model end to end,
Greatly reduce the identification difficulty of the Complex Nonlinear System of Bao Yun removal problems；
2. the thin cloud removal model designed by the present invention is a production model.The deep learning method of traditional discriminate
Simple study maps end to end, is unable to the true distribution situation of learning data, the interpretation of model is poor, and generalization ability is opposite
It is weaker.And the present invention is removed thin cloud using production model, designed criterion function combines data distribution and again
Error of both precision is built, uses a differentiation network-assisted system identification process so that identification system can be learned preferably
Practise the distribution of the feature and data of image.
3. the removal effect to thin cloud is good, shows more remarkably, outclass on uniform non-uniform thin cloud atlas picture
Method of the tradition based on simplified model or priori；Compared to the method for conventional depth learning method, i.e. discriminate, can handle more
Uniform uneven thin cloud under more complex scenes has the ability for adaptively removing thin cloud, and restoring image has good color
Color and texture homogeneity；
4. system has good robustness and universality.Though training set used herein is based on Landsat-8 remote sensing
Image data is established, but shows to be equally applicable to a variety of satellite remote sensings, the boats such as high score No.1, Google Earth by verification
Clap class image；The method of the present invention can also be generalized in the thin cloud removal task of multi-spectral remote sensing image, only need to be according to step 4 weight
It is new to prepare training dataset, and the port number of network inputs is changed to selected wave band number, model is trained again,
More multiwave thin cloud can be realized and go division operation.
Description of the drawings
Fig. 1 is that thin cloud of the present invention removes systematic schematic diagram.
Fig. 2 a, Fig. 2 b are network architecture schematic diagram designed by the present invention.
Fig. 3 a, Fig. 3 b are that the present invention emulates thin cloud remote sensing images example.
Fig. 4 a-1, Fig. 4 a-2, Fig. 4 b-1, Fig. 4 b-2, Fig. 4 c-1, Fig. 4 c-2 are the thin cloud on different remote sensing satellite images
Division result is gone to show.
Specific implementation mode
Technical solution for a better understanding of the present invention is made embodiments of the present invention below in conjunction with attached drawing further
Description：
The present invention systematic schematic diagram as shown in Figure 1, designed network structure as shown in Fig. 2 a, Fig. 2 b.Computer is matched
Set use：Intel Core i7-4709k processors, Nvidia GeForce GTX1080Ti graphics processors, dominant frequency
4.0GHz, memory 16GB, operating system are ubuntu 14.04.The training of network model is based on Pytorch frames.The present invention is
A kind of thin cloud in remote sensing image minimizing technology based on generation confrontation network, specifically includes following steps：
Step 1：Establish thin cloud removal system model
The present invention carries out thin cloud removal using the remote sensing images that the land Landsat-8 OLI imager is acquired.
The land the OLI imager of Landsat-8 includes 9 wave bands, selects the 4th, 3,2 wave band (red wave band, green wave band, Lan Bo therein
Section) synthesis obtains true color image, and it carries out follow-up thin cloud and goes division operation.
Enable what y represented input system to have a cloud atlas picture, z represents system noise and the measurement noise of introducing, and g, which is represented, generates mould
Type goes cloud to operate,The clear cloudless image restored is represented, then thin cloud removal problem can be modeled as following form：
The system model established according to above formula fights network identification by generation and obtains the equivalent model G of model g
Afterwards, using the functional relation, input picture is converted, there can be cloud atlas picture to predict clear cloudless image from input,
It realizes and the thin cloud of Mono temporal remote sensing images is removed.
Step 2：Network model designs
Fig. 1 gives the present invention the general frame of thin cloud minimizing technology.System according to Fig. 1 writes entire project
All Files.It is broadly divided into following function module file：Data loader, model file, optimization module, training daily record note
Record and some visualization tools etc..
1. data loader：Stochastical sampling a batch paired data from data set carries out the pre- place of mean variance standardization
Reason, the input as model and desired output.Upset data again before carrying out the sampling of a new round to entire data set, ensures to adopt
Sample has randomness.
2. model file：Model file, that is, network model is divided into the positive differentiation net for generating network model and auxiliary identification
Network.Network structure shown in a, Fig. 2 b establishes the calculating figure of whole network according to fig. 2, and when defining forward-propagating data flow flowing
Mode.
3. optimization module：Define the costing bio disturbance of criterion function, optimizer optimisation strategy, training method.
4. tool model：Including training log recording, network model and parameter preserve, network heavy duty tool (is used for breakpoint
Training or small parameter perturbations), data type conversion tool and data visualization tool etc..
Step 3：Recognize criterion function structure
Designed criterion function is two-part adduction：Data distribution difference function and quadratic form reconstructed error function.
For a forward direction operation, criterion function loss, which is calculate by the following formula, to obtain：
For one group of input training image sample, the clear image of operation output prediction is carried out by generation network, differentiates net
Network carries out one group of differentiation to prognostic chart picture and desired image, calculates distributional difference loss；Prognostic chart picture is calculated simultaneously and it is expected defeated
The Euclidean distance gone out, i.e. reconstructed error lose, and sum of the two is under current network parameter according to the calculated damage of criterion function
Mistake value.
Step 4：Bao Yun removes System Discrimination
1) training sample obtains
This method obtains pairs of thin cloud atlas picture and true value image when building training dataset by the way of emulation.
Concrete operations are：
1. obtain clearly cloudless remote sensing images as true value figure.From the official websites USGS (United States Geological Survey, United
States Geological Survey,https://earthexplorer.usgs.gov/) download free Landsat-8
Remote sensing image data.The cloudless clear image of more scapes is chosen, radiation calibration and atmospheric correction are carried out to it in 5.1 softwares of ENVI.
Taking the 4th, 3,2, (blue, green, red) wave band carries out True color synthesis, obtains panorama clear image.It is cut and is screened again,
So that data set includes a variety of landforms such as vegetation, farmland, waters, desert, mountain area, city as possible.This step is obtained 256 ×
The original image 2000 of 256 sizes is opened.
2. thin cloud image simulation synthesis.Suitable image is found in the official websites USGS again, this purpose is the extraction (volume of wave band 9
Cloud wave band) synthesis of artificial the image with thin cloud, therefore in the area that atural object is more single, interference of texture is less, such as grassland, sea
Ocean is chosen with the moderate more scape images of large stretch of thin cloud, cloud amount.The image of download ENVI softwares are loaded into, wave band 9 is independent
Gray level image is saved as, the cutting of 256 × 256 sizes is equally carried out, obtains 2000 emulation transmission plots.Will emulation transmission plot and
Formula (1) is brought in clear image into, adjusts different air light attenuation coefficients, and 8000 thin cloud remote sensing images of emulation are always obtained,
With true value figure at corresponding, totally 8000 pairs of band true value label images pair.
Fig. 3 a, Fig. 3 b illustrate two thin cloud remote sensing images obtained using emulation mode of the present invention.As can be seen that
The cloud strong sense of reality simulated.
Data set is divided into three parts：6000 pairs of images are for the training to network, and 1500 images are for carrying out parameter
Fine tuning, 500 images as emulation collect on verification, testing model whether over-fitting.
2) identification system network
The training of two networks of identification system is completed using the newer mode of alternating iteration.Data loader is from training set
It is sampled in pairs of emulating image, is first inputted to and generates in network, carry out a propagated forward, cloud atlas picture is removed in generation；
It is then input in differentiation network and carries out a propagated forward, calculate distribution loss and reconstruction error loses；Gradient is finally calculated,
An iteration update is completed to generating network.It generates the cloudless image of neural network forecast and the input of true value figure differentiates network, carry out respectively
Propagated forward, calculates distribution loss and gradient, and passback is completed to update an iteration for differentiating network.Total data completes a wheel
After iteration, training set is upset again, starts next round iteration.When hands-on, 100 wheel iteration, loss function have been carried out altogether
I.e. basic convergence.
The initialization of convolution kernel uses the Gaussian Profile that mean value is 0.01 for 0, variance, the initialization of biasing to use fixed value
0.01.Using Adam optimization algorithms, initial learning rate is set as 0.001, and iteration 50 is reduced to half after taking turns.The finger of single order moments estimation
Number attenuation rate is set as 0.9, and the exponential decay rate of second order moments estimation is set as 0.999.Adam optimization algorithms are by calculating the one of gradient
Rank and second moment are estimated as the independent autoadapted learning rate of different parameter designings so that network can efficiently restrain.
Step 5：Thin cloud in remote sensing image removes
There is the input-output characteristic that systems compliant is removed with practical thin cloud by the trained network that generates of step 4.System
After the completion of identification, differentiate that network just no longer needs.It, only need to be by the remote sensing figure with thin cloud when carrying out thin cloud in remote sensing image removal
As input generation network, a propagated forward is carried out, network output is the cloudless image restored.
Identification gained model is the model that task is removed suitable for remote sensing satellite visible light wave range image cloud, other are defended
Star sensor image only need to choose corresponding visible light wave range and be synthesized, as mode input, you can obtain after removing thin cloud
It influences.Fig. 4 a-1, Fig. 4 a-2, Fig. 4 b-1, Fig. 4 b-2, Fig. 4 c-1, Fig. 4 c-2 illustrate to Landsat-8 satellite images,
Google Earth images, high score No.1 satellite image carry out the result of thin cloud removal.
The system equally can be generalized to the thin cloud removal task of multi-spectrum remote sensing image, only need to choose wave band interested,
True value figure is established according to step 4 and has the training data of cloud atlas picture pair, training is re-started to model, you can obtains being suitable for more
The thin cloud of spectral remote sensing image removes system.
Claims (7)
1. a kind of based on the thin cloud in remote sensing image minimizing technology for generating confrontation network, which is characterized in that steps are as follows：
Step 1：Establish thin cloud removal system total model
Thin cloud removal is carried out for the remote sensing images that the land Landsat-8OLI imager is acquired；Landsat-8 is that the U.S. exists
The satellite of transmitting on 2 11st, 2013, carries the land OLI imager and TIRS thermal infrared sensors, the OLI of Landsat-8
Land imager include 9 wave bands, wherein the 4th, 3,2 wave bands corresponded to three visible light wave ranges of red, green, blue；Select the 4th, 3,2
Wave band obtains True color synthesis image, carries out thin cloud removal；
Bao Yun imaging distortion models are described as：
S (i, j)=aLr (i, j) t (i, j)+L (1-t (i, j)) (1)
Wherein, s (i, j) is the signal that sensor receives at point (i, j), and L is air light radiation, and r (i, j) is that ground is true
Reflection, i.e., desired cloudless image, t (i, j) is transmission plot, and a is air light attenuation coefficient, in 0~1 range；
According to the imaging distortion model, when known to transmission plot, establishes cloudless clear image and have the linear pass between cloud atlas picture
System；Enable what y was denoted as condition entry to have a cloud atlas picture, z represents the noise introduced,Indicate that the cloudless image restored, g represent life
At model, i.e., thin cloud goes division operation, then thin cloud removal system can be indicated with following simplification mathematical model：
System model is removed according to the thin cloud that formula (2) is established, as long as the generation parameterized by the network identification of design
Model g then just has cloud atlas picture to predict cloudless clear image by functional relation from input, to realize to the thin of remote sensing images
Cloud removes；
Step 2：Network model designs
Neural network is fought using one group of generation to recognize thin cloud removal system model；Network used in identification system is by two
Sub-network forms：It generates network G and differentiates network D；It generates network G and predicts cloudless image using input information, that is, receiving has cloud
Image and noise are driven with criterion function and are trained so that the cloudless image of generation, which can allow, differentiates that network D is determined as very；Differentiate
The effect of network D is to differentiate that auxiliary generates the training of network G to the cloudless image of generation and true cloudless image；Two
A sub-network is contended with binary minimax game loss, in the process, generates the condition of network G study truthful data
Distribution；By the confrontation study of two sub-networks, identified parameters are obtained；After the completion of identification, generates network G and gone as with thin cloud
There is the alternative model of identical input-output characteristic except system generates model g；When going division operation to the thin cloud of remote sensing images progress,
It generates network G and carries out a propagated forward, calculated by the parameter picked out and recover cloudless image, be not required to again by sentencing
Other network D；
Step 3：Recognize criterion function structure
Model g input-output characteristics having the same are generated to make generation network G and thin cloud remove system, it is also necessary to the mistake of the two
Training supervisory signals of the difference as network；This signal is by the output and supervision sample of differentiation network D to providing；Correspondingly, accurate
Then function is also made of two parts；
Step 4：Bao Yun removes System Discrimination
4.1 training samples obtain
The identification that thin cloud removal system generates model g is carried out using confrontation type neural network model is generated, is a supervised learning
Process, need the data of tape label to generating network and differentiating that network is trained；However remote sensing satellite is to same place
Paying a return visit has period distances, and atmospheric radiation, geomorphic feature can great changes will take place within the time period, and pinpoints shooting for a long time
Cost it is again very high, so ideal, to have cloud atlas picture and its cloudless true value figure in pairs be difficult to obtain；Therefore, supervision is based on to learn
The means for the thin cloud minimizing technology generally use emulation practised obtain enough training samples；Equally, using in clear image
The method for emulating thin cloud is generated, using clear image as cloudless true value figure, obtains tape label data；
4.2 identification system network trainings
Decline thought using gradient to carry out parametric solution to identification system, specific optimization uses Adam optimization algorithms；Using imitative
Input and ideal output of the pairs of sample really obtained as identification system, to generating network G and differentiating that network D is iterated
Training, and the parameter of two sub-networks is constantly updated, when criterion function loss tends towards stability, two network trainings finish；
Obtained generation network G is to generate model g with identical defeated with the thin cloud in remote sensing image removal system described in formula (2)
Enter the alternative model of output characteristics, identification system model training finishes namely determine the parameter for generating network G, so far, Bao Yun
System Discrimination is removed to complete；
Step 5：Thin cloud in remote sensing image removes
After the completion of the identification training process of step 4, it is defeated that network model is provided with input identical with the thin cloud removal system of reality
Go out characteristic；The remote sensing images for having Bao Yun for a web are inputted and generate a network G propagated forward of progress, utilize training
Good parameter carries out operation, that is, exports the cloudless image being resumed.
2. according to claim 1 a kind of based on the thin cloud in remote sensing image minimizing technology for generating confrontation network, feature exists
In：Generation network G is the depth convolutional neural networks being made of several residual error study modules, passes through several convolution-specifications
Change-nonlinear activation operation can realize the fine fitting to nonlinear model, and residual error structure is made up of identical mapping
It practises；The local module is described as：
Wherein,Indicate that obtaining n-th layer i-th by (n-1) layer jth characteristic pattern opens convolution kernel used in characteristic pattern,Table
Show i-th characteristic pattern of network n-th layer,Indicate the jth characteristic pattern of network (n-1) layer,For i-th spy of n-th layer
Levy the bias vector of figure；NnFor the standardized operation to n-th layer output response, ΦnFor non-linear the swashing to n-th layer output response
Operation living；The identical mapping that characteristic pattern opens to n-th layer i-th characteristic pattern is opened for (n-1) layer i-th, this structure makes convolutional layer
Residual error between input and desired output learns so that fitting is more easy, to reduce the training hardly possible for generating network
Degree；Designed generation network structure is：
DownsampleBlock×3→ResBlock×6→UpsampleBlock×3
DownsampleBlock:Conv (4 × 4, stride=2) → BN → LeakyRelu
UpsampleBlock:TransposedConv (4 × 4, stride=2) → BN → LeakyRelu
Wherein, Conv represents convolutional layer, and TransposedConv is warp lamination, in bracket numerical value for used convolution kernel ruler
Very little and step-length uses the convolution kernel of 4 × 4 sizes herein, and step-length is 2；DownsampleBlock is down-sampled module, is reduced
The dimension of image reduces the parameter amount learnt required for network；ResBlock volumes with identical mapping shown in formula (3)
Product-standardization-nonlinear activation layer forms, and also uses the convolution kernel of 4 × 4 sizes, has used 6 residual blocks；
UpsampleBlock is up-sampling module, corresponding DownsampleBlock modules, for restoring image original size；Using compared with
On the one hand small convolution kernel reduces training parameter amount, on the other hand so that the reconstruction process of image reinforces the profit to local message
With；With doing and standardizing to response output there are one standardization layer BN after each convolutional layer, keeps its mean value consistent with variance, avoid
The explosion of local gradient or disappearance in training；It is a nonlinear activation layer after standardization layer, nonlinear activation function uses
LeakyRelu functions.
3. according to claim 1 a kind of based on the thin cloud in remote sensing image minimizing technology for generating confrontation network, feature exists
In：When differentiating to generation image and true picture, differentiate that network D needs the sense of reality from generation image simultaneously and is
It is no that thin cloud differentiate in terms of effectively removing two；This just needs to differentiate that network D is not concerned only with local texture spy
Sign, judges the continuity at edge and texture, the sharp keen degree of texture, while also to drop and carry out to the feature of image entirety
Judge；Its structure is as follows, the differentiation network differentiated in conjunction with Analysis On Multi-scale Features for one；
DownsampleBlock×3→MultiscaleBlock×3
Wherein, down-sampled module is structurally and functionally identical with generation network G；Each MultiscaleBlock is one
Multi resolution feature extraction module is made of three branches, and each branch uses various sizes of convolution kernel, to obtain different size
Receptive field, from different scales obtain image feature, i.e.,：
MultiscaleBlock=[conv (1 × 1), conv (3 × 3), conv (5 × 5)]
It is 1 convolution operation that each convolutional layer, which use step-length, and carries out edge filling, so that exporting the holding of characteristic pattern size
Unanimously；Then, it connects to the activation output that each branch obtains, obtains the feature of different scale；With generation network G phase
Together, there are one output standardization layer and a nonlinear activation layers for connection after each convolutional layer.
4. according to claim 1 a kind of based on the thin cloud in remote sensing image minimizing technology for generating confrontation network, feature exists
In：
First part, data distribution criterion function；This part is intersected to differentiating that the output result of network D calculates true and false two classification
Entropy is as criterion function；For differentiation network, need to maximize criterion function, cloudless true value figure and generation network is extensive
Multiple cloudless image strict differences are opened；For generating for network G, needs to minimize this criterion function, be differentiated with " deception "
Network makes differentiation network restore the classification that image carries out mistake to it；Ideally, differentiate true picture of the network to input
It is 0.5 with image discriminating is generated as genuine probability；Therefore, this criterion function is write：
J1=E [logD (x)]+E [log (1-D (G (y, z)))] (4)
Wherein, x is true cloudless image, and y is represented has a cloud atlas picture as condition entry, and z represents the noise of introducing, D () and
G () respectively represents the output for differentiating network and generating network, and E () indicates corresponding to each batch sample inputted when training
Mathematic expectaion is sought in output；Pass through the two minimax alternative optimization so that generating network can extract indirectly by network is differentiated
High frequency detail feature and global characteristics.
5. according to claim 1 a kind of based on the thin cloud in remote sensing image minimizing technology for generating confrontation network, feature exists
In：
Second part, response error function；For input sample, structure quadratic form error function is as follows：
Generate the Euclidean distance between prognostic chart and cloudless true value figure；In formula subscript k indicate to network carry out small lot with
Machine gradient declines the serial number that sample in each small lot sample is taken when training；xk, ykAnd zkRespectively represent the cloudless sample of kth group
Originally the noise for, having cloud sample and this group of input sample being introduced；It is flat to obtain by minimizing this loss function to generate network
The smooth characteristics of low-frequency of equal colouring information, regression forecasting is done to image；
Therefore, the criterion function of final design is：
J=J1+J2 (6)。
6. according to claim 1 a kind of based on the thin cloud in remote sensing image minimizing technology for generating confrontation network, feature exists
In：
Clear cloudless true value figure obtains：Cloudless clear image is chosen from remote sensing satellite image, choosing image should be as much as possible
Covering farm land, forest, vegetation, mountain area, waters, city various landforms, to establish a feature remote sensing images complete as possible
Tranining database so that finally obtained thin cloud removal system can be widely used in very color acquired in various satellite sensors
Color image；Since remote sensing images picture is all bigger, it is also necessary to carry out cutting screening operation to image；To original multispectral remote sensing
Image opens it in ENVI softwares, and selection the 4th, 3,2 wave bands obtain True color synthesis image, and manually with 256 × 256
Fixed size is cut and is preserved to image.
7. according to claim 1 a kind of based on the thin cloud in remote sensing image minimizing technology for generating confrontation network, feature exists
In：
Thin cloud image simulation synthesis：The thin cloud atlas used is emulated as emulation mode, using Landsat-8 satellite bands 9 into racking；
At 1.36~1.38 μm, high-altitude steam has it compared with strong reflection spy the wavelength of Landsat-8 satellite OLI sensor cirrus wave bands
Property, low latitude atural object and dry environment reflection are very weak, therefore cloud layer responds highly significant in the wave band, translucent white are presented, ground
Object response is faint, and large stretch of black is presented in the picture, is used for cloud detection task；Using the wave band, extraction cloud layer reflects gray scale
Figure, as transmission plot t, brings into the imaging distortion model of formula (1), and combines the atmosphere light of different level to get to synthesis
Thin cloud remote sensing images.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810174430.3A CN108460739A (en) | 2018-03-02 | 2018-03-02 | A kind of thin cloud in remote sensing image minimizing technology based on generation confrontation network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810174430.3A CN108460739A (en) | 2018-03-02 | 2018-03-02 | A kind of thin cloud in remote sensing image minimizing technology based on generation confrontation network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108460739A true CN108460739A (en) | 2018-08-28 |
Family
ID=63217577
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810174430.3A Pending CN108460739A (en) | 2018-03-02 | 2018-03-02 | A kind of thin cloud in remote sensing image minimizing technology based on generation confrontation network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108460739A (en) |
Cited By (32)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109191400A (en) * | 2018-08-30 | 2019-01-11 | 中国科学院遥感与数字地球研究所 | A method of network, which is generated, using confrontation type removes thin cloud in remote sensing image |
| CN109215123A (en) * | 2018-09-20 | 2019-01-15 | 电子科技大学 | Unlimited landform generation method, system, storage medium and terminal based on cGAN |
| CN109685863A (en) * | 2018-12-11 | 2019-04-26 | 帝工(杭州)科技产业有限公司 | A method of rebuilding medicine breast image |
| CN109859147A (en) * | 2019-03-01 | 2019-06-07 | 武汉大学 | A kind of true picture denoising method based on generation confrontation network noise modeling |
| CN109872278A (en) * | 2018-12-18 | 2019-06-11 | 深圳先进技术研究院 | Image cloud layer removing method based on U-shape network and generation confrontation network |
| CN109949242A (en) * | 2019-03-19 | 2019-06-28 | 内蒙古工业大学 | The generation method of image defogging model, device and image defogging method, device |
| CN110211045A (en) * | 2019-05-29 | 2019-09-06 | 电子科技大学 | Super-resolution face image method based on SRGAN network |
| CN110599401A (en) * | 2019-08-19 | 2019-12-20 | 中国科学院电子学研究所 | Remote sensing image super-resolution reconstruction method, processing device and readable storage medium |
| WO2020093792A1 (en) * | 2018-11-08 | 2020-05-14 | Guangdong Oppo Mobile Telecommunications Corp., Ltd. | Method, system, and computer-readable medium for improving color quality of images |
| CN111210483A (en) * | 2019-12-23 | 2020-05-29 | 中国人民解放军空军研究院战场环境研究所 | Simulated satellite cloud picture generation method based on generation of countermeasure network and numerical mode product |
| CN111337929A (en) * | 2020-03-26 | 2020-06-26 | 上海眼控科技股份有限公司 | Meteorological cloud picture prediction method and device, computer equipment and storage medium |
| CN111351502A (en) * | 2018-12-21 | 2020-06-30 | 赫尔环球有限公司 | Method, apparatus and computer program product for generating an overhead view of an environment from a perspective view |
| CN111667431A (en) * | 2020-06-09 | 2020-09-15 | 云南电网有限责任公司电力科学研究院 | Method and device for manufacturing cloud and fog removing training set based on image conversion |
| CN111784787A (en) * | 2019-07-17 | 2020-10-16 | 北京沃东天骏信息技术有限公司 | Image generation method and device |
| CN111951284A (en) * | 2020-08-12 | 2020-11-17 | 湖南神帆科技有限公司 | Optical remote sensing satellite image refined cloud detection method based on deep learning |
| CN112529788A (en) * | 2020-11-13 | 2021-03-19 | 北京航空航天大学 | Multispectral remote sensing image thin cloud removing method based on thin cloud thickness map estimation |
| CN112565756A (en) * | 2020-11-26 | 2021-03-26 | 西安电子科技大学 | Cloud-containing remote sensing image compression method based on quantization strategy |
| CN112634169A (en) * | 2020-12-30 | 2021-04-09 | 成都星时代宇航科技有限公司 | Color homogenizing method and device for remote sensing image |
| CN112862670A (en) * | 2021-02-08 | 2021-05-28 | 中国科学院信息工程研究所 | 360-degree image countermeasure sample generation method based on multi-view fusion |
| CN113240208A (en) * | 2021-06-25 | 2021-08-10 | 国网浙江省电力有限公司双创中心 | Photovoltaic power generation output power prediction method and device and computer readable storage medium |
| CN113298836A (en) * | 2021-06-29 | 2021-08-24 | 天津市测绘院有限公司 | Remote sensing image thin cloud removing method and system considering element contour intensity |
| CN113628098A (en) * | 2021-07-28 | 2021-11-09 | 北京航空航天大学 | Multi-temporal cloud-free satellite image reconstruction method |
| CN113724149A (en) * | 2021-07-20 | 2021-11-30 | 北京航空航天大学 | Weak supervision visible light remote sensing image thin cloud removing method |
| CN113935917A (en) * | 2021-10-14 | 2022-01-14 | 中国石油大学(华东) | Optical remote sensing image thin cloud removing method based on cloud picture operation and multi-scale generation countermeasure network |
| CN114066755A (en) * | 2021-11-11 | 2022-02-18 | 南京航空航天大学 | Remote sensing image thin cloud removing method and system based on full-band feature fusion |
| CN114936972A (en) * | 2022-04-06 | 2022-08-23 | 航天东方红卫星有限公司 | Remote sensing image thin cloud removing method based on multi-path perception gradient |
| CN115661002A (en) * | 2022-12-14 | 2023-01-31 | 北京数慧时空信息技术有限公司 | Multi-temporal remote sensing data restoration method based on GAN |
| CN116823664A (en) * | 2023-06-30 | 2023-09-29 | 中国地质大学（武汉） | Remote sensing image cloud removal method and system |
| CN117152622A (en) * | 2023-10-30 | 2023-12-01 | 中国科学院空天信息创新研究院 | Boundary optimization model training, boundary optimization method, device, equipment and medium |
| CN117611494A (en) * | 2024-01-24 | 2024-02-27 | 北京理工大学 | Panchromatic remote sensing image thin cloud removing method |
| CN117851927A (en) * | 2024-03-07 | 2024-04-09 | 南京大学 | Cloud prediction method based on random forest and meteorological data |
| CN114936972B (en) * | 2022-04-06 | 2024-07-02 | 航天东方红卫星有限公司 | Remote sensing image thin cloud removing method based on multipath perception gradient |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106127702A (en) * | 2016-06-17 | 2016-11-16 | 兰州理工大学 | A kind of image mist elimination algorithm based on degree of depth study |
-
2018
- 2018-03-02 CN CN201810174430.3A patent/CN108460739A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106127702A (en) * | 2016-06-17 | 2016-11-16 | 兰州理工大学 | A kind of image mist elimination algorithm based on degree of depth study |
Non-Patent Citations (4)
| Title |
|---|
| OREST KUPYN: "DeblurGAN: Blind Motion Deblurring Using Conditional Adversarial Networks", 《ARXIV:1711.07064V3》 * |
| QIAOBO CHEN: "Rain Removal via Residual Generation Cascading", 《IEEE》 * |
| 李策: "生成对抗映射网络下的图像多层感知去雾算法", 《计算机辅助设计与图形学学报》 * |
| 阿布: "《电子工业出版社》", 31 August 2017 * |
Cited By (49)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109191400A (en) * | 2018-08-30 | 2019-01-11 | 中国科学院遥感与数字地球研究所 | A method of network, which is generated, using confrontation type removes thin cloud in remote sensing image |
| CN109215123A (en) * | 2018-09-20 | 2019-01-15 | 电子科技大学 | Unlimited landform generation method, system, storage medium and terminal based on cGAN |
| CN109215123B (en) * | 2018-09-20 | 2022-07-29 | 电子科技大学 | Method, system, storage medium and terminal for generating infinite terrain based on cGAN |
| WO2020093792A1 (en) * | 2018-11-08 | 2020-05-14 | Guangdong Oppo Mobile Telecommunications Corp., Ltd. | Method, system, and computer-readable medium for improving color quality of images |
| US11972543B2 (en) | 2018-11-08 | 2024-04-30 | Guangdong Oppo Mobile Telecommunications Corp., Ltd. | Method and terminal for improving color quality of images |
| CN109685863A (en) * | 2018-12-11 | 2019-04-26 | 帝工(杭州)科技产业有限公司 | A method of rebuilding medicine breast image |
| CN109872278B (en) * | 2018-12-18 | 2020-11-10 | 深圳先进技术研究院 | Image cloud layer removing method based on U-shaped network and generation countermeasure network |
| CN109872278A (en) * | 2018-12-18 | 2019-06-11 | 深圳先进技术研究院 | Image cloud layer removing method based on U-shape network and generation confrontation network |
| US11720992B2 (en) | 2018-12-21 | 2023-08-08 | Here Global B.V. | Method, apparatus, and computer program product for generating an overhead view of an environment from a perspective image |
| CN111351502A (en) * | 2018-12-21 | 2020-06-30 | 赫尔环球有限公司 | Method, apparatus and computer program product for generating an overhead view of an environment from a perspective view |
| CN109859147A (en) * | 2019-03-01 | 2019-06-07 | 武汉大学 | A kind of true picture denoising method based on generation confrontation network noise modeling |
| CN109949242A (en) * | 2019-03-19 | 2019-06-28 | 内蒙古工业大学 | The generation method of image defogging model, device and image defogging method, device |
| CN110211045A (en) * | 2019-05-29 | 2019-09-06 | 电子科技大学 | Super-resolution face image method based on SRGAN network |
| CN110211045B (en) * | 2019-05-29 | 2022-09-06 | 电子科技大学 | Super-resolution face image reconstruction method based on SRGAN network |
| CN111784787B (en) * | 2019-07-17 | 2024-04-09 | 北京沃东天骏信息技术有限公司 | Image generation method and device |
| CN111784787A (en) * | 2019-07-17 | 2020-10-16 | 北京沃东天骏信息技术有限公司 | Image generation method and device |
| CN110599401A (en) * | 2019-08-19 | 2019-12-20 | 中国科学院电子学研究所 | Remote sensing image super-resolution reconstruction method, processing device and readable storage medium |
| CN111210483A (en) * | 2019-12-23 | 2020-05-29 | 中国人民解放军空军研究院战场环境研究所 | Simulated satellite cloud picture generation method based on generation of countermeasure network and numerical mode product |
| CN111337929A (en) * | 2020-03-26 | 2020-06-26 | 上海眼控科技股份有限公司 | Meteorological cloud picture prediction method and device, computer equipment and storage medium |
| CN111667431A (en) * | 2020-06-09 | 2020-09-15 | 云南电网有限责任公司电力科学研究院 | Method and device for manufacturing cloud and fog removing training set based on image conversion |
| CN111667431B (en) * | 2020-06-09 | 2023-04-14 | 云南电网有限责任公司电力科学研究院 | Method and device for manufacturing cloud and fog removing training set based on image conversion |
| CN111951284A (en) * | 2020-08-12 | 2020-11-17 | 湖南神帆科技有限公司 | Optical remote sensing satellite image refined cloud detection method based on deep learning |
| CN111951284B (en) * | 2020-08-12 | 2022-04-22 | 湖南神帆科技有限公司 | Optical remote sensing satellite image refined cloud detection method based on deep learning |
| CN112529788B (en) * | 2020-11-13 | 2022-08-19 | 北京航空航天大学 | Multispectral remote sensing image thin cloud removing method based on thin cloud thickness map estimation |
| CN112529788A (en) * | 2020-11-13 | 2021-03-19 | 北京航空航天大学 | Multispectral remote sensing image thin cloud removing method based on thin cloud thickness map estimation |
| CN112565756B (en) * | 2020-11-26 | 2021-09-03 | 西安电子科技大学 | Cloud-containing remote sensing image compression method based on quantization strategy |
| CN112565756A (en) * | 2020-11-26 | 2021-03-26 | 西安电子科技大学 | Cloud-containing remote sensing image compression method based on quantization strategy |
| CN112634169B (en) * | 2020-12-30 | 2024-02-27 | 成都星时代宇航科技有限公司 | Remote sensing image color homogenizing method and device |
| CN112634169A (en) * | 2020-12-30 | 2021-04-09 | 成都星时代宇航科技有限公司 | Color homogenizing method and device for remote sensing image |
| CN112862670A (en) * | 2021-02-08 | 2021-05-28 | 中国科学院信息工程研究所 | 360-degree image countermeasure sample generation method based on multi-view fusion |
| CN112862670B (en) * | 2021-02-08 | 2023-12-05 | 中国科学院信息工程研究所 | 360-degree image countermeasure sample generation method based on multi-view fusion |
| CN113240208A (en) * | 2021-06-25 | 2021-08-10 | 国网浙江省电力有限公司双创中心 | Photovoltaic power generation output power prediction method and device and computer readable storage medium |
| CN113298836B (en) * | 2021-06-29 | 2023-03-14 | 天津市测绘院有限公司 | Remote sensing image thin cloud removing method and system considering element contour intensity |
| CN113298836A (en) * | 2021-06-29 | 2021-08-24 | 天津市测绘院有限公司 | Remote sensing image thin cloud removing method and system considering element contour intensity |
| CN113724149A (en) * | 2021-07-20 | 2021-11-30 | 北京航空航天大学 | Weak supervision visible light remote sensing image thin cloud removing method |
| CN113724149B (en) * | 2021-07-20 | 2023-09-12 | 北京航空航天大学 | Weak-supervision visible light remote sensing image thin cloud removing method |
| CN113628098A (en) * | 2021-07-28 | 2021-11-09 | 北京航空航天大学 | Multi-temporal cloud-free satellite image reconstruction method |
| CN113935917A (en) * | 2021-10-14 | 2022-01-14 | 中国石油大学(华东) | Optical remote sensing image thin cloud removing method based on cloud picture operation and multi-scale generation countermeasure network |
| CN114066755A (en) * | 2021-11-11 | 2022-02-18 | 南京航空航天大学 | Remote sensing image thin cloud removing method and system based on full-band feature fusion |
| CN114936972A (en) * | 2022-04-06 | 2022-08-23 | 航天东方红卫星有限公司 | Remote sensing image thin cloud removing method based on multi-path perception gradient |
| CN114936972B (en) * | 2022-04-06 | 2024-07-02 | 航天东方红卫星有限公司 | Remote sensing image thin cloud removing method based on multipath perception gradient |
| CN115661002A (en) * | 2022-12-14 | 2023-01-31 | 北京数慧时空信息技术有限公司 | Multi-temporal remote sensing data restoration method based on GAN |
| CN116823664A (en) * | 2023-06-30 | 2023-09-29 | 中国地质大学（武汉） | Remote sensing image cloud removal method and system |
| CN116823664B (en) * | 2023-06-30 | 2024-03-01 | 中国地质大学（武汉） | Remote sensing image cloud removal method and system |
| CN117152622A (en) * | 2023-10-30 | 2023-12-01 | 中国科学院空天信息创新研究院 | Boundary optimization model training, boundary optimization method, device, equipment and medium |
| CN117152622B (en) * | 2023-10-30 | 2024-02-23 | 中国科学院空天信息创新研究院 | Boundary optimization model training, boundary optimization method, device, equipment and medium |
| CN117611494B (en) * | 2024-01-24 | 2024-04-30 | 北京理工大学 | Panchromatic remote sensing image thin cloud removing method |
| CN117611494A (en) * | 2024-01-24 | 2024-02-27 | 北京理工大学 | Panchromatic remote sensing image thin cloud removing method |
| CN117851927A (en) * | 2024-03-07 | 2024-04-09 | 南京大学 | Cloud prediction method based on random forest and meteorological data |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108460739A (en) | A kind of thin cloud in remote sensing image minimizing technology based on generation confrontation network | |
| Karra et al. | Global land use/land cover with Sentinel 2 and deep learning | |
| Hsu et al. | Precipitation estimation from remotely sensed information using artificial neural networks | |
| Zhuo et al. | Physics-augmented deep learning to improve tropical cyclone intensity and size estimation from satellite imagery | |
| US20160104059A1 (en) | Identifying visual storm signatures from satellite images | |
| CN106355151A (en) | Recognition method, based on deep belief network, of three-dimensional SAR images | |
| CN109063569A (en) | A kind of semantic class change detecting method based on remote sensing image | |
| CN110046415A (en) | A kind of soil organic matter content remote sensing dynamic playback method of space-time fining | |
| CN106605225A (en) | Simulation device, simulation method, and memory medium | |
| CN103268607B (en) | A kind of common object detection method under weak supervision condition | |
| CN114564982B (en) | Automatic identification method for radar signal modulation type | |
| CN111160268A (en) | Multi-angle SAR target recognition method based on multi-task learning | |
| Hagen et al. | Identifying major drivers of daily streamflow from large-scale atmospheric circulation with machine learning | |
| Otsuka et al. | Nowcasting with data assimilation: A case of global satellite mapping of precipitation | |
| CN116403048B (en) | Crop growth estimation model construction method based on multi-mode data fusion | |
| CN114529097A (en) | Multi-scale crop phenological period remote sensing dimensionality reduction prediction method | |
| CN114399073A (en) | Ocean surface temperature field prediction method based on deep learning | |
| Bai et al. | Optimal window size selection for spectral information extraction of sampling points from UAV multispectral images for soil moisture content inversion | |
| Lou et al. | An effective method for canopy chlorophyll content estimation of marsh vegetation based on multiscale remote sensing data | |
| Moschos et al. | Computer Vision for Ocean Eddy Detection in Infrared Imagery | |
| Vietinghoff et al. | Visual analysis of spatio-temporal trends in time-dependent ensemble data sets on the example of the North Atlantic Oscillation | |
| Lou et al. | Recognition of thaw slumps based on machine learning and UAVs: A case study in the Qilian Mountains, northeastern Qinghai-Tibet Plateau | |
| CN116863341B (en) | Crop classification and identification method and system based on time sequence satellite remote sensing image | |
| CN114971041A (en) | Sea ice area prediction method and system based on residual error network | |
| Astola et al. | Highforest-forest parameter estimation from high resolution remote sensing data |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180828 |
|
| WD01 | Invention patent application deemed withdrawn after publication |