CN108629297A - A kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics - Google Patents
A kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics Download PDFInfo
- Publication number
- CN108629297A CN108629297A CN201810352969.3A CN201810352969A CN108629297A CN 108629297 A CN108629297 A CN 108629297A CN 201810352969 A CN201810352969 A CN 201810352969A CN 108629297 A CN108629297 A CN 108629297A
- Authority
- CN
- China
- Prior art keywords
- super
- pixel block
- pixel
- waveform parameter
- cloud
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- 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
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
- G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
- G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
Abstract
The present invention provides a kind of remote sensing images cloud detection method of optic counted based on spatial domain natural scene, and since cloud sector domain is relatively flat, the value of asymmetric generalized Gaussian distribution concentrates near zero, so the asymmetric generalized Gaussian distribution approximation kurtosis in cloud sector domain;And city high-brightness region, it is more dispersed, so asymmetric generalized Gaussian distribution is than shallower;Then when the asymmetric generalized Gaussian distribution of super-pixel block to be measured is consistent with the asymmetric generalized Gaussian distribution of cloud atlas picture, i.e. waveform parameter f1With waveform parameter f2The distance between, it is less than waveform parameter f1With waveform parameter f3The distance between, then super-pixel block to be measured is judged as the regions Yun Xue, otherwise it is judged as that non-cloud avenges region;The method first Application of this statistical parameter distance can accurately extract in cloud sector domain in cloud detection from visible light wave range remote sensing images.
Description
Technical field
The invention belongs to technical field of remote sensing image processing more particularly to a kind of remote sensing based on spatial domain natural scene statistics
Image cloud detection method of optic.
Background technology
Remote sensing be grow up the sixties in last century it is close with the science and technology such as electronics, space, computer, optics, geography
Cut relevant one emerging interdisciplinary science technology.It is not only a very active subject in current information field, and
One important component of current science and technology.With resource three, high score No.1, high score two for representative, domestic optics is distant
Sense satellite items design objective has progressivelyed reach international most advanced level, and earth observation systems are gradually improved, the data of satellite image
Amount increases sharply, and Market Orientation improves year by year.However, and not all remote sensing image can meet image information intelligent processing
Requirement, one of them critically important factor is exactly the cloud cover on image.In the case of having a large amount of cloud covers, satellite
Imaging device cannot be correctly received the reflective character from ground target, and corresponding visible light satellite image data includes
Ground target available information is greatly decreased, and is produced to being subsequently based on the processing such as image classification, target detection and identification and application
Greatly negative effect, reduces the use value of visible light satellite remote sensing images.In conclusion cloud detection is visible light satellite
One of top priority and key technology of image processing application.
Present cloud detection algorithm can be divided into two classes:The cloud detection algorithm of different time and the cloud detection of same time are calculated
Method.The cloud detection of different time, the principle that this kind of algorithm mainly follows are the reflectivity and temperature of remote sensing atural object in a short time
Variation is little, 2013, and SumingJin proposes the image using the same place in two days, to compare the difference before them,
To detect the position of cloud, Nicholas R.Goodwin by extract spectrum and the texture feature information of Different climate image come
It is compared with image to be detected, to realize the cloud detection algorithm of image.2014, Zhe Zhu proposed to use several not
With the remote sensing images of time, termporal filter more than one is established, to realize the automatic detection in remote sensing image clouds region.However
Method mentioned above, computation complexity is high, and testing result depends critically upon the effect of registration, is not easily applied in practice on satellite
Face.The cloud detection algorithm of same time, the principle that this kind of algorithm mainly follows are that cloud is typically that and temperature brighter than ground is lower
, this kind of algorithm can be divided into two kinds again, and one is the algorithm based on multichannel image characteristic, another kind is logical based on visible light
The algorithm of road picture characteristics.The algorithm of multichannel image characteristic needs that temperature threshold is arranged using Medium wave infrared channel, to use up
Possible differentiation cloud and other regions, but which limits the satellite moneys of the application environment of such algorithm, such as China's transmitting
Source three, high score No.1, high score two all do not have medium-wave infrared wave band, so above-mentioned algorithm all fails.Another algorithm
It needs, extracts visible light image information, such as H.K.Chethan proposed to extract cloud using Gabor filter in 2009
Texture information, and classified to the image of cloud by SVM.In 2015, Yi Yuan propose using super-pixel come to image into
Row segmentation, and the result of segmentation classifies to image by SVM, obtain final cloud detection result.Cloud detection algorithm
Method is numerous, but still has following difficult point needs to focus on solving at present,
1) fractus various shapes bring challenge to the cloud detection method of optic based on texture and structure;
2) cloud has the differentiation of thin and thick, simple Threshold segmentation that the detection of Bao Yun can be caused to lose;
3) interference of similar atural object, such as highlighted building, snow can all be affected to cloud detection.
Invention content
To solve the above problems, the present invention provides a kind of remote sensing images cloud detection side counted based on spatial domain natural scene
Method can accurately extract in cloud sector domain from visible light wave range remote sensing images.
A kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics, includes the following steps:
S1:It will be seen that optical band remote sensing images carry out super-pixel segmentation, super-pixel block is obtained;
S2:The waveform parameter f of the corresponding asymmetric generalized Gaussian distribution of each super-pixel block is obtained respectively1It is corresponding with cloud atlas picture
Asymmetric generalized Gaussian distribution waveform parameter f2, the corresponding asymmetric generalized Gaussian distribution of cloudless image waveform parameter f3
The distance between, if waveform parameter f1With waveform parameter f2The distance between, it is less than waveform parameter f1With waveform parameter f3Between
Distance, then waveform parameter f1Corresponding super-pixel block is the regions Yun Xue, otherwise avenges region for non-cloud;
S3:Obtain the Gabor characteristic of the corresponding super-pixel block in the regions Yun Xue;
S4:The Gabor characteristic is inputted in support vector machines and is judged, the detection to cloud sector domain is completed;Its
In, the support vector machines train to obtain according to the Gabor characteristic and the Gabor characteristic of snow image of cloud atlas picture.
Optionally, described to will be seen that optical band remote sensing images carry out super-pixel segmentation and are specially:
Super-pixel segmentation is carried out to visible light wave range remote sensing images using simple linear iteration clustering procedure, wherein simple
During the enhancing connectivity of linear iteraction clustering procedure, if the pixel mean value of super-pixel block of the area less than given threshold T2
More than given threshold T3, then super-pixel block of the area less than given threshold T2 is not incorporated in the area nearest with it not less than setting threshold
The super-pixel block of value T2;If the pixel mean value of super-pixel block of the area less than given threshold T2 is not more than given threshold T3,
Super-pixel block of the area less than given threshold T2 is incorporated to the super-pixel block that the area nearest with it is not less than given threshold T2.
Optionally, it will be seen that optical band remote sensing images progress super-pixel segmentation, after obtaining super-pixel block, it is small to reject area
Step S2 and step S3 is executed in the super-pixel block of given threshold T4, then by remaining super-pixel block.
Optionally, the waveform parameter f for obtaining the corresponding asymmetric generalized Gaussian distribution of each super-pixel block1Specially:
Obtain pixel value, local mean value and the Local standard deviation of each pixel of super-pixel block;
According to the pixel value of each pixel, local mean value and Local standard deviation, the inequality contrast of each pixel is obtained
Normalization coefficient;
According to the inequality contrast normalization coefficient of each pixel, the corresponding inequality contrast normalizing of each super-pixel block is obtained
Change coefficient matrix;
According to the inequality contrast normalization coefficient matrix, the corresponding asymmetric Generalized Gaussian point of each super-pixel block is obtained
Cloth, and obtain the waveform parameter f of asymmetric generalized Gaussian distribution1。
Optionally, the waveform parameter f1With waveform parameter f2The distance between acquisition methods be:
Obtain the waveform parameter of the corresponding asymmetric generalized Gaussian distribution of super-pixel block, wherein the waveform parameter includes
Waveform halfwidth, left offset, right offset and mean value;
By pixel centered on each pixel of super-pixel block, then the inequality contrast of each central pixel point is normalized into system
Number is compared with the inequality of the horizontal direction of central pixel point, vertical direction, diagonally opposed and dihedral direction pixel respectively
Normalization coefficient is spent to be multiplied, to obtain super-pixel block in the horizontal direction, vertical direction, diagonally opposed and dihedral direction correspond to
Inequality contrast normalization coefficient matrix;
According to super-pixel block in the horizontal direction, vertical direction, the corresponding inequality contrast in diagonally opposed and dihedral direction
Normalization coefficient matrix, obtain respectively super-pixel block in the horizontal direction, vertical direction, diagonally opposed and dihedral direction it is corresponding
Asymmetric generalized Gaussian distribution;
Obtain respectively super-pixel block in the horizontal direction, vertical direction, diagonally opposed and dihedral direction it is corresponding asymmetric
The waveform parameter of generalized Gaussian distribution;
Obtain the corresponding covariance matrix of 20 characteristic values of super-pixel block;Wherein, 20 characteristic values are respectively super picture
The waveform parameter and super-pixel block of the corresponding asymmetric generalized Gaussian distribution of plain block in the horizontal direction, vertical direction, diagonal side
To the waveform parameter of, the corresponding asymmetric generalized Gaussian distribution in dihedral direction;
According to the corresponding covariance matrix of 20 characteristic values, the cloud atlas picture of the 20 of super-pixel block characteristic values, super-pixel block
20 characteristic values, cloud atlas picture the corresponding covariance matrix of 20 characteristic values, obtain the corresponding asymmetric broad sense of super-pixel block
The waveform parameter f of Gaussian Profile1With cloud atlas as the waveform parameter f of corresponding asymmetric generalized Gaussian distribution2The distance between.
Optionally, the waveform parameter f1With waveform parameter f3The distance between acquisition methods be:
According to the 20 of super-pixel block characteristic values, 20 spies of the corresponding covariance matrix of 20 characteristic values, non-cloud image
Value indicative, the corresponding covariance matrix of 20 characteristic values obtain the waveform ginseng of the corresponding asymmetric generalized Gaussian distribution of super-pixel block
Number f1The waveform parameter f of asymmetric generalized Gaussian distribution corresponding with non-cloud image3The distance between.
Optionally, the Gabor characteristic for obtaining the corresponding super-pixel block in the regions Yun Xue is specially:
Side of more than two angles as the kernel function of Gabor filter is randomly selected in 0 °~360 ° of range
To;
It is that the kernel function determines more than two SIN function wavelength based on experience value, wherein the SIN function wavelength
Determine the frequency of the kernel function;
According to the direction and frequency, the corresponding super-pixel block in the regions Yun Xue is obtained on different directions and frequency
Gabor values;
The average value for obtaining multiple Gabor values, using the average value as the corresponding super-pixel block in the regions Yun Xue
Gabor characteristic.
Advantageous effect:
1, the present invention provides a kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics, due to cloud sector domain ratio
Flatter, the value of asymmetric generalized Gaussian distribution concentrates near zero, so the asymmetric generalized Gaussian distribution in cloud sector domain is approximate
Kurtosis;And city high-brightness region, it is more dispersed, so asymmetric generalized Gaussian distribution is than shallower;Then when to be measured super
The asymmetric generalized Gaussian distribution of block of pixels is consistent with the asymmetric generalized Gaussian distribution of cloud atlas picture, i.e. waveform parameter f1With waveform
Parameter f2The distance between, it is less than waveform parameter f1With waveform parameter f3The distance between, then super-pixel block to be measured is judged as cloud
Region is avenged, is otherwise judged as that non-cloud avenges region;The method first Application of this statistical parameter distance, can be accurately in cloud detection
Cloud sector domain is extracted from visible light wave range remote sensing images;
Since cloud is to float in the air, being formed by the steam constantly polymerization vaporized, then texture transition is smooth around cloud,
And it is to be close to that transition around earth's surface is apparent, and texture is larger to avenge;The grain direction of the cloud of vaporization does not have rule simultaneously, and due to by ground
The influence of face texture, snow zone-texture have apparent directionality.Based on this two attributes, the present invention proposes special using Gabor
It levies come into the differentiation in snow region of racking, Gabor characteristic is obtained by Gabor filter, and Gabor filter can be extracted well again
The texture information of specific direction, therefore the texture that can be used for distinguishing the edges Yun Xue is strong and weak;The last present invention is again by SVM come area
The other cloud sector regions Yu Hexue so that the difference of the invention to cloud snow is apparent, and cloud detection rate is high, meets the in-orbit cloud of satellite system and comments
The problem of estimating..
2, the present invention carries out super-pixel point using simple linear iteration cluster (SLIC) to visible light wave range remote sensing images
It cuts, small area region can be merged into close region, fractus can thus be merged by traditional SLIC in order to ensure to divide quality
To other regions;And the present invention faces during SLIC enhances connectivity whether small area region is merged into large area
Near field is judged, if the pixel mean value of the super-pixel block of small area is more than the threshold value set, the small area of high brightness
Super-pixel block nonjoinder thus remains fractus to other regions, improves the result accuracy of segmentation.
3, the Gabor values of the present invention extraction different directions and frequency, using the average value of multiple Gabor values as Gabor spies
Sign, so that Yun Xue is distinguished in region in different directions, can further increase cloud detection rate.
Description of the drawings
Fig. 1 is a kind of flow of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics provided by the invention
Figure;
Fig. 2 is a kind of flow effect of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics provided by the invention
Fruit is schemed;
Fig. 3 is the inequality contrast normalization coefficient distribution curve of cloud atlas picture provided by the invention, cloudless image;
Fig. 4 be central pixel point provided by the invention with it is horizontal, vertical, diagonally and the pixel in dihedral direction shows
It is intended to;
Fig. 5 (a) is visible light wave range cloud sector provided by the invention domain remote sensing images schematic diagram;
Fig. 5 (b) is that visible light wave range provided by the invention avenges regional remote sensing image schematic diagram;
Fig. 6 (a) is that remote sensing images schematic diagram corresponding gradient in visible light wave range cloud sector provided by the invention domain counts histogram
Figure;
Fig. 6 (b) is that visible light wave range provided by the invention avenges the corresponding gradient statistics histogram of regional remote sensing image schematic diagram
Figure;
Fig. 7 (a) is visible light wave range remote sensing images schematic diagram provided by the invention;
Fig. 7 (b) is the corresponding Gabor characteristic schematic diagram of visible light wave range remote sensing images schematic diagram provided by the invention.
Specific implementation mode
In order to make those skilled in the art more fully understand application scheme, below in conjunction in the embodiment of the present application
Attached drawing, technical solutions in the embodiments of the present application are clearly and completely described.
Referring to Fig. 1, which is a kind of remote sensing images cloud detection counted based on spatial domain natural scene provided in this embodiment
The flow chart of method.Fig. 2 is a kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics provided in this embodiment
Flow design sketch.
A kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics, includes the following steps:
S1:It will be seen that optical band remote sensing images carry out super-pixel segmentation, super-pixel block is obtained.
It is described below and will be seen that optical band remote sensing images carry out super-pixel segmentation using simple linear iteration clustering procedure SLIC
Method, include the following steps:
S101:Initialization seed point (cluster centre):According to the super-pixel block number of setting, in visible light wave range remote sensing figure
As interior uniform distribution seed point.Assuming that a total of N number of pixel of visible light wave range remote sensing images, pre-segmentation is K identical rulers
Very little super-pixel block, then the size of each super-pixel block is N/K, then the distance (step-length) of neighboring seeds point is approximately S=
Sqrt (N/K) (dividing exactly automatically).
S102:Seed point (generally taking n=3) is reselected in n × n neighborhoods of seed point.Specific method is:Calculating should
Seed point is moved on to the place of the neighborhood manhole ladder angle value minimum by the Grad of all pixels point in neighborhood.The purpose for the arrangement is that
In order to avoid seed point is fallen in the larger profile and border of gradient, in order to avoid influence follow-up Clustering Effect.
S103:It is which each pixel initialization distribution class label (belongs in the neighborhood around each seed point
Cluster centre).Different, the present embodiment is searched in whole visible light wave range remote sensing images from the k-means clustering methods of standard
The search range of the simple linear iteration clustering procedure SLIC of use is limited to 2S × 2S, can be restrained with accelerating algorithm.
S104:Distance metric.Each pixel searched is calculated separately including color distance and space length
The distance of it and the seed point.Distance calculating method is as follows:
Wherein, it is seen that the colouring information of optical band remote sensing images u, v are (lu, au, bu), (lv, av, bv), lab color spaces
It is by the space after rgb space linear transformation, such to handle the extraction for being conducive to colouring information, spatial positional information is (xu,
yu), (xv, yv), dcRepresent color distance, dsRepresent space length, NsIt is maximum space distance in class, is defined as Ns=S=
Sqrt (N/K) is suitable for each cluster.Maximum color distance Nc, it is different and different with visible light wave range remote sensing images, also with
Cluster is different and different, so the present invention takes a fixed constant m, (value range [Isosorbide-5-Nitrae 0] generally takes and 10) replaces.Final
Distance metric D ' are as follows:
Since each pixel can be searched by multiple seed points, thus each pixel can there are one with surrounding kind
The distance of son point, is minimized cluster centre of the corresponding seed point as the pixel.
S105:Iteration optimization.Theoretically the continuous iteration of above-mentioned steps (can be understood as each pixel until error convergence
Until cluster centre no longer changes), practice finds that 10 iteration can to most visible light wave range remote sensing images
More satisfactory effect is obtained, so general iterations take 10.
S106:Enhance connectivity.It is likely to occur following flaw by above-mentioned iteration optimization:There is more connection situations, super picture
Plain block size is too small, and single super-pixel block is cut into multiple discontinuous super-pixel etc., these situations can be connected to by enhancing
Property solve, judge this area size whether with peripheral region together with, if be not connected to, be directly merged on other regions,
On the basis of this merges, present invention addition is following to be judged, if area is equal less than the pixel of the super-pixel block of given threshold T2
Value is more than given threshold T3, then super-pixel block of the area less than given threshold T2 is not incorporated in the area nearest with it not less than setting
The super-pixel block of threshold value T2;If the pixel mean value of super-pixel block of the area less than given threshold T2 is not more than given threshold T3,
Then super-pixel block of the area less than given threshold T2 is incorporated to the super-pixel block that the area nearest with it is not less than given threshold T2, from
And retain the detection to fractus.
Optionally, it will be seen that optical band remote sensing images progress super-pixel segmentation, after obtaining super-pixel block not of uniform size,
The super-pixel block less than given threshold T4 is rejected, then remaining super-pixel block is executed into step S2 and step S3.It can be seen that surplus
Remaining super-pixel block size is not quite similar.
Optionally, divided using Otsu threshold and obtain given threshold T4, included the following steps:
Wherein, T is Otsu threshold, pigIt is tonal gradation igIt is general
Rate value, g=0,1, last given threshold T4 is defined as follows:
S2:The waveform parameter f of the corresponding asymmetric generalized Gaussian distribution of each super-pixel block is obtained respectively1It is corresponding with cloud atlas picture
Asymmetric generalized Gaussian distribution waveform parameter f2, the corresponding asymmetric generalized Gaussian distribution of cloudless image waveform parameter f3
The distance between, if waveform parameter f1With waveform parameter f2The distance between, it is less than waveform parameter f1With waveform parameter f3Between
Distance, then waveform parameter f1Corresponding super-pixel block is the regions Yun Xue, otherwise avenges region for non-cloud.
It should be noted that cloud atlas picture and cloudless image can be provided by known image library, and the known image library provides
Cloud atlas picture and cloudless image be by calibration image.
Optionally, the acquisition methods of the corresponding asymmetric generalized Gaussian distribution of each super-pixel block are:
S201:Obtain pixel value, local mean value and the Local standard deviation of each pixel of super-pixel block.
Specifically, the local mean value μ (i, j) of each pixel and Local standard deviation σ (i, j) computational methods are respectively:
Wherein, i, j are the positions of pixel, if the size of super-pixel block is M0×N0, then have i ∈ { 1,2 ..., M0, j
∈{1,2,…,N0, ω={ ωk,l| k=-K ..., K, l=-L ..., L } be a two-dimensional symmetric unit Gauss weight core,
General setting K=L, value is depending on super-pixel block size.Commonly it is set as K=L=3.
S202:According to the pixel value of each pixel, local mean value and Local standard deviation, the inequality pair of each pixel is obtained
Than degree normalization coefficient.
Specifically, the computational methods of the inequality contrast normalization coefficient MSCN of each pixel are:
Wherein, I (i, j) indicates the pixel value of normalization preceding pixel point,Indicate the pixel of pixel after normalizing
Value, i.e. inequality contrast normalization coefficient M SCN.
Referring to Fig. 3, which is the inequality contrast normalization coefficient distribution of cloud atlas picture, cloudless image provided in this embodiment
Curve.Since the regions Yun Xue are relatively flat, after the normalization of inequality contrast, the value in region concentrates near zero, so Yun Xuequ
The approximate kurtosis of inequality contrast normalization coefficient distribution in domain, such as dotted line in Fig. 3, city high-brightness region, i.e. non-cloud snow
Region is more dispersed, and inequality contrast normalization coefficient is distributed approximate Gaussian state, as solid line circle is distributed in Fig. 3.Thus may be used
Know, the corresponding distribution curve in the regions Yun Xue is higher, and the corresponding distribution curve in non-cloud snow region is short, and numerical value is more dispersed.When waiting for
The inequality contrast normalization coefficient distribution for surveying super-pixel block is consistent with the distribution in the regions Yun Xue, we are judged as the regions Yun Xue
Super-pixel block, be otherwise judged as non-cloud avenge super-pixel block.
S203:According to the inequality contrast normalization coefficient of each pixel, the corresponding inequality comparison of each super-pixel block is obtained
Spend normalization coefficient matrix.
S204:According to the inequality contrast normalization coefficient matrix, the corresponding asymmetric broad sense of each super-pixel block is obtained
Gaussian Profile.
Specifically, asymmetric generalized Gaussian distributionFor:
Wherein, γ is the waveform halfwidth (" fat or thin " of waveform) of asymmetric generalized Gaussian distribution, βl,βrIt is waveform respectively
Left offset, right offset can solve to obtain by the method for match by moment.Meanwhile the mean value η of distribution is also to asymmetric broad sense
Gaussian Profile has an impact, therefore also one of the feature as the modeling of asymmetric generalized Gaussian distribution:
The inequality contrast normalization coefficient distribution in the regions Yun Xue known to distribution curve as shown in Figure 3 is not a rule
Gaussian Profile.In order to calculate the deviation of distribution, need to close on coefficient product to horizontal, vertical, diagonal and dihedral direction, then
The result of product is made to the modeling of asymmetric generalized Gaussian distribution, the structural relation between super-pixel block pixel can thus be carried out
Characterization.Referring to Fig. 4, which is central pixel point provided in this embodiment and its horizontal, vertical, diagonal and dihedral direction pixel
The schematic diagram of point.As shown in Figure 4, central pixel point horizontal, vertical, diagonal and dihedral direction closes on coefficient product difference with it
For:With
Central pixel point is corresponding with the inequality contrast normalization coefficient on its four direction to be multiplied, and is to point all in super-pixel block
It is operated.Therefore, each super-pixel block can obtain 20 characteristic values.Based on the above theoretical foundation, it is described in detail below
Waveform parameter f1With waveform parameter f2, waveform parameter f1With waveform parameter f3The distance between acquisition methods.
Optionally, the waveform parameter f1With waveform parameter f2, waveform parameter f1With waveform parameter f3The distance between obtain
The method is taken to be:
S205:Obtain the waveform parameter of the corresponding asymmetric generalized Gaussian distribution of super-pixel block, wherein the waveform parameter
Including waveform halfwidth, left offset, right offset and mean value.
S206:By pixel centered on each pixel of super-pixel block, then the inequality contrast of each central pixel point is returned
One change coefficient respectively with the horizontal direction of central pixel point, vertical direction, diagonally opposed and dihedral direction pixel it is equal
Poor contrast normalization coefficient is multiplied, to obtain super-pixel block in the horizontal direction, vertical direction, diagonally opposed and dihedral side
To corresponding inequality contrast normalization coefficient matrix.
S207:According to super-pixel block in the horizontal direction, vertical direction, the diagonally opposed and corresponding inequality pair in dihedral direction
Than degree normalization coefficient matrix, obtain respectively super-pixel block in the horizontal direction, vertical direction, diagonally opposed and dihedral direction pair
The asymmetric generalized Gaussian distribution answered.
S208:Obtain respectively super-pixel block in the horizontal direction, vertical direction, diagonally opposed and dihedral direction it is corresponding non-
Waveform halfwidth, left offset, right offset and the mean value of symmetrical generalized Gaussian distribution.
S209:Obtain the corresponding covariance matrix of 20 characteristic values of super-pixel block;Wherein, 20 characteristic values are respectively
Super-pixel block and super-pixel block are in the horizontal direction, vertical direction, diagonally opposed, the corresponding asymmetric broad sense in dihedral direction is high
Waveform halfwidth, left offset, right offset and the mean value of this distribution.
S210:According to the corresponding covariance matrix of 20 characteristic values, the cloud of the 20 of super-pixel block characteristic values, super-pixel block
It is corresponding asymmetric to obtain super-pixel block for 20 characteristic values of image, the corresponding covariance matrix of 20 characteristic values of cloud atlas picture
The waveform parameter f of generalized Gaussian distribution1With cloud atlas as the waveform parameter f of corresponding asymmetric generalized Gaussian distribution2Between away from
From.
S211:According to the 20 of super-pixel block characteristic values, the corresponding covariance matrix of 20 characteristic values, non-cloud image 20
A characteristic value, the corresponding covariance matrix of 20 characteristic values obtain the wave of the corresponding asymmetric generalized Gaussian distribution of super-pixel block
Shape parameter f1The waveform parameter f of asymmetric generalized Gaussian distribution corresponding with non-cloud image3The distance between.
For example, first finding out the regions Yun Xue, the i.e. feature of cloud atlas picture in known image library apart from acquisition methods based on above-mentioned
v1The covariance matrix ∑ of (20 features), cloud characteristics of image1, non-cloud snow region, i.e. the feature v of non-cloud image2, non-cloud image
The covariance matrix ∑ of feature2, then seek the feature v of super-pixel block to be detected3, the covariance matrix ∑ of super-pixel block feature3,
Be brought into it is following seek range formula, if D1<D2, wherein D1For waveform parameter f1With waveform parameter f2The distance between, D2For
Waveform parameter f1With waveform parameter f3The distance between, then super-pixel block to be detected is judged for the regions Yun Xue, is otherwise avenged for non-cloud
Region.
S3:Obtain the Gabor characteristic of the corresponding super-pixel block in the regions Yun Xue.
It should be noted that Gabor characteristic is generated by Gabor filter, it is seen that optical band remote sensing images pass through
It will produce a Gabor value after Gabor filter, which is Gabor characteristic.Gabor filter is a kind of highly important
Directional filters are particularly suitable for the stronger edge of detection direction in two dimensional image texture analysis.Gabor filter has
Direction and the adjustable property of frequency domain, kernel function g (x, y;λ, θ, ψ, σ, ε) it is as follows:
X '=xcos θ+ysin θ
Y '=- xsin θ+ycos θ
Wherein λ is SIN function wavelength, determines the frequency of Gabor kernel functions, θ is the direction of Gabor kernel functions, and ψ is phase
Position offset, the standard deviation of σ Gaussian functions, ε is the ratio of width to height in space.It can be seen that by by the corresponding super-pixel in the regions Yun Xue
Different directions and SIN function wavelength is arranged by Gabor filter, and for Gabor filter kernel function in block, then can obtain
Take Gabor value of the corresponding super-pixel block in the regions Yun Xue on different directions and frequency.Then being averaged for multiple Gabor values is obtained
Value, using the average value as the Gabor characteristic of the corresponding super-pixel block in the regions Yun Xue.
Optionally, in the present embodiment, the direction θ of Gabor kernel functions is specially with SIN function wavelength X:
θ∈{0°,45°,90°,135°,180°,225°,270°,315°},λ∈{2,4}
It should be noted that the direction of Gabor kernel functions in addition to use above-mentioned setting method, can also be at 0 °~360 °
More than two angles, such as 4 angles, 6 angles or 9 angles are randomly selected in range.Similarly, Gabor kernel functions
SIN function wavelength can also choose according to actual needs or based on experience value, the present embodiment does not repeat this.
S4:The Gabor characteristic is inputted in support vector machines and is judged, the detection to cloud sector domain is completed;Its
In, the support vector machines train to obtain according to the Gabor characteristic and the Gabor characteristic of snow image of cloud atlas picture.
It should be noted that support vector machines are built upon the VC dimensions theory and Structural risk minization of the meter theories of learning
On basis, in the complexity (i.e. to the study precision of specific training sample) of model and learned according to limited sample information
Seek optimal compromise between habit ability (ability for identifying arbitrary sample without error), to obtain best classification capacity.
It needs to be trained support vector machines before use, the present invention obtains 100 Zhang Yun's images and 100 snow in image library
Image, and the Gabor characteristic of the Gabor characteristic and 100 snow images of this 100 Zhang Yun image is extracted, to support vector machines
SVM is trained.Wherein, in the Gabor characteristic of cloud atlas picture and snow image in extracting image library, Gabor filter core letter
Directions and the SIN function wavelength of number setting, when the Gabor characteristic of super-pixel block corresponding with the regions Yun Xue are extracted in step S3,
The direction of Gabor filter kernel function setting is identical as SIN function wavelength.That is, in the present embodiment, support vector machines
The feature that SVM training uses is exactly the 16 dimension Gabor that 8 above-mentioned Gabor kernel functions directions are determined with 2 SIN function wavelength
Feature.The kernel function used when training is gaussian kernel function, remaining parameters selection model default value.
It should be noted that since cloud is to float in the air, formed by the steam constantly polymerization vaporized, then around cloud
Gradient transition is smooth, and it is to be close to that transition around earth's surface is apparent, and Edge texture is larger to avenge.The grain direction of the cloud of vaporization does not have simultaneously
Regular, and due to the influence of ground texture, snow zone-texture has an apparent directionality, for example, Fig. 5 (a) be the present embodiment provides
Visible light wave range cloud sector domain remote sensing images schematic diagram, Fig. 5 (b) be visible light wave range provided in this embodiment avenge regional remote sensing figure
As schematic diagram;Fig. 6 (a) is the corresponding gradient statistic histograms of Fig. 5 (a);Fig. 6 (b) is that the corresponding gradients of Fig. 5 (b) count histogram
Figure.The Gabor characteristic of the corresponding super-pixel block in the regions Yun Xue is further extracted in the regions Yun Xue of acquisition based on step S2.Ginseng
See that Fig. 7 (a), Fig. 7 (b), visible light wave range remote sensing images schematic diagram and visible light wave range respectively provided in this embodiment are distant
Feel the corresponding Gabor characteristic schematic diagram of image schematic diagram.Wherein, Fig. 7 (a) white portions are the regions Yun Xue, and grey parts are mountain
Area or shade.Fig. 7 (b) is Gabor value figures, and Gabor values are smaller, and color is deeper, then aterrimus region is cloud sector domain in Fig. 7 (b),
White border is snow region.Comparison chart 7 (a) and Fig. 7 (b), the present embodiment based on spatial domain natural scene statistics and Gabor characteristic
Remote sensing images cloud detection method of optic, cloud sector domain can be accurately extracted from visible light wave range remote sensing images.
Certainly, the invention may also have other embodiments, without deviating from the spirit and substance of the present invention, ripe
Various corresponding change and deformations can be made according to the present invention certainly by knowing those skilled in the art, but these it is corresponding change and
Deformation should all belong to the protection domain of appended claims of the invention.
Claims (7)
1. a kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics, which is characterized in that include the following steps:
S1:It will be seen that optical band remote sensing images carry out super-pixel segmentation, super-pixel block is obtained;
S2:The waveform parameter f of the corresponding asymmetric generalized Gaussian distribution of each super-pixel block is obtained respectively1With cloud atlas as corresponding non-
The waveform parameter f of symmetrical generalized Gaussian distribution2, the corresponding asymmetric generalized Gaussian distribution of cloudless image waveform parameter f3Between
Distance, if waveform parameter f1With waveform parameter f2The distance between, it is less than waveform parameter f1With waveform parameter f3Between away from
From then waveform parameter f1Corresponding super-pixel block is the regions Yun Xue, otherwise avenges region for non-cloud;
S3:Obtain the Gabor characteristic of the corresponding super-pixel block in the regions Yun Xue;
S4:The Gabor characteristic is inputted in support vector machines and is judged, the detection to cloud sector domain is completed;Wherein, institute
Support vector machines are stated to train to obtain according to the Gabor characteristic and the Gabor characteristic of snow image of cloud atlas picture.
2. a kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics as described in claim 1, feature exist
In described it will be seen that optical band remote sensing images progress super-pixel segmentation is specially:
Super-pixel segmentation is carried out to visible light wave range remote sensing images using simple linear iteration clustering procedure, wherein in simple linear
During the enhancing connectivity of iteration clustering procedure, if the pixel mean value of super-pixel block of the area less than given threshold T2 is more than
Given threshold T3, then area less than given threshold T2 super-pixel block be not incorporated in the area nearest with it be not less than given threshold T2
Super-pixel block;If the pixel mean value of super-pixel block of the area less than given threshold T2 is not more than given threshold T3, area
Super-pixel block less than given threshold T2 is incorporated to the super-pixel block that the area nearest with it is not less than given threshold T2.
3. a kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics as described in claim 1, feature exist
In, will be seen that optical band remote sensing images carry out super-pixel segmentation, obtain super-pixel block after, reject area be less than given threshold T4
Super-pixel block, then by remaining super-pixel block execute step S2 and step S3.
4. a kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics as described in claim 1, feature exist
In the waveform parameter f for obtaining the corresponding asymmetric generalized Gaussian distribution of each super-pixel block1Specially:
Obtain pixel value, local mean value and the Local standard deviation of each pixel of super-pixel block;
According to the pixel value of each pixel, local mean value and Local standard deviation, the inequality contrast normalizing of each pixel is obtained
Change coefficient;
According to the inequality contrast normalization coefficient of each pixel, each super-pixel block corresponding inequality contrast normalization system is obtained
Matrix number;
According to the inequality contrast normalization coefficient matrix, the corresponding asymmetric generalized Gaussian distribution of each super-pixel block is obtained,
And obtain the waveform parameter f of asymmetric generalized Gaussian distribution1。
5. a kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics as claimed in claim 4, feature exist
In the waveform parameter f1With waveform parameter f2The distance between acquisition methods be:
Obtain the waveform parameter of the corresponding asymmetric generalized Gaussian distribution of super-pixel block, wherein the waveform parameter includes waveform
Halfwidth, left offset, right offset and mean value;
By pixel centered on each pixel of super-pixel block, then by the inequality contrast normalization coefficient of each central pixel point point
Inequality contrast not with the horizontal direction of central pixel point, vertical direction, diagonally opposed and dihedral direction pixel is returned
One changes multiplication, to obtain super-pixel block in the horizontal direction, vertical direction, diagonally opposed and dihedral direction it is corresponding
Poor contrast normalization coefficient matrix;
According to super-pixel block in the horizontal direction, vertical direction, the corresponding inequality contrast normalizing in diagonally opposed and dihedral direction
Change coefficient matrix, obtain respectively super-pixel block in the horizontal direction, vertical direction, diagonally opposed and dihedral direction it is corresponding non-right
Claim generalized Gaussian distribution;
Respectively obtain super-pixel block in the horizontal direction, vertical direction, the corresponding asymmetric broad sense in diagonally opposed and dihedral direction
The waveform parameter of Gaussian Profile;
Obtain the corresponding covariance matrix of 20 characteristic values of super-pixel block;Wherein, 20 characteristic values are respectively super-pixel block
The waveform parameter and super-pixel block of corresponding asymmetric generalized Gaussian distribution in the horizontal direction, vertical direction, it is diagonally opposed,
The waveform parameter of the corresponding asymmetric generalized Gaussian distribution in dihedral direction;
According to the 20 of super-pixel block characteristic values, the corresponding covariance matrix of 20 characteristic values of super-pixel block, cloud atlas picture 20
The corresponding covariance matrix of 20 characteristic values of a characteristic value, cloud atlas picture obtains the corresponding asymmetric Generalized Gaussian of super-pixel block
The waveform parameter f of distribution1With cloud atlas as the waveform parameter f of corresponding asymmetric generalized Gaussian distribution2The distance between.
6. a kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics as claimed in claim 5, feature exist
In the waveform parameter f1With waveform parameter f3The distance between acquisition methods be:
According to the 20 of super-pixel block characteristic values, the corresponding covariance matrix of 20 characteristic values, non-cloud image 20 characteristic values,
The corresponding covariance matrix of 20 characteristic values obtains the waveform parameter f of the corresponding asymmetric generalized Gaussian distribution of super-pixel block1With
The waveform parameter f of the corresponding asymmetric generalized Gaussian distribution of non-cloud image3The distance between.
7. a kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics as described in claim 1, feature exist
In the Gabor characteristic for obtaining the corresponding super-pixel block in the regions Yun Xue is specially:
Direction of more than two angles as the kernel function of Gabor filter is randomly selected in 0 °~360 ° of range;
It is that the kernel function determines more than two SIN function wavelength based on experience value, wherein the SIN function wavelength determines
The frequency of the kernel function;
According to the direction and frequency, the corresponding super-pixel block in the regions Yun Xue is obtained on different directions and frequency
Gabor values;
The average value for obtaining multiple Gabor values, using the average value as the Gabor of the corresponding super-pixel block in the regions Yun Xue
Feature.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810352969.3A CN108629297A (en) | 2018-04-19 | 2018-04-19 | A kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810352969.3A CN108629297A (en) | 2018-04-19 | 2018-04-19 | A kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108629297A true CN108629297A (en) | 2018-10-09 |
Family
ID=63705497
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810352969.3A Pending CN108629297A (en) | 2018-04-19 | 2018-04-19 | A kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108629297A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111307291A (en) * | 2020-03-02 | 2020-06-19 | 武汉大学 | Surface temperature anomaly detection and positioning method, device and system based on unmanned aerial vehicle |
CN112053371A (en) * | 2020-09-29 | 2020-12-08 | 成都星时代宇航科技有限公司 | Water body extraction method and device in remote sensing image |
CN112183505A (en) * | 2020-11-30 | 2021-01-05 | 江苏久智环境科技服务有限公司 | Unmanned aerial vehicle aerial photography target identification method for smart gardens |
CN113449788A (en) * | 2021-06-24 | 2021-09-28 | 湖南大学 | Visible light and near-infrared remote sensing image cloud detection method |
CN117152637A (en) * | 2023-10-30 | 2023-12-01 | 中国海洋大学 | Strong convection cloud identification method based on FY-4A satellite cloud image prediction sequence |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9058517B1 (en) * | 2014-04-16 | 2015-06-16 | I.R.I.S. | Pattern recognition system and method using Gabor functions |
CN106503739A (en) * | 2016-10-31 | 2017-03-15 | 中国地质大学(武汉) | The target in hyperspectral remotely sensed image svm classifier method and system of combined spectral and textural characteristics |
CN106920231A (en) * | 2017-02-17 | 2017-07-04 | 北京理工大学 | A kind of remote sensing image clouds appraisal procedure based on full-colour image statistical nature |
CN107610114A (en) * | 2017-09-15 | 2018-01-19 | 武汉大学 | Optical satellite remote sensing image cloud snow mist detection method based on SVMs |
-
2018
- 2018-04-19 CN CN201810352969.3A patent/CN108629297A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9058517B1 (en) * | 2014-04-16 | 2015-06-16 | I.R.I.S. | Pattern recognition system and method using Gabor functions |
CN106503739A (en) * | 2016-10-31 | 2017-03-15 | 中国地质大学(武汉) | The target in hyperspectral remotely sensed image svm classifier method and system of combined spectral and textural characteristics |
CN106920231A (en) * | 2017-02-17 | 2017-07-04 | 北京理工大学 | A kind of remote sensing image clouds appraisal procedure based on full-colour image statistical nature |
CN107610114A (en) * | 2017-09-15 | 2018-01-19 | 武汉大学 | Optical satellite remote sensing image cloud snow mist detection method based on SVMs |
Non-Patent Citations (2)
Title |
---|
LANTX_SYSU: "SLIC超像素分割算法", 《CSDN》 * |
杨羚等: "基于图割模型的高分全色影像自动云检测", 《地理空间信息》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111307291A (en) * | 2020-03-02 | 2020-06-19 | 武汉大学 | Surface temperature anomaly detection and positioning method, device and system based on unmanned aerial vehicle |
CN112053371A (en) * | 2020-09-29 | 2020-12-08 | 成都星时代宇航科技有限公司 | Water body extraction method and device in remote sensing image |
CN112183505A (en) * | 2020-11-30 | 2021-01-05 | 江苏久智环境科技服务有限公司 | Unmanned aerial vehicle aerial photography target identification method for smart gardens |
CN113449788A (en) * | 2021-06-24 | 2021-09-28 | 湖南大学 | Visible light and near-infrared remote sensing image cloud detection method |
CN117152637A (en) * | 2023-10-30 | 2023-12-01 | 中国海洋大学 | Strong convection cloud identification method based on FY-4A satellite cloud image prediction sequence |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108629297A (en) | A kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics | |
CN107610114B (en) | optical satellite remote sensing image cloud and snow fog detection method based on support vector machine | |
CN110287932B (en) | Road blocking information extraction method based on deep learning image semantic segmentation | |
CN102722891B (en) | Method for detecting image significance | |
Cerutti et al. | A parametric active polygon for leaf segmentation and shape estimation | |
CN102324032B (en) | Texture feature extraction method for gray level co-occurrence matrix in polar coordinate system | |
CN105335966B (en) | Multiscale morphology image division method based on local homogeney index | |
CN103727930B (en) | A kind of laser range finder based on edge matching and camera relative pose scaling method | |
CN107092871B (en) | Remote sensing image building detection method based on multiple dimensioned multiple features fusion | |
CN107392968B (en) | The image significance detection method of Fusion of Color comparison diagram and Color-spatial distribution figure | |
CN104392240A (en) | Parasite egg identification method based on multi-feature fusion | |
CN107862698A (en) | Light field foreground segmentation method and device based on K mean cluster | |
CN109118528A (en) | Singular value decomposition image matching algorithm based on area dividing | |
CN110992341A (en) | Segmentation-based airborne LiDAR point cloud building extraction method | |
CN110309781B (en) | House damage remote sensing identification method based on multi-scale spectrum texture self-adaptive fusion | |
CN103984953A (en) | Cityscape image semantic segmentation method based on multi-feature fusion and Boosting decision forest | |
JP2006107462A (en) | Method for detecting road in aerial image of ground topology | |
Hu et al. | Local edge distributions for detection of salient structure textures and objects | |
CN108256467B (en) | Traffic sign detection method based on visual attention mechanism and geometric features | |
Yuan et al. | Learning to count buildings in diverse aerial scenes | |
Wang et al. | A novel multi-scale segmentation algorithm for high resolution remote sensing images based on wavelet transform and improved JSEG algorithm | |
CN111709430B (en) | Ground extraction method of outdoor scene three-dimensional point cloud based on Gaussian process regression | |
CN108052886A (en) | A kind of puccinia striiformis uredospore programming count method of counting | |
Herrera et al. | A segmentation method using Otsu and fuzzy k-Means for stereovision matching in hemispherical images from forest environments | |
Galsgaard et al. | Circular hough transform and local circularity measure for weight estimation of a graph-cut based wood stack measurement |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20181009 |