CN109635748A - The extracting method of roadway characteristic in high resolution image - Google Patents
The extracting method of roadway characteristic in high resolution image Download PDFInfo
- Publication number
- CN109635748A CN109635748A CN201811532429.XA CN201811532429A CN109635748A CN 109635748 A CN109635748 A CN 109635748A CN 201811532429 A CN201811532429 A CN 201811532429A CN 109635748 A CN109635748 A CN 109635748A
- Authority
- CN
- China
- Prior art keywords
- generator
- sample
- discriminator
- image
- roadway characteristic
- 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
- G06V20/182—Network patterns, e.g. roads or rivers
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/30—Noise filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
Abstract
The extracting method of roadway characteristic in a kind of high resolution image.It obtains the generator G that can accurately identify roadway characteristic by the training of optimization and parameter to generator G, discriminator D and generation confrontation network V (D, G) structure.In training process, the present invention generates sample P by assessing using Wasserstein distanceg(x) with authentic specimen Pdata(x) gap can effectively avoid the problem that the gradient of the loss function occurred when the existing KL divergence utilized or JS divergence are assessed does not restrain.Thus obtained generator G identifies obtained roadway characteristic, and closer to actual conditions, accuracy is higher.
Description
Technical field
The present invention relates to a kind of extracting methods of roadway characteristic in field of image processing more particularly to high resolution image.
Background technique
High-definition remote sensing technology grows continuously and fast with aeroplane photography obtaining means, has driven high resolution image
Extensive use.Its obtain ground resolution less than 10m remote sensing observations image for Urban Traffic Planning, environmental monitoring,
Equal civil fields are also well worth doing.
To 2018, has more high score satellites and come into operation.High-resolution earth observation systems traffic transport industry number
It is supported according to center also synchronous opening traffic transport industry high score data.It is contemplated that the combination of " high score+traffic " will be into
It is horizontal that one step promotes the fine-grained management to path resource.
Road is one of most basic terrestrial object information in the field GIS, is not only important state-owned assets and connection area
The crucial tie of domain economy and resident trip.Timely and effectively grasping road status and its situation of change seems most important.Road
The aobvious acquisition of road status and its situation of change need road geographical feature is information-based, realize the Optimal Management of path resource with
It utilizes.
Since, without complete data report flow, especially rural area is remote in early stage road construction and recondition
The inferior grade road in area lacks data required for informationization and supports.For the management for realizing this part path resource, pass through
The technology that high resolution image extracts roadway characteristic is come into being.That is, path resource manager wants to from existing
Reflect in the data source of road geographic information, effective road basic data is gone out according to road feature extraction, formation can be managed remotely
The information resources of reason.
Work is extracted in the automation for carrying out road based on high resolution image, since it can substantially use manpower and material resources sparingly,
The always research emphasis of domestic and foreign scholars.Domestic and foreign scholars also propose a variety of feasible schemes to this demand.In recent years,
The deep development of deep learning research also extracts work for road network and provides new thinking and inspiration, this to lead in image procossing
Domain achieves the artificial intelligence technology greatly broken through, provides new thinking and inspiration for road extraction work.Due to depth
Habit technology is based on great amount of samples training, it is thus possible to so that machine has learning ability, found from object common feature and according to
Model after training makes a policy judgement.The application of depth learning technology further improves the precision of recognition detection
And efficiency.
Mainstream data source currently used for road feature extraction is all types of satellite remote-sensing image, i.e., by from number
The one kind or multiclass feature for meeting specified conditions are found in image, and restore the geographical letter of road within the pixel from feature
Breath.Common roadway characteristic is summarized as following four class in the research of Vosselman and Knecht by nineteen ninety-five:
1) geometrical characteristic: physical feature of the roadway characteristic on geometry.It reflects in the picture, to keep certain
The elongated strip shaped line segment of width, slow along road extending direction change procedure, road width reflects road etc. to a certain extent
Grade, it is neat that how straight advanced road is, and inferior grade road wriggles rugged more；
2) spectral signature: physical feature of the roadway characteristic in material.It is reflected as, the road in same scape remote sensing image
Road internal color is evenly distributed, and spectral characteristic is close, and there are certain contrasts with its two sides scenery.It is easily observed under high-resolution
To more apparent spectrum line of demarcation；
3) topological characteristic: linked character of the road topology feature in structure, reflection are that two road can in the picture
Can there are the intersection of " X ", " Y " or T-shape, and the angle of infall has certain degree limitation, meanwhile, between each road mutually
It is connected to form one whole road network；
4) contextual feature: other elements in image can be used for assisting in identifying roadway characteristic, be roadway characteristic
Linked character.Such as, continuous vegetation or building information on region, village or town information in the overall situation.
Road extraction research based on remote sensing image originates in 1970s.However, due to early stage Computing
Horizontal and digital image processing techniques fall behind relatively, and the large sample size that hardware design conditions can not cope with remote sensing image image mentions
Take task.The remote sensing images of early stage more take manual method to extract.Hereafter with calculate equipment performance promotion and
The deep development of research theory, lot of domestic and foreign scholar propose all kinds of roadway characteristic identification sides based on image characteristics extraction
Method.
According to the difference for the characteristic properties that road feature extraction is relied on, existing extracting method can be divided into based on pixel side
Method and object-oriented method.Road feature extraction method based on pixel depends on the geometrical characteristic of road, with image picture
Member be used as basic unit, start with from the directions such as wave spectrum and space characteristics, find satisfactory pixel region, thus segmentation and
Identify site of road.The road feature extraction method of object-oriented depends on the spectral signature of road, with road itself
Modeling Research is carried out as a whole, is focused on data and operation all in object, i.e., is identified road from image.
Although above-mentioned traditional road feature extraction method significantly reduces the cost manually extracted and greatlys improve
Extraction efficiency, but precision of its identification still has a biggish gap compared to artificial identification method.In particular, blocking
Or on the section of shade, above-mentioned conventional method extraction result generated, which still needs manually to post-process, is repaired.Especially:
1) it for specific region or the road of type, although each traditional methods multipotency enough obtains good effect, lacks
Weary universality, needs the road for the region or type to carry out special parameter designing, and migration application effect is bad；
2) when facing low, intermediate-resolution remote sensing image, although all kinds of methods can obtain certain effect,
It is commonly used ineffective when being required in face of the fining extraction that high-resolution remote sensing image is proposed；
3) conventional method is unable to fully carry out comprehensive descision, conventional method using each category feature present in remote sensing image
Often only focus on the processing to Partial Feature and level in remote sensing impression；
4) the degree of automation is relatively low, and artificial pretreatment and artificial post-processing intervention are frequent.
Summary of the invention
In order to solve the shortcomings of the prior art, the purpose of the present invention is to provide one kind to be directed to high resolution image
The extracting method of middle roadway characteristic, simple to the pre-processing requirements of image, recognition effect is more accurate and is suitable for a variety of figures
Image field scape.
Firstly, to achieve the above object, a kind of extracting method of roadway characteristic in high resolution image, step packet are proposed
Include: the first step obtains remote sensing images, it is carried out slicing treatment according to fixed dimension, remote sensing images is obtained and is sliced x；Second
Remote sensing images slice is input to progress propagated forward operation in trained generator G, obtains road therein by step
Feature exports the roadway characteristic.Wherein, in second step, the generator G is obtained according to following steps training: step s1, right
Sample image carries out slicing treatment according to fixed dimension, obtains image slice sample z, and in each image slice sample z
Road element be marked；Step s2, constructs generator G and discriminator D, and initialization generates confrontation network V (D, G)；Its
In, the generator G is a residual error network, and the discriminator D is convolutional network, and the loss function of building generator G isConstruct discriminator D loss function be-((1-t) log (1- (D (G (z)))+
ylogD(x))；Wherein, t=1 represents input as remote sensing images slice, and t=0 represents input as sample image slice；It represents
Convolutional network after rounding exports result；Generally taking threshold value is 0.5, i.e., model judgement is assert when exporting result and being greater than 0.5
Input data is authentic specimen, and rounding up is 1, on the contrary assert input data be generate sample, downwards be rounded be 0, this two
It is the feedback being arranged be because of needing consideration to input the source being distributed and output result in generator training；Step s3 is formulated
Optimization aim isWherein,For
Majorized function and its optimization direction, Pdata(x) indicate that whole remote sensing images slice x's is distributed as authentic specimen, Pz(z) it indicates
The prior distribution of image slice sample z, E, which is represented, seeks loss function to overall data distribution in training process, then
Wherein, PgIt (x) is the distribution of the generator G generation sample obtained；Step s4 assesses generation using Wasserstein distance
Sample Pg(x) with authentic specimen Pdata(x) gap；Wherein, Wasserstein distance W (Pdata(x)',Pg(x)) it isWherein, Pdata(x) ' indicate Pg(x) and Pdata(x) it combines
The all possible Joint Distribution set got up, unitary sampling (x, G (z))~γ obtain an authentic specimen x and generate sample
G (z), | | x-G (z) | | for the distance between the authentic specimen x and generation sample G (z), E(x, G (z))~γ[||x-G(z)||]
For the desired value of the distance between the authentic specimen x and generation sample G (z)；Step s6, by marked good image slice
Sample inputs generator G, according toThe loss for calculating generator, according toCalculate the loss of discriminator；Step s7 carries out propagated forward fortune to generator G in step s6
Obtained loss carries out BP backpropagation operation；Generator and discriminator are alternately trained respectively, optimize network parameter；Step
S8, repeating said steps s6 optimize its network parameter, until the generator G to step s7, training generator and discriminator
Reach Nash Equilibrium with the discriminator D, or until the loss of the generator G and discriminator D is constant, output is at this time
Generator G as trained generator G.
Optionally, in above-mentioned extracting method, in the step s8, it is D (G (z)) ≈ that discriminator D, which reaches Nash Equilibrium,
0.5。
Optionally, in above-mentioned extracting method, in the second step, first respectively with different parameters and structure building life
Grow up to be a useful person G, discriminator D and generate confrontation network V (D, G) carry out step s2 to step s8 training, obtain different generator G；
Then advantageous one group of generator is selected from the different generator G, according to the parameter and structure of this group of generator
Different weights is set, fusion recombination is carried out according to respective weight to this group of generator, is ultimately generated with what is formed after fusion
Device G ' is used as trained generator G, ultimately generates device G ' with this and carries out propagated forward operation acquisition roadway characteristic.
Optionally, in above-mentioned extracting method, the structure type of the discriminator D is
Optionally, in above-mentioned extracting method, in the step s1, to the road element in described image slice sample z
The label of progress specifically includes: the serial number of road element, the width of road element, the material of road element, belonging to road element
Environment.
Optionally, in above-mentioned extracting method, in the step s1, described image is sliced positive sample and negative sample in sample z
This ratio is close.
Optionally, in above-mentioned extracting method, described image be sliced sample z or the remote sensing images slice without
Even color processing.
Optionally, in above-mentioned extracting method, it is size phase that described image, which is sliced sample z or remote sensing images slice,
Deng bianry image.
Beneficial effect
The present invention, by generator G, discriminator D and generating the optimization and parameter of confrontation network V (D, G) structure
Training obtains the generator G that can accurately identify roadway characteristic.It is close by selection percentage in training process, without even
The positive negative sample of color processing, and sample is marked according to the characteristics of road.Thus the roadway characteristic at identification, more connects
Nearly actual conditions, accuracy are higher.
Further, in the calculating process to loss function, it is contemplated that the KL divergence or JS divergence usually utilized is being commented
Estimate and generates sample Pg(x) with authentic specimen Pdata(x) it will lead to P when gapg(x) and Pdata(x) two are distributed in weight itself
When folded region is less, causes the gradient of generator loss function to be fixed as a constant, influence subsequent training process.Therefore this hair
It is bright to replace the two divergences using Wasserstein distance, to evaluate sample Pg(x) with authentic specimen Pdata(x) gap.By
Gap between Wasserstein distance still can smoothly be reacted when two distributions are not overlapped, thus this
It is not in the problem of model parameter does not restrain in the training process of invention.
Further, the present invention also utilizes different parameters and structure constructs generator G, discriminator D and generates confrontation network
V (D, G), carries out the fusion of different generators.The process of fusion can sufficiently receive different parameters and model structure to be had
Advantage, filter random noise by different generators while retaining inevitable road, guarantee that the image of generation result is net
Degree, the present invention can improve generation precision as much as possible as a result,.
Other features and advantages of the present invention will be illustrated in the following description, also, partly becomes from specification
It obtains it is clear that understand through the implementation of the invention.
Detailed description of the invention
Attached drawing is used to provide further understanding of the present invention, and constitutes part of specification, and with it is of the invention
Embodiment together, is used to explain the present invention, and is not construed as limiting the invention.In the accompanying drawings:
Fig. 1 is the flow diagram of road feature extraction method according to the present invention；
Fig. 2 is that condition according to the present invention generates confrontation network architecture figure；
Fig. 3 is WGAN Model Fusion frame structure according to the present invention
Fig. 4 is the extracted roadway characteristic of WGAN Model Fusion according to the present invention；
Fig. 5 is the data source tile of the invention obtained in a kind of concrete application scene；
Fig. 6 is the roadway characteristic marked in above-mentioned data source tile；
Fig. 7 is the roadway characteristic data by marking in above-mentioned data source tile；
Fig. 8 is the sample data set by marking acquisition in above-mentioned data source tile；
Fig. 9 is by the above-mentioned extracted roadway characteristic of WGAN model.
Specific embodiment
Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings, it should be understood that preferred reality described herein
Apply example only for the purpose of illustrating and explaining the present invention and is not intended to limit the present invention.
Fig. 1 is the extracting method of roadway characteristic in high resolution image according to the present invention, and step includes:
The first step obtains remote sensing images, it is carried out slicing treatment according to fixed dimension, remote sensing images is obtained and is sliced x；
Remote sensing images slice is input to progress propagated forward operation in trained generator G, obtained by second step
Roadway characteristic therein is obtained, the roadway characteristic is exported.
Above-mentioned extracting method can extract road element from original high score remote sensing image, and this method is in WGAN mould
Related link is optimized on the basis of type, method mainly includes following innovative point:
1:WGAN structure optimization and sample labeling process
2:WGAN sample fights process
3:WGAN Model Fusion scheme.
Analyze above-mentioned innovative point one by one below.
1: the present invention is the process of WGAN sample labeling calculated, main improvement: applied by this patent
Deep learning model is using standard WGAN structure as prototype, and according to road extraction task and high score image feature to network structure
Adjustment is optimized with training process.Its structure is as shown in Figure 2.The structure includes the essential characteristic in standard GAN structure,
There is an a generator G and discriminator D.It is raw in order to preferably learn each level high score image road feature and data
At potential principle, generator G is made of a more complicated residual error network in model used in this patent, by by big spirogram
The remote sensing images that piece size is 512 × 512 are sliced sample z as input and are put into generator to get to according to raw video feature
The road element of generation is distributed sample G (z).It is corresponding to it, the real roads Elemental redistribution manually marked according to sample z
Sample is denoted as x, and two classes distribution sample is indicated with the bianry image for being sliced same size with raw video, wherein with road element
As positive sample, it is expressed as 1 in the picture, using non-rice habitats element as negative sample, is expressed as 0 in the picture.Discriminator D by
One basic convolutional network is constituted, to judge data source in generation sample G (z) or authentic specimen x, final output
One differentiates as a result, the sample data being input in arbiter is the probability of truthful data.Thus one is obtained completely
WGAN model structure, the application flow of the model include training and extraction two parts, and wherein training department's subpackage contains sample set system
Make, three links of sample training and Model Fusion.
2: as other deep learning models, WGAN model needs just be put into reality by a large amount of sample training
It is applied in the scene of border, and makes the necessary condition for meeting that the sample set that model structure requires is progress sample training.Therefore, originally
In invention, for second step, the generator G specifically can be according to following steps training, to be sufficiently carried out WGAN sample pair
It is anti-, obtain feature extraction effect more preferably generator G:
Step s1 carries out slicing treatment according to fixed dimension to sample image, obtains image slice sample z, and to described
Road element in each image slice sample z is marked；
Step s2, constructs generator G and discriminator D, and initialization generates confrontation network V (D, G)；Wherein, the generator
G is a residual error network, and the discriminator D is convolutional network, and the loss function of building generator G isConstruct discriminator D loss function be-((1-t) log (1- (D (G (z)))+
ylogD(x))；Wherein, t=1 represents input as remote sensing images slice, and t=0 represents input as sample image slice；It represents
Convolutional network after rounding exports result；
Step s3, formulating optimization aim isWherein,For majorized function and its optimization direction, Pdata(x) indicate that whole remote sensing images slice x's is distributed as true sample
This, Pz(z) prior distribution of image slice sample z is indicated, E, which is represented, seeks loss function to overall data distribution in training process,
Then
Wherein, PgIt (x) is the distribution of the generator G generation sample obtained；
Step s4 is assessed using Wasserstein distance and is generated sample Pg(x) with authentic specimen Pdata(x) gap；
Wherein, Wasserstein distance W (Pdata(x)',Pg(x)) it is Wherein, Pdata(x) ' indicate Pg(x) and Pdata(x) what is combined is all
Possible Joint Distribution set, unitary sampling (x, G (z))~γ obtain an authentic specimen x and generate sample G (z), | | x-G
(z) | | for the distance between the authentic specimen x and generation sample G (z), E(x, G (z))~γ[| | x-G (z) | |] it is described true
Sample x and the desired value for generating the distance between sample G (z)；
Marked good image slice sample is inputted generator G by step s6, according toThe loss for calculating generator, according to
Calculate the loss of discriminator；
Step s7 carries out the loss that propagated forward operation obtains to generator G in step s6 and carries out BP backpropagation fortune
It calculates；Generator and discriminator are alternately trained respectively, optimize network parameter；
To step s7, training generator and discriminator optimize its network parameter by step s8, repeating said steps s6, until
The generator G and discriminator D reaches Nash Equilibrium, or until the loss of the generator G and discriminator D not
Become, exports generator G at this time as trained generator G.
3: on this basis, the application effect of deep learning model is inseparable with sample training degree, and road is mentioned
For taking target, made training sample concentrates the type comprising topography and geomorphology and category of roads abundanter, the migration of model
Application effect is better, and the training samples number under identical type is more, and the road extraction effect of model is better.But due to remote sensing
Complexity in image, situations such as misidentifying and being not detected, are still inevitable, even if using identical training sample, for not
Same training parameter and model structure, the result optimized also has bigger difference.Different training parameters and model structure
It often cuts both ways, the present invention has particularly formulated the Model Fusion scheme based on WGAN thus:
Refering to what is shown in Fig. 3, the present invention is repeatedly trained by being carried out with the model of different parameters and structure to same sample
To several groups training result, therefrom selects advantageous result and melted according to its initial parameter and structure setting weight
Recombination is closed, fused model is finally obtained.By rational model fusion process, different parameters and model can be sufficiently received
The advantage of structure institute band filters random noise while retaining inevitable road, guarantees the image cleanliness for generating result, to the greatest extent may be used
Can ground improve and generate precision, Maker model namely be applied to the final mask that specific road element extracts task after fusion,
Effect is as shown in Figure 4.
Under a kind of concrete implementation mode of the invention:
Using the part tile in No. two images of Xiamen City's high score in nine scape images as sample data source, extract therein
Roadway characteristic.
With reference to data source tile shown in fig. 5.In order to make model that there is better adaptability, remote sensing image is not carried out
Even color processing.It is that sample slice in each scape image is labeled in ArcGIS software, forms the real roads member of sample
Plain data distribution sample x, i.e. remote sensing images slice, as shown in Figure 6.Meanwhile model is generated in varying environment in order to count respectively
With the extraction accuracy under road element classification, attribute labeling is carried out to road element in x, marks data obtained such as Fig. 7 institute
Show.Parameter therein is specifically as shown in table 1, comprising:
1 label parameters table of table
Later, according to all kinds of topography and geomorphologies of model needs selection covering and 3000 512 × 512 comprising road element
Resolution remote sense image slice is used as training data z.In order to enable model to obtain better training effect, when selecting sample
It should guarantee that positive and negative sample proportion should not differ excessively in break area as far as possible, finally obtain sample data set example such as Fig. 8 institute
Show.
The present invention is trained according to above-mentioned sample as a result, is fought by WGAN sample, acquisition can accurately identify
The generator G of road feature.The purpose of sample training be in order to allow model that can optimize self structure in a large amount of learning processes,
The generator with adaptability and accurate rate is finally obtained to generate road element according to remote sensing image.In reality
In the training process on border, needs respectively to give generator and discriminator intensive training using the sample data set made, allow
The two continues to optimize upgrading in confrontation each other, and detailed process is as follows for model training:
Step 1: setting loss function: loss function generates data and real data gap for assessing in training process
Variation tendency, thus reflect model training optimization whether towards desired target adjustment, with reference first to standard GAN
Model is respectively the training process setting loss function of discriminator and generator respectively as shown in formula 1,2:
-((1-t)log(1-(D(G(z)))+ylog D(x)) (1)
T represents input data source in current optimization process in formula, when input data comes from authentic specimen, i.e. remote sensing
Image slice, Shi You t=1；The output after being rounded is represented as a result, generally taking threshold value is 0.5, i.e., when output result is greater than 0.5
When assert model judge input data for authentic specimen, rounding up is 1, on the contrary assert input data be generation sample, downwards
Being rounded is 0, this two are that setting is because of the anti-of the source and output result for needing consideration input to be distributed in generator training
Feedback.
Step 2: formulate optimization aim: dual training is substantially the zero-sum game of a maximin strategy, and model is wanted
Ask discriminator that can distinguish to input data source.Therefore when input data is authentic specimen, D (x) is the bigger the better, when
Input data is when generating sample, and (D's 1- (G (z)) is the bigger the better, i.e. discriminator loss is the bigger the better, and similarly has generator damage
It loses the smaller the better, then has final training optimization aim as shown in Equation 3:
In formulaRepresenting optimized function and optimization direction, wherein Pdata(x) indicate whole authentic specimen x's
Distribution, that is, the distribution of whole remote sensing images slice x；Pz(z) input image is represented, that is, the priori of image slice sample z point
Cloth；E, which is represented, seeks loss function to overall data distribution in training process, as shown in formula 4,5:
In formula, Pg(x) distribution for representing generation sample can derive theoretically optimal mirror according to trained final purpose
Other device answers shape as shown in Equation 6:
The formula indicates x from generation sample Pg(x) with authentic specimen Pdata(x) relative scale of possibility, ideal shape
Under state, discriminator should can not identify generator sample generated, i.e. Pg(x) and Pdata(x) approximately equal, then generation at this time
Table generator has preferable generation effect, and generally this state is referred to as Nash Equilibrium, there is D (G (z)) ≈ 0.5.
Step 3: being modified according to WGAN model to loss function: in the calculating process to loss function, needing to answer
Sample P is assessed with KL divergence or JS divergence calculation formulag(x) with authentic specimen Pdata(x) gap, and tradition GAN model
In it is some training difficulty on deficiency derive from the two divergence calculation formula itself defect: they will lead to Pg(x) with
Pdata(x) two be distributed in overlapping region itself it is less when, the gradient of generator loss function is fixed as constant, and then influences
Subsequent training process.Therefore the present invention replaces the two divergences by introducing a Wasserstein distance conception, such as formula 7
It is shown:
In formula, Pdata(x) ' indicate Pg(x) and Pdata(x) all possible Joint Distribution set to combine, single are adopted
Sample (x, G (z))~γ can obtain an authentic specimen x and generate sample G (z), pass through | | x-G (z) | | this is calculated to sample
Distance to obtaining desired value E(x, G (z))~γ[| | x-G (z) | |], it is arrived finally by obtaining lower bound to the desired value
Wasserstein distance W (Pdata(x)',Pg(x)), relative to KL and JS divergence, Wasserstein distance does not have in two distributions
There is the gap between still can smoothly reacting when overlapping.
In calculating the loss function of Wasserstein distance applications to two models, prove to need by deriving
Construction one contains parameter w, and the last layer does not have the discriminator network f of nonlinear activation layer iwReplace original model, together
When limitation w be no more than a certain particular range, thus obtain revised discriminator and generator loss function as shown in formula 8,9.
Wherein sigmoid layers, it is responsible on artificial neural network neuron for the input of neuron being mapped to the function of output end, WGAN
In by removing this layer convert recurrence computational problem for last true and false two points of problems:
Step 4: propagated forward calculates loss: the sample training collection marked being inputted network, is distinguished according to formula 8,9
Calculate the loss of generator and discriminator.
Step 5: backpropagation optimizes network: being lost according to propagated forward and carries out BP backpropagation, respectively alternately training life
It grows up to be a useful person and discriminator, optimizes network parameter.
Step 6: repetition training is to reaching optimization aim: step 4, the training process in 5 is repeated, until optimum results connect
Nearly Nash Equilibrium or model loss function gradient no longer change, and discriminator can not identify the data sample of generator output at this time
This true and false, obtains the generator that high-precision road element can be generated from remote sensing image, and test effect is as shown in Figure 9.By
The extracted roadway characteristic of the generator substantially conforms to real roads situation.
The artificial intelligence technologys such as the deep learning that the present invention is based on learn road element by using multitiered network
Multi-level processing is fused in internal model optimization by the feature in remote sensing image different levels, essence and above-mentioned road
The basic thought of extracting method fits., it can be said that depth learning technology, is applied to road feature extraction, is by face
A kind of feature extracting method combined to the extracting method of object with the feature extracting method based on pixel.Technology reduction
The link of manual intervention, takes full advantage of all kinds of elements inside high resolution image.It is supported with enough sample sizes
Under the premise of, Feature Selection Model provided by the present invention can also obtain stronger migration application power simultaneously, cope with well
Deficiency of the conventional method on the pretreatment of image, identification precision.
Those of ordinary skill in the art will appreciate that: the foregoing is only a preferred embodiment of the present invention, and does not have to
In the limitation present invention, although the present invention is described in detail referring to the foregoing embodiments, for those skilled in the art
For, still can to foregoing embodiments record technical solution modify, or to part of technical characteristic into
Row equivalent replacement.All within the spirits and principles of the present invention, any modification, equivalent replacement, improvement and so on should all wrap
Containing within protection scope of the present invention.
Claims (8)
1. the extracting method of roadway characteristic in a kind of high resolution image, which is characterized in that step includes:
The first step obtains remote sensing images, it is carried out slicing treatment according to fixed dimension, remote sensing images is obtained and is sliced x；
Remote sensing images slice is input to progress propagated forward operation in trained generator G, obtained wherein by second step
Roadway characteristic, export the roadway characteristic；
Wherein, in second step, the generator G is obtained according to following steps training:
Step s1 carries out slicing treatment according to fixed dimension to sample image, obtains image slice sample z, and to each figure
As the road element in slice sample z is marked；
Step s2, constructs generator G and discriminator D, and initialization generates confrontation network V (D, G)；Wherein, the generator G is one
A residual error network, the discriminator D are convolutional network, and the loss function of building generator G isConstruct discriminator D loss function be-((1-t) log (1- (D (G (z)))+
ylogD(x))；Wherein, t=1 represents input as remote sensing images slice, and t=0 represents input as sample image slice；Representative takes
Convolutional network after whole exports result；
Step s3, formulating optimization aim isWherein,For majorized function and its optimization direction, Pdata(x) indicate that whole remote sensing images slice x's is distributed as true sample
This, Pz(z) prior distribution of image slice sample z is indicated, E, which is represented, seeks loss function to overall data distribution in training process,
Then
Wherein, PgIt (x) is the distribution of the generator G generation sample obtained；
Step s4 is assessed using Wasserstein distance and is generated sample Pg(x) with authentic specimen Pdata(x) gap；Wherein,
Wasserstein distance W (Pdata(x)',Pg(x)) it isWherein,
Pdata(x) ' indicate Pg(x) and Pdata(x) all possible Joint Distribution set to combine, unitary sampling (x, G (z))~
γ obtains an authentic specimen x and generates sample G (z), | | x-G (z) | | for the authentic specimen x and generate between sample G (z)
Distance, E(x, G (z))~γ[| | x-G (z) | |] is the authentic specimen x and the desired value that the distance between generates sample G (z)；
Marked good image slice sample is inputted generator G by step s6, according toThe loss for calculating generator, according to
Calculate the loss of discriminator；
Step s7 carries out the loss that propagated forward operation obtains to generator G in step s6 and carries out BP backpropagation operation；Respectively
Alternately training generator and discriminator optimize network parameter；
Step s8, repeating said steps s6 optimize its network parameter, until described to step s7, training generator and discriminator
The generator G and discriminator D reaches Nash Equilibrium, or until the loss of the generator G and discriminator D is constant, defeated
Generator G at this time is as trained generator G out.
2. the extracting method of roadway characteristic in high resolution image as described in claim 1, which is characterized in that the step s8
In, it is D (G (z)) ≈ 0.5 that discriminator D, which reaches Nash Equilibrium,.
3. the extracting method of roadway characteristic in high resolution image as described in claim 1, which is characterized in that the second step
In, step s2 is first carried out with different parameters and structure building generator G, discriminator D and generation confrontation network V (D, G) respectively
To the training of step s8, different generator G is obtained；Then advantageous one group is selected from the different generator G
Generator, according to the parameter of this group of generator weight different from structure setting, to this group of generator according to respective weight into
Row fusion recombination ultimately generates device G ' as trained generator G using what is formed after merging, ultimately generates device G ' with this and carry out
Propagated forward operation obtains roadway characteristic.
4. the extracting method of roadway characteristic in high resolution image as claimed in claim 3, which is characterized in that the discriminator
The form of D is
5. the extracting method of roadway characteristic in high resolution image as claimed in claim 3, which is characterized in that in step s1,
The label carried out to the road element in described image slice sample z specifically includes: the width of the serial number of road element, road element
Degree, the material of road element, the affiliated environment of road element.
6. the extracting method of roadway characteristic in high resolution image as claimed in claim 3, which is characterized in that in step s1,
It is close that described image is sliced positive sample and the ratio of negative sample in sample z.
7. the extracting method of roadway characteristic in high resolution image as claimed in claim 3, which is characterized in that described image is cut
Piece sample z or remote sensing images slice are handled without even color.
8. the extracting method of roadway characteristic in high resolution image as claimed in claim 7, which is characterized in that described image is cut
Piece sample z or remote sensing images slice are equal sized bianry image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811532429.XA CN109635748B (en) | 2018-12-14 | 2018-12-14 | Method for extracting road characteristics in high-resolution image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811532429.XA CN109635748B (en) | 2018-12-14 | 2018-12-14 | Method for extracting road characteristics in high-resolution image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109635748A true CN109635748A (en) | 2019-04-16 |
| CN109635748B CN109635748B (en) | 2021-09-03 |
Family
ID=66074043
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811532429.XA Active CN109635748B (en) | 2018-12-14 | 2018-12-14 | Method for extracting road characteristics in high-resolution image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109635748B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110211046A (en) * | 2019-06-03 | 2019-09-06 | 重庆邮电大学 | A kind of remote sensing image fusion method, system and terminal based on generation confrontation network |
| CN110263612A (en) * | 2019-04-25 | 2019-09-20 | 北京工业大学 | Based on the multi-spectral remote sensing image method for extracting roads for generating confrontation network |
| CN110598673A (en) * | 2019-09-24 | 2019-12-20 | 电子科技大学 | Remote sensing image road extraction method based on residual error network |
| CN111192221A (en) * | 2020-01-07 | 2020-05-22 | 中南大学 | Aluminum electrolysis fire hole image repairing method based on deep convolution generation countermeasure network |
| CN112580721A (en) * | 2020-12-19 | 2021-03-30 | 北京联合大学 | Target key point detection method based on multi-resolution feature fusion |
| 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 |
| CN112733756A (en) * | 2021-01-15 | 2021-04-30 | 成都大学 | Remote sensing image semantic segmentation method based on W divergence countermeasure network |
| CN112734849A (en) * | 2021-01-18 | 2021-04-30 | 上海市城市建设设计研究总院（集团）有限公司 | Computer-based urban road network intersection angle detection method |
| CN112906459A (en) * | 2021-01-11 | 2021-06-04 | 甘肃省公路局 | Road network checking technology based on high-resolution remote sensing image and deep learning method |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101915990A (en) * | 2009-04-02 | 2010-12-15 | 通用汽车环球科技运作公司 | Enhancing road vision on the full-windscreen head-up display |
| US20150148448A1 (en) * | 2013-11-27 | 2015-05-28 | The Goodyear Tire & Rubber Company | Pneumatic tire |
| CN107563510A (en) * | 2017-08-14 | 2018-01-09 | 华南理工大学 | A kind of WGAN model methods based on depth convolutional neural networks |
| CN107909621A (en) * | 2017-11-16 | 2018-04-13 | 深圳市唯特视科技有限公司 | It is a kind of based on it is twin into confrontation network medical image synthetic method |
| CN108596141A (en) * | 2018-05-08 | 2018-09-28 | 深圳大学 | A kind of depth network generates the detection method and system of facial image |
| US20180307947A1 (en) * | 2017-04-25 | 2018-10-25 | Nec Laboratories America, Inc. | Cyclic generative adversarial network for unsupervised cross-domain image generation |
| CN108763857A (en) * | 2018-05-29 | 2018-11-06 | 浙江工业大学 | A kind of process soft-measuring modeling method generating confrontation network based on similarity |
| CN108764173A (en) * | 2018-05-31 | 2018-11-06 | 西安电子科技大学 | The hyperspectral image classification method of confrontation network is generated based on multiclass |
| CN108805188A (en) * | 2018-05-29 | 2018-11-13 | 徐州工程学院 | A kind of feature based recalibration generates the image classification method of confrontation network |
| CN108830209A (en) * | 2018-06-08 | 2018-11-16 | 西安电子科技大学 | Based on the remote sensing images method for extracting roads for generating confrontation network |
| CN108922518A (en) * | 2018-07-18 | 2018-11-30 | 苏州思必驰信息科技有限公司 | voice data amplification method and system |
-
2018
- 2018-12-14 CN CN201811532429.XA patent/CN109635748B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101915990A (en) * | 2009-04-02 | 2010-12-15 | 通用汽车环球科技运作公司 | Enhancing road vision on the full-windscreen head-up display |
| US20150148448A1 (en) * | 2013-11-27 | 2015-05-28 | The Goodyear Tire & Rubber Company | Pneumatic tire |
| US20180307947A1 (en) * | 2017-04-25 | 2018-10-25 | Nec Laboratories America, Inc. | Cyclic generative adversarial network for unsupervised cross-domain image generation |
| CN107563510A (en) * | 2017-08-14 | 2018-01-09 | 华南理工大学 | A kind of WGAN model methods based on depth convolutional neural networks |
| CN107909621A (en) * | 2017-11-16 | 2018-04-13 | 深圳市唯特视科技有限公司 | It is a kind of based on it is twin into confrontation network medical image synthetic method |
| CN108596141A (en) * | 2018-05-08 | 2018-09-28 | 深圳大学 | A kind of depth network generates the detection method and system of facial image |
| CN108763857A (en) * | 2018-05-29 | 2018-11-06 | 浙江工业大学 | A kind of process soft-measuring modeling method generating confrontation network based on similarity |
| CN108805188A (en) * | 2018-05-29 | 2018-11-13 | 徐州工程学院 | A kind of feature based recalibration generates the image classification method of confrontation network |
| CN108764173A (en) * | 2018-05-31 | 2018-11-06 | 西安电子科技大学 | The hyperspectral image classification method of confrontation network is generated based on multiclass |
| CN108830209A (en) * | 2018-06-08 | 2018-11-16 | 西安电子科技大学 | Based on the remote sensing images method for extracting roads for generating confrontation network |
| CN108922518A (en) * | 2018-07-18 | 2018-11-30 | 苏州思必驰信息科技有限公司 | voice data amplification method and system |
Non-Patent Citations (3)
| Title |
|---|
| MARTIN ARJOVSKY.ET.: "Wasserstein GAN", 《ARXIV 2017》 * |
| 王双: "高分辨率遥感影像道路提取方法研究", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》 * |
| 肖旭: "基于深度学习的图像语义分割研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110263612A (en) * | 2019-04-25 | 2019-09-20 | 北京工业大学 | Based on the multi-spectral remote sensing image method for extracting roads for generating confrontation network |
| CN110211046A (en) * | 2019-06-03 | 2019-09-06 | 重庆邮电大学 | A kind of remote sensing image fusion method, system and terminal based on generation confrontation network |
| CN110211046B (en) * | 2019-06-03 | 2023-07-14 | 重庆邮电大学 | Remote sensing image fusion method, system and terminal based on generation countermeasure network |
| CN110598673A (en) * | 2019-09-24 | 2019-12-20 | 电子科技大学 | Remote sensing image road extraction method based on residual error network |
| CN111192221A (en) * | 2020-01-07 | 2020-05-22 | 中南大学 | Aluminum electrolysis fire hole image repairing method based on deep convolution generation countermeasure network |
| CN111192221B (en) * | 2020-01-07 | 2024-04-16 | 中南大学 | Aluminum electrolysis fire hole image repairing method based on deep convolution generation countermeasure network |
| CN112580721B (en) * | 2020-12-19 | 2023-10-24 | 北京联合大学 | Target key point detection method based on multi-resolution feature fusion |
| CN112580721A (en) * | 2020-12-19 | 2021-03-30 | 北京联合大学 | Target key point detection method based on multi-resolution feature fusion |
| 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 |
| 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 |
| CN112906459A (en) * | 2021-01-11 | 2021-06-04 | 甘肃省公路局 | Road network checking technology based on high-resolution remote sensing image and deep learning method |
| CN112733756A (en) * | 2021-01-15 | 2021-04-30 | 成都大学 | Remote sensing image semantic segmentation method based on W divergence countermeasure network |
| CN112733756B (en) * | 2021-01-15 | 2023-01-20 | 成都大学 | Remote sensing image semantic segmentation method based on W divergence countermeasure network |
| CN112734849A (en) * | 2021-01-18 | 2021-04-30 | 上海市城市建设设计研究总院（集团）有限公司 | Computer-based urban road network intersection angle detection method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109635748B (en) | 2021-09-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109635748A (en) | The extracting method of roadway characteristic in high resolution image | |
| CN106877237B (en) | A method of insulator lacks in the detection transmission line of electricity based on Aerial Images | |
| CN109376605B (en) | Electric power inspection image bird-stab-prevention fault detection method | |
| CN106356757A (en) | Method for inspecting electric power lines by aid of unmanned aerial vehicle on basis of human vision characteristics | |
| CN109977921A (en) | A kind of transmission line of electricity perils detecting method | |
| CN106127815A (en) | A kind of tracking merging convolutional neural networks and system | |
| CN106203498B (en) | Urban scene garbage detection method and system | |
| CN109902677A (en) | A kind of vehicle checking method based on deep learning | |
| CN110163213B (en) | Remote sensing image segmentation method based on disparity map and multi-scale depth network model | |
| CN108573222A (en) | The pedestrian image occlusion detection method for generating network is fought based on cycle | |
| CN106127204A (en) | A kind of multi-direction meter reading Region detection algorithms of full convolutional neural networks | |
| CN110084165A (en) | The intelligent recognition and method for early warning of anomalous event under the open scene of power domain based on edge calculations | |
| CN106909886B (en) | A kind of high-precision method for traffic sign detection and system based on deep learning | |
| CN108776772A (en) | Across the time building variation detection modeling method of one kind and detection device, method and storage medium | |
| CN105258704A (en) | Multi-scale space-time hot point path detection method based on rapid road network modeling | |
| CN113160062B (en) | Infrared image target detection method, device, equipment and storage medium | |
| CN103729620B (en) | A kind of multi-view pedestrian detection method based on multi-view Bayesian network | |
| CN105426908A (en) | Convolutional neural network based substation attribute classification method | |
| CN108305260A (en) | Detection method, device and the equipment of angle point in a kind of image | |
| CN111723657A (en) | River foreign matter detection method and device based on YOLOv3 and self-optimization | |
| CN107945210A (en) | Target tracking algorism based on deep learning and environment self-adaption | |
| CN110084181A (en) | A kind of remote sensing images Ship Target Detection method based on sparse MobileNetV2 network | |
| CN109035223A (en) | A kind of intelligent evaluation method for satellite remote sensing images availability | |
| CN105303258B (en) | Solar burst event prediction method based on machine learning techniques forecasting model | |
| CN114240796A (en) | Remote sensing image cloud and fog removing method and device based on GAN and storage medium |
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 |