CN100370486C - Typhoon center positioning method based on embedded type concealed Markov model and cross entropy - Google Patents
Typhoon center positioning method based on embedded type concealed Markov model and cross entropy Download PDFInfo
- Publication number
- CN100370486C CN100370486C CNB2006100134478A CN200610013447A CN100370486C CN 100370486 C CN100370486 C CN 100370486C CN B2006100134478 A CNB2006100134478 A CN B2006100134478A CN 200610013447 A CN200610013447 A CN 200610013447A CN 100370486 C CN100370486 C CN 100370486C
- Authority
- CN
- China
- Prior art keywords
- typhoon
- image
- status block
- wavelet
- state
- 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.)
- Expired - Fee Related
Links
Images
Landscapes
- Image Processing (AREA)
Abstract
The text of the present invention uses a differential cloud picture of a time sequence satellite cloud picture as an investigated subject, and provides a typhoon center positioning method that (1). a differential cloud picture sequence is generated by the time sequence of the cloud picture, wavelet transformation is carried out to the differential cloud picture, and a wavelet domain serves as a data source and an information extraction source of a bottom layer; (2). the division to an adaptive state block of an image is realized by an embedded type concealed Markov model according to the space correlation of the typhoon differential cloud picture; the divided state blocks are changed as the change of the typhoon, and have certain adaptability; (3). the motion tracking and measurement of energy of motion are carried out to the state blocks by a cross entropy according to the time correlation of the sequence differential cloud picture; (4) an energy function is configured, and the typhoon center is primarily ascertained in a low resolution layer of the wavelet transform through looking for minimum energy. The guidance of initial positioning to accurate positioning is realized by the correlation of the low resolution layer and a high resolution layer. The typhoon center positioning method provided by the present invention has the advantages of high universality, high positioning accuracy and rapid calculation speed.
Description
Technical field
The present invention relates to a kind of center of typhoon localization method, belong to the weather forecast field.
Background technology
According to whether existing typhoon eye, localization method to be divided in the typhoon cloud atlas typhoon and anophthalmia center of typhoon location are arranged.
Having a typhoon to have tangible typhoon eye, be reflected on the satellite cloud picture, is in the cloud sector of a slice high brightness, has a darker zone.Because eye district feature is obvious, locatees more or less freely.But typhoon eye results from the maturity stage of typhoon usually, at this moment typhoon is forecast to have lost the best opportunity.
The anophthalmia typhoon promptly be can't see the typhoon that eye is distinguished in cloud atlas.Divide two kinds of typical case, the one, when typhoon is in generation phase, decline phase, typhoon anophthalmia district; The 2nd, high-rise thin cloud blocks the eye district.Anophthalmia center of typhoon location has proposed following method:
(1) utilize shape information to position
1. be similar to bowlder when the central airtight cloud sector of typhoon, the geometric center in cloud sector is decided to be center of typhoon.
2. extract local helix,, obtain center of typhoon local helix match based on cloud bar.
(2) utilize half-tone information to position
When the airtight cloud sector of typhoon central authorities distributes when asymmetric, the center of typhoon orientation problem can be sought by asking for the big zone of cloud cluster shade of gray.
(3) utilize rotatory information location
Typhoon is envisioned for rigid body, analyzes the rotational structure of typhoon, determine center of typhoon by estimation.
(4) utilize expert system
System can carry out machine learning to source book, obtains the knowledge base about typhoon forecast.Under the support of background knowledge, adopt the inference mechanism center that detects a typhoon.
Owing to the formation reason of typhoon and geographic position etc. vary, make the form of each typhoon cloud system obviously different.The common drawback of said method is that adaptability is poor, and versatility is undesirable, and the localization method that is adopted can not satisfy all types typhoon in all periods.In addition, big multi-method is based on single width cloud atlas rather than sequence cloud atlas, and the information utilization is little, does not consider that typhoon self has one from producing to the evolution of withering away.
Summary of the invention
The present invention according to center of typhoon for the other parts of typhoon, the characteristics that kinergety is minimum, difference cloud atlas with the time series satellite cloud picture is a research object, utilize the multidate information in the typhoon motion process, the spatial coherence of combining image, the temporal correlation of sequence image propose the more intense center of typhoon localization method of a cover versatility.It comprises the following steps:
(1) generate difference cloud atlas sequence by the time series of cloud atlas, and the difference cloud atlas carried out wavelet transformation, with wavelet field as bottom data source and information extraction source.
(2) according to the spatial coherence of typhoon difference cloud atlas, the self-adaptation status block that uses built-in type hidden Markov model that the wavelet field of difference image is carried out is cut apart, and the status block that is divided into changes along with the variation of typhoon.
(3), use cross entropy status block to be carried out the tolerance of dynamic tracking and kinergety according to the temporal correlation of differential of sequence cloud atlas.
(4) cross entropy that calculates according to above-mentioned steps (3), structure is tentatively determined center of typhoon in the low-resolution layer of wavelet transformation by seeking energy-minimum in order to describe the energy function of kinergety.
(5) utilize the correlativity of low-resolution layer and resolution layer,, instruct, realize the locating information fusion of low-resolution layer and resolution layer in the accurate location of resolution layer according to the center of typhoon of tentatively determining by step (4).
In the present invention, can adopt the Haar wavelet basis that difference image is done wavelet transformation.
Step in the technique scheme (2) can be carried out according to the following step, and wherein, step (21), (22), (23) are unified to cut apart to image, and step (24), (25), (26) are cut apart with the self-adaptation that realizes image revising unified cutting apart:
(21) set the number of super state and the number of each super state neutron state, determine the structure of image built-in type hidden Markov model;
(22) setting according to step (21) is divided into several status blocks with the difference image unification, and with each status block as observing vector, cutting apart of this moment is a kind ofly on average to cut apart, and do not consider the spatial coherence of image;
(23) according to super state number, in the vertical direction is determined super state parameter, will surpass state parameter again and press from left to right horizontal direction, distributes unitedly in the corresponding sub-state, determines sub-state parameter, the status block parameter initialization after realization is cut apart;
(24) use dual embedded Viterbi algorithm, determine the best embedded type state sequence and the super status switch of image;
(25) use the K averaging method pixel position is adjacent and that state is identical to be divided in the same status block, and the parameter of this piece is revised;
(26) repeat above-mentioned steps (24) and (25) until meeting the condition of convergence: in the subsequent iteration process, the probability difference that twice calculates before and after the use Viterbi algorithm is less than threshold value.
Step in the technique scheme (3) can be carried out according to the following step:
(31) the present invention is with the difference of two status block probability distribution of cross entropy tolerance, and it is defined as:
P=(p wherein
1, p
2..., p
N), Q=(q
1, q
2..., q
N), be respectively the probability distribution of two status blocks;
(32) test set of definite tracked status block in the one adjacent seasonal effect in time series cloud atlas of back;
(33) by the distance of each status block in tracked of cross entropy assessment and the test set, finish dynamic tracking.
Among the present invention, the energy function that is defined in the status block in the whole motion process can be the quadratic sum of the cross entropy of this status block in each time interval.
The present invention has following beneficial effect:
(1) by using difference image to embody multidate information in the typhoon motion process, finished conversion from the static information source to the multidate information source.
(2) locating with respect to single image, is research object with the multiple image of difference image sequence, more meets the evolution of typhoon, has enlarged the information utilization, has improved bearing accuracy.
(3), can utilize the multiresolution analysis of wavelet transformation and wavelet coefficient to improve travelling speed in the continuation of interlayer based on the analytical approach of wavelet field with respect to analysis based on pixel grey scale.
(4) aspect correlativity, analyzed the spatial coherence of image, the temporal correlation of sequence image, both combine and can improve bearing accuracy.Built-in type hidden Markov model combines with cross entropy, is suitable for describing the typhoon sequence image.
(5) propose a kind of built-in type hidden Er Kefu model adaptation training patterns that is directed to typhoon, realized that finally status block changes along with the variation of typhoon, image has been carried out adaptive status block cut apart.Status block with adaptivity is suitable for the typhoon of different shape, can improve the versatility of the inventive method, and makes the subsequent dynamic tracking have good followability.
(6) to the model training of adjacent time series cloud atlas, adopt the strategy that develops on the basis of inheriting, can reduce the model training time of the inventive method.
(7), be more suitable for having reduced matching error in typhoon based on the coupling of status block with respect to coupling based on pixel.
(8) based on minimum this universal law of the persistent movement energy of center of typhoon,, locate center of typhoon, can improve the versatility of the inventive method by the implicit kinergety information of measuring period sequence typhoon difference cloud atlas.
(9) utilize the multiresolution characteristic of wavelet transformation, realize that the locating information of low-resolution layer and resolution layer merges, under the prerequisite that guarantees bearing accuracy, avoided global search, can reduce the working time of the inventive method.
Description of drawings
The satellite cloud picture in 04 on August 1,13: 32 of Figure 120
The satellite cloud picture in 004 on August 1,14: 32 of Figure 22
Fig. 3 difference image (difference image of Fig. 1 and Fig. 2)
The repeatedly wavelet transformation synoptic diagram of Fig. 4 difference image
Fig. 5 difference image wavelet transformation is figure as a result, and wherein, figure (a) is one, two, three wavelet transformation result of glh, and figure (b) is g
HlOne, two, three wavelet transformation result, figure (c) be one, two, three wavelet transformation of ghh
The two condition zero-mean mixed Gauss model of Fig. 6 wavelet coefficient
The spatial coherence of Fig. 7 difference image is cut apart process flow diagram
The embedded Viterbi algorithm structure of Fig. 8 figure
The segmentation result synoptic diagram of Fig. 9 typhoon difference image
The sub-piece dynamic tracking process synoptic diagram of Figure 10 difference sequence figure
Embodiment
No matter how typhoon changes in motion process, the wind speed minimum of center of typhoon, and it is the most weak to move, and this point is all unexceptional to the typhoon of any kind.In addition, typhoon has one from producing to the change procedure of withering away, and its movable information is present among the time series cloud atlas of typhoon.Invention is a research object with the difference cloud atlas of time series satellite cloud picture, proposes the more intense center of typhoon localization method of a cover versatility, and this method mainly comprises the following steps:
(1) generate difference cloud atlas sequence by the time series of cloud atlas, and the difference cloud atlas carried out wavelet transformation, with wavelet field as bottom data source and information extraction source;
(2), use the self-adaptation status block of built-in type hidden Markov model realization image to cut apart according to the spatial coherence of typhoon difference cloud atlas.The status block of cutting apart changes along with the variation of typhoon, has certain adaptivity;
(3), use cross entropy status block to be carried out the tolerance of dynamic tracking and kinergety according to the temporal correlation of differential of sequence cloud atlas:
(4) structure energy function is tentatively determined center of typhoon in the low-resolution layer of wavelet transformation by seeking energy-minimum.Utilize the correlativity of low-resolution layer and resolution layer, realize that Primary Location is to pinpoint guidance.
Following content has been carried out labor to above-mentioned various piece.
1. ask for difference image, difference image is carried out wavelet transformation
Foundation of the present invention is: center of typhoon is for the other parts of typhoon, and kinergety is minimum.Therefore, all the elements are all based on motion analysis.And difference image can embody movable information, by asking for adjacent seasonal effect in time series difference image, can realize by the conversion of static information source to the multidate information source.Detailed process is as follows:
(1) asks for the difference image of adjacent time series satellite cloud picture
Earlier single width typhoon image is carried out stretching, removing as contrast the pre-service of noise etc.Get wherein adjacent seasonal effect in time series two two field pictures then and make difference, obtain difference image.Calculus of differences is: the gray-scale value of establishing two width of cloth images is respectively f
1(x, y) and f
2(x, y).For drawing gradation of image after avoiding subtracting each other for negative, with f
12(x, y)=| f
1(x, y)-f
2(x, y) | be defined as difference result.Difference image can reflect the movable information that comprises in the image preferably: motion Shaoxing opera is strong, and the image difference is big more, and the gray-scale value of the difference image that obtains like this is just big more, otherwise more little.When not having motion in the image, then the gray-scale value of difference image is zero.Fig. 1,2,3 is for making the synoptic diagram of calculus of differences: Fig. 1 is the satellite cloud picture in 13: 32 on the 1st August in 2004, and Fig. 2 is the satellite cloud picture in 2004 on August 1,14: 32.Both do calculus of differences, and difference image as shown in Figure 3.
(2) difference image is carried out wavelet transformation
The present invention is directly vectorial as observing with difference image, but at first difference image is carried out wavelet transformation, and its wavelet field is vectorial as observing.This mainly is owing to following reason:
1. utilize wavelet transformation can reduce the dimension of observing vector, to adapt to the requirement of mass data processing.
Along with science and technology development, satellite cloud picture is tending towards the data scale of magnanimity gradually, therefore is necessary to consider to observe the problem of vectorial dimension compression, and wavelet transformation can address this problem.The wavelet transformation process is: during each conversion, the data on the preceding low pass frequency band are descended one deck conversion as input, the image zone that just to be decomposed into 4 sizes be the different frequency bands of original size 1/4 after the conversion.After the conversion, image is for original image each time, and resolution has reduced, but simultaneously data scale to be processed also reduces to original 1/4.Can carry out Primary Location to typhoon comparatively reasonably on the level at one, the present invention is that the image behind wavelet transformation for the third time carries out Primary Location to typhoon.This image size is original
, the dimension of observing vector is reduced to original
Therefore can use wavelet transformation to reduce and observe vectorial dimension, reduce working time.
2. observation space is converted to wavelet field by the gray scale territory, the convenient built-in type hidden Markov model analysis wavelet territory of using.
Built-in type hidden Markov model is the focus of current international research, and this model has some good character.Wavelet transformation has had the application of built-in type hidden Markov model may.
3. utilize the continuation of wavelet coefficient between layer to realize the fusion of locating information.
In a plurality of layers of wavelet transformation, low-resolution layer statistics global information, resolution layer embodies detailed information, and the wavelet coefficient between adjacent layer has continuation.Referring to Fig. 4, behind the two-dimensional wavelet transformation, the wavelet coefficient of each frequency band has constituted a natural tree structure.Wherein lower floor is a resolution layer, and the upper strata is a low-resolution layer.Wavelet coefficient shows as continuation in the correlativity between the layer: in this quaternary tree, has father and son's dependence between adjacent layer, the wavelet coefficient of four child nodes of related next detailed level of the wavelet coefficient of each father node, the value by father node can the predictor node value.Utilize this continuation, realize the guidance that the low Primary Location of differentiating on the layer is located resolution layer, locating information can be merged.Avoid full search, reduced working time.
It is a lot of that image is carried out method of wavelet, and the present invention uses the Haar wavelet basis that image is carried out wavelet transformation.The Haar wavelet basis is made of following four wavelet basiss:
Local level and smooth wavelet basis:
Horizontal edge detects wavelet basis:
Vertical rim detection wavelet basis:
Tilted direction rim detection wavelet basis:
Wavelet transformation is a kind of decomposition to input signal, and four different wavelet basiss carry out conversion to image in the use Haar wavelet basis, obtain the information of different frequency bands.h
LLBe a kind of low-pass filter, use it that image is carried out conversion after, obtain the low frequency part information of image, HFS is filtered, consequently the average information of image.And g
LH, g
HL, g
HHBe Hi-pass filter, image is carried out conversion after, obtain the HFS information of image, low frequency part is filtered, its result be respectively image level, vertically, the change information on the tilted direction, the result after the conversion is called g
LH, g
HL, g
HHWavelet field, the present invention is with g
LH, g
HL, g
HHWavelet field is a research object.The result of three wavelet transformations as shown in Figure 5.
2. use the spatial coherence of built-in type hidden Markov model analysis image and carry out the self-adaptation status block and cut apart
The cloud body belongs to non-rigid body, does not have tangible boundary information, shows the character of regional gradual change, so two adjacent its image similarities of cloud of position, and the present invention is called spatial coherence with this character, and according to spatial coherence image segmentation is become several status blocks.These status blocks have adaptivity, along with its size of the variation of cloud and layout all in variation.
As mentioned above, be converted into wavelet field through observation space behind the wavelet transformation by the gray scale territory, the spatial coherence of cloud atlas is represented by the aggregation of wavelet coefficient in same yardstick.This aggregation is the correlativity of wavelet coefficient within same yardstick, if promptly a certain wavelet coefficient is bigger or little, in the same yardstick, the wavelet coefficient around it is also bigger probably or little.
The present invention uses the aggregation of built-in type hidden Markov model analysis wavelet coefficient, and image has been carried out adaptive status block cuts apart.Use the reason of this model as follows:
Can use the probability density function of theory of random processes analysis wavelet coefficient, the probability density function of wavelet coefficient be regarded as once realized at random.And the present invention analyzes the instrument of this stochastic process is built-in type hidden Markov model.The more detailed introduction of relevant Wavelet-Domain Hidden Markov Model has had report in the literature, can consult following document:
(1) S é bastien Aupetit, Nicolas Monmarch é, the paper of Mohamed Slimane: Hidden Markov ModelsTraining by a Particle Swarm Optimization Algorithm (source: Mathematical Modelling andAlgorithms, Online First, February 2006)
(2) G.B.Di Masi, the paper of L.Stettner: Ergodicity of hidden Markov models (source: Mathematics ofControl, Signals, and Systems (MCSS), Volume 17, and Number 4, October 2005 Pages:269-296)
(3) paper of Ath.Kehagias: A hidden Markov model segmentation procedure for hydrological andenvironmental time series (source: Stochastic Environmental Research and Risk Assessment (SERRA), Volume 18, Number 2, April 2004, Pages:117-130)
(4) paper of T.D.Pham: Perception-based hidden Markov models:a theoretical framework fordata mining and knowledge discovery (source: Soft Computing-A Fusion of Foundations, Methodologies and Applications, Volume 6, Number 5, August 2002, Pages:400-405)
(5) Paolo Frasconi, Giovanni Soda, the paper of Alessandro Vullo: Hidden Markov Models for TextCategorization in Multi-Page Documents (source: Intelligent Information Systems, Volume 18, Numbers 2-3, March 2002, Pages:195-217)
Hidden Markov model is that state variable with wavelet coefficient is a research object.The generation of state variable mainly is because wavelet transformation has shrinkability, and promptly type signal is behind wavelet transformation, and the minority wavelet coefficient is big value, and majority is little value.The wavelet coefficient that value is bigger has been concentrated the main information of signal, and common fast-changing position in the correspondence image is as the edge etc.And the quantity of information that the less coefficient of value contains is fewer, often corresponding details composition.In addition, wavelet coefficient can be regarded as the realization at random of signal distributions.Based on above-mentioned 2 reasons, the state variable of wavelet coefficient is expressed as " height " or " low ".Wherein, state variable is " height ", contains main signal energy; State variable is " low ", contains a spot of signal energy.If every kind of state variable and a probability density are interrelated, then: " high " state is represented by the probability density of zero-mean, big variance; " low " state is represented by the probability density of zero-mean, little variance.Each wavelet coefficient is described by the mixture model of above-mentioned two states, as shown in Figure 6.
What Fig. 6 represented is the two condition zero-mean mixed Gauss model of wavelet coefficient as stochastic variable W.Represent state variable S with white point
i, i=1,2.Corresponding probability is p
Si(i),
, pairing variance is σ
Ij 2, j=1,2.Represent wavelet coefficient W with stain.The non-Gaussian probability density of wavelet coefficient W is represented by the mixture model of the Gaussian probability density of W|S.Although W|S has Gaussian probability density, because the randomness of variable S, so W|S can describe the non-Gaussian probability density of W.This has exactly embodied the concentration of energy character of wavelet coefficient.Therefore, non-Gauss model has been explained the statistical property of wavelet coefficient more exactly.
Like this, the probability density function of W can be represented with following formula:
The value of wavelet coefficient is observable, and its state value is hidden, and is immesurable.Hidden Markov model is based on the state of wavelet coefficient, and its basic thought is to remove the perception state value by observed value.
Image belongs to 2-D data, and built-in type hidden Markov model is suitable for describing 2-D data, and calculated amount is smaller.Built-in type hidden Markov model is a kind of specific type in the hidden Markov model, is made of one group of super status switch and the some groups of sub-status switches that embed.The corresponding one group of sub-status switch that embeds of each super state.Embed the formula state from horizontal direction according to from left to right order data of description.Super state is summarized to its pairing one group of sub-status switch, from the vertical direction data of description.
The more detailed introduction of relevant built-in type hidden Markov model, report has been arranged in the literature, can consult following document: Radford M.Neal, Matthew J.Beal, the paper of and Sam T.Roweis: Inferring State Sequences forNon-linear Systems with Embedded Hidden Markov Models (source: ADVANCES IN NEURALINFORMATION PROCESSING SYSTEMS 1616:401-408,2004)
The inventive method uses built-in type hidden Markov model that image is cut apart according to the spatial coherence of image, and particular content is as follows:
(1) use built-in type hidden Markov model that the self-adaptation that typhoon difference cloud atlas carries out status block is cut apart
With built-in type hidden Markov (E-HMM) model the wavelet field of difference image is trained, and according to spatial coherence image segmentation is become several status blocks, its process flow diagram as shown in Figure 7.Its basic thought is: at first image is unified to cut apart, promptly do not consider spatial coherence, image is averaged cut apart.Then,, initial segmentation result is adjusted, cut apart with the self-adaptation that realizes status block according to the spatial coherence of image.Concrete training process is as follows:
1. determine the structure of image built-in type hidden Markov model.Set the number of super state and the number of each super state neutron state.
2. according to 1. setting the difference image unification is divided into several status blocks, and with each status block as observing vector.Be not consider preliminary on average the cutting apart of spatial coherence at this moment, M * N big status block, M=N=5 among the present invention such as just image be divided into.
3. according to the result after cutting apart, to parameter initialization:
A. according to super state number, in the vertical direction is determined the parameter of super state;
B. will surpass state parameter by from left to right horizontal direction, distribute unitedly in the corresponding sub-state, determine sub-state parameter.
4. adjust initial segmentation according to spatial coherence
I. use dual embedded Viterbi algorithm to determine the best embedded type state sequence and the super status switch of image,, infer the internal state sequence that to observe promptly according to the observable values sequence.Its training process as shown in Figure 8.
A. with the Viterbi algorithm application in image, calculate the distributions, state-transition matrix, observation matrix of each sub-state in the super state etc.;
B. each is embedded the formula state and summarize, calculate the probability of each super state, determine best super status switch.
P(O
t0，1...O
t0，T1，q
1，1 (t 0 )，...，q
1，T1 (t 0 )|λ
(k))，1≤k≤N
0(2)
λ wherein
(k)Represent the parameter value that k is capable, comprise initial super distributions, super state-transition matrix, observation matrix etc., q
T0, t1 (k), 1≤t
1≤ T
1Represent t in k the super state
0Row t
1Row are observed vectorial O
T0, t1State.
Ii. image is cut apart again with the K averaging method, and parameter is revised.
Use the K averaging method pixel position is adjacent and that state is identical to be divided in the same status block, and the parameter of this piece is revised.
Iii. repeat the process of i and ii, until convergence.The condition that iteration finishes is: in the subsequent iteration process, the probability difference that twice calculates before and after the use Viterbi algorithm is less than threshold value.
Following content is introduced with regard to the K means clustering algorithm:
The K means clustering algorithm is one of the most frequently used unsupervised learning algorithm, and whole data set is divided into K gathering.Reserve classification in advance and count K,, on any a plurality of sample sets, obtain K cluster by continuous adjustment cluster barycenter.Its central idea is to minimize distance in total class.Detailed process is as follows:
Picked at random K sample is as kind of a child node;
To remaining set of records ends, calculate each record and K distance of planting child node, each record is grouped into nearest that kind child node, whole like this record set is divided into K gathering for the first time;
Each is assembled, calculate the barycenter of assembling (gathering central point);
With each barycenter is kind of a child node, repeats above-mentioned steps, changes until assembling no longer.
As for the more detailed introduction to the Viterbi algorithm, in the prior art appearing in the newspapers led more, can consult following document:
(1) Luis Javier Rodr í guez, the paper of In é s Torres: Comparative Study of the Baum-Welch and ViterbiTraining Algorithms Applied to Read and Spontaneous Speech Recognition (source: Lecture Notes inComputer Science, Volume 2652,2003, Pages:847-857)
(2) A.V Korolev, the paper of A.M.Silaev: The Viterbi algorithm for models of hidden Markovprocesses with the unknown moment of appearance of parameter jump (source: Radiophysics andQuantum Electronics, Volume 48, Number 4, April 2005, Pages:321-328)
The result of finally cutting apart is as shown in Figure 9: difference image is divided into 5 * 5 status blocks according to spatial coherence.Along vertical direction, the typhoon difference image is divided into 5 super states, constitutes and embed the formula state along each super state of horizontal direction by 5.Corresponding one of each status block embeds the formula state, and wavelet coefficient adjacent by several positions and that state value is identical constitutes.Variation on the horizontal direction between the adjacent states is described by the state transition probability that embeds the formula state.State by sub-state transition probability and super state embodies the spatial coherence of image, and is according to image segmentation is become several status blocks with the spatial coherence.These status blocks have adaptivity, along with its size of the variation of typhoon and layout all in dynamic change, so cutting procedure is a dynamic process based on built-in type hidden Markov model.
(2) adopt the strategy that on the basis of inheriting, develops, fpn image is carried out status block cut apart.
The image change of the adjacent sequential of typhoon is not very big, and therefore the model of front and back sequential cloud atlas also has certain correlativity, and this correlativity is called temporal correlation.Temporal correlation has determined the similarity of sequence image model.Because the typhoon image of adjacent sequential has this temporal correlation, can adopt the mechanism of not only inheriting but also developing during model training, to reduce the training time.
Detailed process is as follows:
After the model training of last sequential chart picture is good, the model of back one sequential can be inherited the model of last sequential when initial, on the basis of inheriting, develop, the built-in type hidden Markov model parameter is carried out online recursion to be estimated, to realize the self-adaptation of status block, status block will the dynamic change along with the variation of typhoon like this.
This training mode had both reduced the training time, had guaranteed that again dynamic change to embody adaptivity, makes method of the present invention can be suitable for the typhoon of different shape, had strengthened versatility.In addition, the adaptivity of status block makes follow-up dynamic tracking have good followability, has strengthened the antijamming capability and the stability of the inventive method.
3. use cross entropy that typhoon difference image sequence is carried out the status block dynamic tracking along time shaft
Use cross entropy to measure the difference of two status block probability distribution, this species diversity is called two distances between status block, this distance has reflected the dynamic change of status block indirectly.The definition of cross entropy:
Wherein,
P＝(p
1，p
2，…，p
N)，Q＝(q
1，q
2，…，q
N)。P, Q are respectively the probability distribution of two status blocks.
All status blocks all have corresponding probable value after training through built-in type hidden Markov model.Use the difference degree between two status blocks of cross entropy assessment, cross entropy is more little, and then two status block differences are more little, and are similar more.By cross entropy the status block in former and later two time serieses is followed the tracks of, detailed process is as follows:
At first in the one time sequence of back, determine in the last time series tracked position.Then with 9 status blocks around this position as test set.Use the distance of tracked of cross entropy assessment and interior each status block of test set.With the cross entropy minimum, promptly the status block of distance minimum is as tracking results.Concrete tracing process as shown in figure 10, every bit is all represented a status block among the figure.Use cross entropy to determine that each status block has been finished the dynamic tracking process of status block along the corresponding relation of time shaft in typhoon difference image sequence like this.
Said method is a research object with the differential of sequence cloud atlas, based on the method for single width or two, three width of cloth cloud atlas, has enlarged the information utilization more in the past.In the present differential of sequence cloud atlas of the persistent movement imformosome of typhoon, this is the method for research object with the differential of sequence cloud atlas, more can embody the evolution of typhoon self, improves bearing accuracy.In addition, because typhoon is dynamic change, its shape information and half-tone information all can dynamic changes in the typhoon motion process, so have many problems based on the tracking of pixel.And overcome in the past based on the tracking shortcoming of pixel based on the tracking of status block, improved bearing accuracy.
4. determine center of typhoon based on the kinergety function
The present invention carries out centralized positioning to typhoon, makes two bites at a cherry.At first determine energy function, on the low resolution layer of wavelet transformation, carry out Primary Location.Adopt then and recall strategy, continuation according to low-resolution layer and resolution layer in the tree of wavelet coefficient, result with Primary Location instructs in the accurate location of resolution layer, realizes the fusion of locating information in wavelet transformation low-resolution layer and resolution layer.
(1) center of typhoon Primary Location
Energy function U in the tectonic movement process in order to describing kinergety, determines the status block of kinergety minimum by this function, and this piece is center of typhoon.
Suppose to have k the time interval, status block M is D at the cross entropy in l the time interval
M l, 1≤l≤k.
Be defined in the whole motion process, the persistent movement energy of status block M is:
According to said process, respectively at glh, ghl, the kinergety of three wavelet field computing modes of ghh piece M is according to the separation property of small echo, with the motion gross energy of its addition as status block M, that is:
Calculate the motion gross energy of all status blocks on the 3rd layer of wavelet transformation, the status block of gross energy minimum is the center of typhoon of Primary Location.
Any typhoon no matter, the persistent movement energy at its center is minimum.With this point is according to typhoon being carried out centralized positioning, having improved the versatility of the inventive method.This process also is indifferent to the definite energy distribution of typhoon, and just seeks persistent movement energy minimum point.This method can reduce the location mistake, and there are errors in computation in section sometime even the energy of status block calculates, and influence also is very little to the persistent movement energy calculation.
(2) center of typhoon Fine Mapping
The 3rd layer of cloud atlas wavelet transformation compared with former figure and dwindled 64 times, thereby bearing accuracy is much smaller with respect to the location on the resolution layer, and therefore final positioning result must come from former figure.Because wavelet coefficient has certain correlativity at interlayer, the locating information of low resolution layer can instruct the location in resolution layer.The present invention's employing is recalled strategy and is finished guidance, and detailed process is as follows:
In the 3rd layer of wavelet transformation, determine the center of typhoon of Primary Location and the relative distance l and the deviation angle θ of picture centre.In former figure, at first determine picture centre, on the direction of slip chart inconocenter θ, determine the exact position of center of typhoon with the distance of 8l then.
This method merges by the locating information of low-resolution layer and resolution layer, has avoided full search, under the prerequisite that guarantees bearing accuracy, has saved working time.
Claims (5)
1. the center of typhoon localization method based on built-in type hidden Markov model and cross entropy is characterized in that, comprises the following steps:
(1) generate difference cloud atlas sequence by the seasonal effect in time series cloud atlas, and the difference cloud atlas carried out wavelet transformation, with wavelet field as bottom data source and information extraction source;
(2) image is averaged cut apart;
(3) according to the spatial coherence of typhoon difference cloud atlas, use built-in type hidden Markov model that the wavelet field of difference image is carried out the self-adaptation status block and cut apart, the status block that is divided into is changed along with the variation of typhoon;
(4), use cross entropy status block to be carried out the tolerance of dynamic tracking and kinergety according to the temporal correlation of differential of sequence cloud atlas;
(5) cross entropy that calculates according to above-mentioned steps (4), structure by seeking the status block of gross energy minimum, is tentatively determined center of typhoon in the low-resolution layer of wavelet transformation in order to describe the energy function of kinergety;
(6) utilize correlativity between wavelet transformation low-resolution layer and the resolution layer,, finish Primary Location, be implemented in the accurate location of resolution layer pinpoint guidance according to the many resolution characteristics and the step (5) of wavelet transformation.
2. center of typhoon localization method according to claim 1 is characterized in that, adopts the Haar wavelet basis that difference image is done wavelet transformation.
3. center of typhoon localization method according to claim 1, it is characterized in that, step wherein (3) is carried out according to the following step, wherein, step (31), (32), (33) average image to be cut apart, step (34), (35), (36) are cut apart with the self-adaptation that realizes image revising unified cutting apart:
(31) set the number of super state and the number of each super state neutron state, determine the structure of image built-in type hidden Markov model;
(32) setting according to step (31) is divided into several status blocks with the difference image unification, and with each status block as observing vector, cutting apart of this moment is a kind ofly on average to cut apart, and do not consider the spatial coherence of image;
(33) according to super state number, in the vertical direction is determined super state parameter, will surpass state parameter again and press from left to right horizontal direction, distributes unitedly in the corresponding sub-state, determines sub-state parameter, the status block parameter initialization after realization is cut apart;
(34) use dual embedded Viterbi algorithm, determine the best embedded type state sequence and the super status switch of image;
(35) use the K averaging method pixel position is adjacent and that state is identical to be divided in the same status block, and the parameter of this piece is revised;
(36) repeat above-mentioned steps (34) and (35) until meeting the condition of convergence: in the subsequent iteration process, the probability difference that twice calculates before and after the use Viterbi algorithm is less than threshold value.
4. center of typhoon localization method according to claim 1 is characterized in that, step wherein (4) is carried out according to following step:
(41) utilize cross entropy to measure the difference of two status block probability distribution, it is defined as:
P=(p wherein
1, p
2..., p
N), Q=(q
1, q
2..., q
N), be respectively the probability distribution of two status blocks;
(42) test set of definite tracked status block in the one adjacent seasonal effect in time series cloud atlas of back;
(43) by the distance of each status block in tracked of cross entropy assessment and the test set, finish dynamic tracking.
5. center of typhoon localization method according to claim 1 is characterized in that, wherein is defined in the quadratic sum of the energy function of the status block in the whole motion process for the cross entropy of this status block in each time interval in the step (5).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CNB2006100134478A CN100370486C (en) | 2006-04-10 | 2006-04-10 | Typhoon center positioning method based on embedded type concealed Markov model and cross entropy |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CNB2006100134478A CN100370486C (en) | 2006-04-10 | 2006-04-10 | Typhoon center positioning method based on embedded type concealed Markov model and cross entropy |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN1828668A CN1828668A (en) | 2006-09-06 |
| CN100370486C true CN100370486C (en) | 2008-02-20 |
Family
ID=36947040
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CNB2006100134478A Expired - Fee Related CN100370486C (en) | 2006-04-10 | 2006-04-10 | Typhoon center positioning method based on embedded type concealed Markov model and cross entropy |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN100370486C (en) |
Families Citing this family (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101515366B (en) * | 2009-03-30 | 2010-12-01 | 西安电子科技大学 | Watershed SAR image segmentation method based on complex wavelet extraction mark |
| CN101510309B (en) * | 2009-03-30 | 2010-12-01 | 西安电子科技大学 | Segmentation method for improving water parting SAR image based on compound wavelet veins region merge |
| CN101662788B (en) * | 2009-07-22 | 2012-01-18 | 山东科技大学 | Link equation constructing method of virus-type mobile communication cross entropy method |
| CN102129566A (en) * | 2011-03-09 | 2011-07-20 | 国家卫星气象中心 | Method for identifying rainstorm cloud cluster based on stationary meteorological satellite |
| CN102509027B (en) * | 2011-12-19 | 2014-11-26 | 国网电力科学研究院 | Wind powder combined predication method based on cross entropy theory |
| CN102645679B (en) * | 2012-03-13 | 2014-07-02 | 天津大学 | Mesocyclone identification method based on Doppler radar echo images |
| CN104992434B (en) * | 2015-06-23 | 2018-03-27 | 国家卫星气象中心 | The method for carrying out typhoon location on present satellites image according to historical satellite image |
| CN107578441B (en) * | 2017-08-24 | 2023-05-09 | 浙江师范大学 | Infrared bright temperature deviation angle gradient distribution uniformity method for tropical cyclone center positioning |
| CN111462230B (en) * | 2020-02-18 | 2023-08-15 | 天津大学 | Typhoon center positioning method based on deep reinforcement learning |
| CN111523087B (en) * | 2020-04-10 | 2021-04-16 | 北京航空航天大学 | Typhoon intensity long-term change trend analysis method |
| CN113421294B (en) * | 2021-06-10 | 2021-12-21 | 温州市气象局 | Subtropical zone high-voltage similarity method analysis based on geometric image processing and neural network |
| CN117036987B (en) * | 2023-10-10 | 2023-12-08 | 武汉大学 | Remote sensing image space-time fusion method and system based on wavelet domain cross pairing |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2003203300A (en) * | 2001-10-19 | 2003-07-18 | Kyowa Shoko Kk | Method and system for forecasting operation limit of ships in harbors |
| JP2005106744A (en) * | 2003-10-01 | 2005-04-21 | Mitsubishi Electric Corp | Image analysis apparatus |
| JP2005164490A (en) * | 2003-12-04 | 2005-06-23 | Mitsubishi Electric Corp | Typhoon center detection device |
-
2006
- 2006-04-10 CN CNB2006100134478A patent/CN100370486C/en not_active Expired - Fee Related
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2003203300A (en) * | 2001-10-19 | 2003-07-18 | Kyowa Shoko Kk | Method and system for forecasting operation limit of ships in harbors |
| JP2005106744A (en) * | 2003-10-01 | 2005-04-21 | Mitsubishi Electric Corp | Image analysis apparatus |
| JP2005164490A (en) * | 2003-12-04 | 2005-06-23 | Mitsubishi Electric Corp | Typhoon center detection device |
Non-Patent Citations (1)
| Title |
|---|
| 基于云导风场的形成期台风定位. 王萍，宋振龙，孔秀梅，卢汉清.模式识别与人工智能，第18卷第6期. 2005 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN1828668A (en) | 2006-09-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN100370486C (en) | Typhoon center positioning method based on embedded type concealed Markov model and cross entropy | |
| CN100595782C (en) | Classification method for syncretizing optical spectrum information and multi-point simulation space information | |
| CN110378308B (en) | Improved port SAR image near-shore ship detection method based on fast R-CNN | |
| CN104376330B (en) | Polarimetric SAR Image Ship Target Detection method based on super-pixel scattering mechanism | |
| CN102800074B (en) | Synthetic aperture radar (SAR) image change detection difference chart generation method based on contourlet transform | |
| Ismail et al. | Objective automated classification technique for marine landscape mapping in submarine canyons | |
| CN106683102B (en) | SAR image segmentation method based on ridge ripple filter and convolutional coding structure learning model | |
| CN106846322B (en) | The SAR image segmentation method learnt based on curve wave filter and convolutional coding structure | |
| CN110163213A (en) | Remote sensing image segmentation method based on disparity map and multiple dimensioned depth network model | |
| CN106611423B (en) | SAR image segmentation method based on ridge ripple filter and deconvolution structural model | |
| CN110555841B (en) | SAR image change detection method based on self-attention image fusion and DEC | |
| CN103236063A (en) | Multi-scale spectral clustering and decision fusion-based oil spillage detection method for synthetic aperture radar (SAR) images | |
| CN101303764A (en) | Method for self-adaption amalgamation of multi-sensor image based on non-lower sampling profile wave | |
| CN106408581B (en) | A kind of quick three-dimensional point cloud lines detection method | |
| CN106611421A (en) | SAR image segmentation method based on feature learning and sketch line constraint | |
| CN102176014B (en) | Method for detecting urban region change based on multi-temporal SAR (synthetic aperture radar) images | |
| CN110414509A (en) | Stop Ship Detection in harbour based on the segmentation of extra large land and feature pyramid network | |
| CN107247927A (en) | A kind of remote sensing images coastline information extracting method and system based on K-T Transformation | |
| CN106157330A (en) | A kind of visual tracking method based on target associating display model | |
| CN111414954A (en) | Rock image retrieval method and system | |
| CN102073867A (en) | Sorting method and device for remote sensing images | |
| CN105205816A (en) | Method for extracting high-resolution SAR image building zone through multi-feature weighted fusion | |
| CN106251375A (en) | A kind of degree of depth study stacking-type automatic coding of general steganalysis | |
| CN109002792B (en) | SAR image change detection method based on layered multi-model metric learning | |
| CN112989940B (en) | Raft culture area extraction method based on high-resolution third satellite SAR image |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| C17 | Cessation of patent right | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20080220 |