CN100416599C - Not supervised classification process of artificial immunity in remote sensing images - Google Patents

Not supervised classification process of artificial immunity in remote sensing images Download PDF

Info

Publication number
CN100416599C
CN100416599C CNB2006100195081A CN200610019508A CN100416599C CN 100416599 C CN100416599 C CN 100416599C CN B2006100195081 A CNB2006100195081 A CN B2006100195081A CN 200610019508 A CN200610019508 A CN 200610019508A CN 100416599 C CN100416599 C CN 100416599C
Authority
CN
China
Prior art keywords
antibody
memory
affinity
image
remote sensing
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Expired - Fee Related
Application number
CNB2006100195081A
Other languages
Chinese (zh)
Other versions
CN1873661A (en
Inventor
钟燕飞
张良培
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuhan University WHU
Original Assignee
Wuhan University WHU
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Wuhan University WHU filed Critical Wuhan University WHU
Priority to CNB2006100195081A priority Critical patent/CN100416599C/en
Publication of CN1873661A publication Critical patent/CN1873661A/en
Application granted granted Critical
Publication of CN100416599C publication Critical patent/CN100416599C/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Abstract

The present invention relates to a remote sensing image artificial immunity non-monitoring classifying method which is characterized in that (1) the remote sensing image to be classified is opened, and the algorithm parameters are inputted; (2) various initial antibody populations and memory antibodies which are classified in a non-monitoring way are selected and stored in various antibody arrays and memory antibody arrays; (3) image antigens can continuously carry out an artificial immunity system training until accomplishing the training of the whole image, and then, (4) a user can judge whether training stop conditions are satisfied or not, if false, then the next iteration is started from the step (3), else the artificial immunity non-monitoring classification is accomplished to output the classification result. The method of the present invention inherits the biological attributes of the bio-immunity system, and moreover, the method has capacities of self-organizing, self-learning, self-recognizing and self-memorizing. Furthermore, the method has the advantages of high intelligence and executing efficiency in practice. The preset invention is suitable for the multi-spectrum and high-spectrum remote sensing image non-monitoring classification, and the method can effectively enhance the accuracy of the remote sensing image non-monitoring classification.

Description

A kind of not supervised classification process of artificial immunity of remote sensing image
Technical field
The invention belongs to the remote sensing image processing technology field, especially a kind of Method for Unsupervised Remote Imagery Classification based on artificial immune system.
Background technology
The remote sensing image unsupervised classification is meant that people do not apply any priori to assorting process in advance, and only the regularity of distribution of the spectral signature of authority remote sensing image atural object is carried out classification blindly naturally with it.Its sorting result is just distinguished different classes of, and the attribute of uncertain classification, and its attribute is by all kinds of spectral response curves being analyzed afterwards, and the back of comparing with on-site inspection is determined.Traditional remote sensing image unsupervised classification mainly adopts K mean algorithm or ISODATA algorithm.
The basic thought of K mean algorithm is: by iteration, move the center of each reference category, till obtaining best cluster result.This algorithm can make that all samples are to distances of clustering centers quadratic sum minimum in the cluster territory.The characteristics of K mean algorithm are that the result of K mean algorithm is subjected to the number K of selected cluster centre and the influence that initial cluster center is selected, also be subjected to the influence of the geometric properties and the ordering of atural object, in fact need sound out different K values and select different initial cluster centers.If the geometrical property of atural object shows that they can form the fritter isolated region of several apart from each others, then algorithm multipotency convergence.
The data analysis algorithm of iteration self-organization (Iterative Selforganizing Data AnalysisTechniques Algorithm) also claims the ISODATA algorithm.This algorithm and K mean algorithm have similarity, be that cluster centre also is that interative computation by sample average decides, but the ISODATA algorithm has added some exploratory step and human-computer interaction functions, can draw the resulting experience of intermediate result, mainly be in iterative process, a class can be divided into two, also two classes may be united two into one, that is " self-organization ", so this algorithm has had didactic characteristics.
Relevant document: Sun Jiabing, oxazepan, Guan Zequn. remote sensing principle methods and applications [M]. Beijing: Mapping Press, 1997; Soup Guoan, Zhang Youshun, Liu Yongmei. process in remote sensing digital image processing etc. [M]. Beijing: Science Press, 2004; Zhao's inch. remote sensing application analysis principle and method [M]. Beijing: Science Press, 2003; Campbell, J.B., Introduction to Remote Sensing[M] .London:Taylor ﹠amp; Francis, 2002.
Traditional Method for Unsupervised Remote Imagery Classification adopts simple average to carry out iteration as cluster centre in non-supervision iterative process, and this cluster centre has locality, and fails to consider the of overall importance of sample.And traditional not supervised classification is intelligent relatively poor, and the nicety of grading that obtains in the image unsupervised classification of reality is often not high.
Artificial immune system (Artificial Immune System, be called for short AIS) is to be subjected to the inspiration of Immune System and a kind of novel intelligence computation method that produces.In the past few years, the application of AIS has expanded to numerous areas such as information security, pattern-recognition, machine learning, data mining gradually, demonstrates powerful information processing of AIS and problem solving ability and wide research and application prospect.Relevant document: D.Dasgupta, Artificial Immune Systems and Their Applications, Germany:Springer, 1999; L.N.de Castro and J.Timmis, Artificial Immune systems:A NewComputational Intelligence Approach, London, U.K.:Springer-Verlag, 2002; J.Timmis, M.Neal, and J.E.Hunt, " An artificial immune system for data analysis, " Biosystem, 55 (1/3), 2000; Xiao Renbin, Wang Lei. artificial immune system: principle, model, analysis and prospect [J]. Chinese journal of computers, 2002,25 (12).
AIS is the very strong optimisation technique of a kind of self-adaptation, numerous attributes of Immune System have been inherited, have self-organization, self study, self-identifying, the ability of memory certainly, therefore it can provide 90% the hunting zone that reaches optimum solution fast, thereby can obtain globally optimal solution comparatively fast more accurately, this be other optimisation technique can't be obtained.The using artificial immune algorithm by the optimizing process that repeats, has highly intelligently, can obtain optimum solution from self-adaptation soon.It can improve the accuracy of classification, the operation time of minimizing algorithm.Can be referring to relevant document: L.N.De Castro and F.J.Von Zuben, " Learning andoptimization using the clonal selection principle; " IEEE Trans. onEvolutionary Computation, Vol.6 (3): 2002; Atkinson P M, Lewis P.Geostatistical classification for remote sensing:an introduction[J] .Computers ﹠amp; Geosciences, 26,2000; Adams D.How the immune system worksand why it causes autoimmune diseases[J] .Immunology Today, 17 (7), 1996; J.H.Carter, " The immune system as a model for pattern recognition andclassification, " Journal of the American Medical InformaticsAssociation, Vol.7 (3), 2000.
Yet in the remote sensing image unsupervised classification, AIS is not also well used.The characteristic of what use is made of artificial immune system provides can the high-speed and high-efficiency intelligent Method for Unsupervised Remote Imagery Classification that obtains classification results, is to realize the intelligentized technical matters that needs to be resolved hurrily of remote sensing image unsupervised classification.
Summary of the invention
The present invention seeks to utilize the advantage of artificial immune system, a kind of classification of remote-sensing images precision not supervised classification that is used to improve is provided.
For achieving the above object, the invention provides the not supervised classification process of artificial immunity of remote sensing image:
Step 1 is opened remote sensing image to be classified by the remote sensing image handling procedure, and the input algorithm parameter;
Step 2 according to the class categories number of setting, is chosen all kinds of initial antibodies populations and the memory antibody of unsupervised classification, deposits each antibody-like array and memory antibody array in;
Step 3 is carried out artificial immune system training to image antigen, is all finished by training up to the view picture image, and the training of each image antigen is comprised following 5 steps,
Step 3.1 is calculated the affinity of image antigen to every class memory antibody, this image antigen is adjudicated under the memory antibody with maximum affinity in the classification, and write down this memory antibody for mating most antibody;
Step 3.2 is calculated the affinity of all antibody in the set of this image antigen and similar memory antibody, and the quantity with the selection antibody number set from the memory antibody set is selected and several the highest antibody of antigen affinity, produces a new antibody and gathers;
Step 3.3 is carried out clone operations to the antibody that chooses with certain cloning efficiency, produces the clonal antibody set, and mutation operation is carried out in clone's set, the antibody that obtains making a variation set;
Step 3.4 is calculated the affinity of all variation back antibody in this image antigen and the variation antibody set, therefrom select have maximum affinity antibody as candidate's memory antibody;
Step 3.5, judge the affinity size of mating antibody most that obtains in candidate's memory antibody and the step 3.1, if greater than coupling memory antibody affinity level, then candidate's memory antibody enters in the antibody memory bank, deposit the memory antibody array in, and then calculate between the two distance, if distance within the allowed band of setting then mate memory antibody and from the memory antibody array, remove, this scope is obtained by the product of distance threshold and distance proportion threshold value;
Step 4 judges whether to satisfy the training stop condition, and is if do not satisfy then begin to carry out next iteration from step 3, satisfied then the artificial immunity unsupervised classification is finished the output category result.
And algorithm parameter includes class categories number, maximum iteration time, selection antibody number, cloning efficiency and distance proportion threshold value, and the Rule of judgment that adopts maximum iteration time to stop as training carries out clone operations according to cloning efficiency to the antibody that chooses.
And, carrying out after step 3.4 calculates the affinity of all variation antibody in the set of variation antibody, from the set of variation antibody, select d the highest variation antibody of affinity to replace d original antibody that affinity is minimum in the original antibody array, wherein d value value is d=β * N, β is a replacement rate, and N is an antibody sum in the original antibody population.
And, adopt the minimax distance to select heart method to choose all kinds of initial antibodies populations and the memory antibody of unsupervised classification.
The present invention carries out the iteration classification by artificial immune system to remote sensing image, has showed the distribution situation of all kinds of atural objects more accurately, has avoided the distribution locality; On obtaining once on the atural object distributed basis, evolving by Immune Clone Selection algorithm in the artificial immune system obtains the cluster centre or the memory antibody of each classification, has avoided adopting each classification average of simple computation and the limitation that obtains cluster centre; Adopt immune threshold value to reduce the redundancy of artificial immune system, accelerate convergence of algorithm speed, reduce the unsupervised classification time of remote sensing image.The inventive method has been inherited the Immune System biological attribute, have self-organization, self study, self-identifying, the ability of memory certainly, intelligent height, carry out the efficient height in the actual motion, be applicable to multispectral, target in hyperspectral remotely sensed image unsupervised classification, can effectively improve the unsupervised classification precision of remote sensing image.
Description of drawings
Fig. 1 Immune Clone Selection principle schematic;
Fig. 2 artificial immunity kind of the present invention group model;
The artificial immunity unsupervised classification parameter of Fig. 3 embodiment of the invention is provided with figure;
Fig. 4 embodiment of the invention principal function flow chart;
Fig. 5 embodiment of the invention initialization function program block diagram;
The non-supervision immunoevolution of Fig. 6 embodiment of the invention function program block diagram;
Fig. 7 embodiment of the invention data base evolution function program block diagram.
Embodiment
For the ease of understanding the present invention, at first provide theoretical foundation of the present invention:
One of critical function of human immune system is to remove external foreign matter by producing antibody (antibody), and foreign matter can be microorganism (bacterium, virus etc.), special-shaped haemocyte, grafting device official rank, and they all are called antigen (antigen).Immune basic composition is lymphocyte or white blood cell.These special cells mainly can be divided into B cell and T cell two big classes.These two kinds of cells all have own unique ecologic structure and produce many Y type antibody from their surface and kill antigen.
In order to explain the formation mechanism of antibody, some scholars propose template theory the earliest, propose side-chain theory afterwards again, but they all can not form mechanism by reasonable dismissal antibody.Proposition up to the Immune Clone Selection theory just makes antibody form the explanation that mechanism obtains satisfaction.
Immune system at first selects cell (autoantigen) and those elements (exotic antigen) that does not belong to self to self to distinguish by feminine gender when identification antigen, eliminate external cell or molecule by distinguishing the back immune system, and the cell of system itself is not handled.After carrying out the feminine gender selection, thereby immune system is carried out Immune Clone Selection generation antibody for exotic antigen.
The Immune Clone Selection main contents are: after the identification of lymphocyte realization to antigen, the B cell is activated, and breeds and duplicate generation B cell clone, and clone cell experiences mutation process subsequently, and the former specific antibody that has creates antagonism.The Immune Clone Selection theoretical description fundamental characteristics of acquired immunity, and have only the immunocyte of successfully discerning antigen just to be bred, the immunocyte after the experience variation is divided into two kinds of thick liquid cell (antibody mediated effect cell) and memory cells.
The principal character of Immune Clone Selection is: the Immune Clone Selection correspondence the process of an affinity maturation, promptly this process be one to the lower individuality of antigen affinity under the effect of Immune Clone Selection mechanism, experience propagation duplicate with mutation process after, because of affinity progressively increases slowly maturescent process.From this process, Immune Clone Selection is the evolutionary process of Darwin's formula in essence, and this process can be by adopting operator and the controlling mechanism realizations of corresponding colony such as intersection, variation.The Immune Clone Selection principle schematic as shown in Figure 1.
Based on this Immune Clone Selection theory, the invention provides the not supervised classification process of artificial immunity of remote sensing image, remote sensing image processing data complexity, workload is big, generally need to adopt computer means to realize, therefore technical solution of the present invention has adopted computer program and computerese to be described, and for example the antibody array comes down in order to explain the antibody set that a content changes according to evolution.The claimed technical scheme of the present invention is not limited to the computer program flow process, and should comprise that other are equal to the replacement means.Technical solution of the present invention is as follows:
Step 1 is opened remote sensing image to be classified by the remote sensing image handling procedure, and the input algorithm parameter;
Step 2 according to the class categories number of setting, is chosen all kinds of initial antibodies populations and the memory antibody of unsupervised classification, deposits each antibody-like array and memory antibody array in;
Step 3 is carried out artificial immune system training to image antigen, is all finished by training up to the view picture image, and the training of each image antigen is comprised following 5 steps,
Step 3.1 is calculated the affinity of image antigen to every class memory antibody, this image antigen is adjudicated under the memory antibody with maximum affinity in the classification, and write down this memory antibody for mating most antibody;
Step 3.2 is calculated the affinity of all antibody in the set of this image antigen and similar memory antibody, and the quantity with the selection antibody number set from the memory antibody set is selected and several the highest antibody of antigen affinity, produces a new antibody and gathers;
Step 3.3 is carried out clone operations to the antibody that chooses with certain cloning efficiency, produces the clonal antibody set, and mutation operation is carried out in clone's set, the antibody that obtains making a variation set;
Step 3.4 is calculated the affinity of all variation back antibody in this image antigen and the variation antibody set, therefrom select have maximum affinity antibody as candidate's memory antibody;
Step 3.5, judge the affinity size of mating antibody most that obtains in candidate's memory antibody and the step 3.1, if greater than coupling memory antibody affinity level, then candidate's memory antibody enters in the antibody memory bank, deposit the memory antibody array in, and then calculate between the two distance, if distance within the allowed band of setting then mate memory antibody and from the memory antibody array, remove, this scope is obtained by the product of distance threshold and distance proportion threshold value;
Step 4 judges whether to satisfy the training stop condition, and is if do not satisfy then begin to carry out next iteration from step 3, satisfied then the artificial immunity unsupervised classification is finished the output category result.
Because the value in image antigen and the antibody on each wave band is a real number, and in order to consider the speed of algorithm, therefore the Immune Clone Selection algorithm adopts Gauss's real number variation mode in the present invention: algorithm is as follows:
The Gaussian mutation operator is adopted more in the variation algorithm of real coding.Specific algorithm is as follows:
In Gaussian mutation, the body one by one of colony is by a solution vector s=(v 1, v 2..., v n) and a disturbance vector σ=(σ 1, σ 2..., σ n) form, this disturbance vector is the control vector of variation solution vector, and it also constantly will make a variation itself, if (s σ) is selected individuality, then can produce new individuality after the variation (s ', σ ') by following formula:
σ′ i＝σ i·exp(N i(0，Δσ))
v′ i＝v i+N(0，σ′ i)，?i＝1，2，…，n
Here Δ σ is called secondary step-length controlled variable, N i(0, Δ σ), i=1,2 ..., n is that separate average is 0, variance is the random number that meets normal distribution of Δ σ.Another adjusts σ ' iPrinciple be exactly the 1/5 successful rule that adopts Rechenberg to propose.That is: successful variation probability should be 1/5, if actual successful ratio then increases σ ' greater than 1/5 iValue, and if be lower than 1/5, then reduce σ ' iValue.According to the research of Schwefel, suppose and will carry out the 10n variation, just carry out once such inspection after then every n variation.σ ' iThe increase and decrease scale factor be respectively (1.0/0.85) and 0.85.
In order to describe algorithm better, the present invention proposes a kind of new artificial immunity cell population model, i.e. the AB model.The antibody aggregation model of a classification of AB model representation, wherein ab describes single antibody, ab ∈ AB.Accompanying drawing 2 has been described the AB model of a classification, and AB comprises such a plurality of immune antiboidy ab (representing with small circle in the accompanying drawing 2) and memory antibody mc (representing with rhombus in the accompanying drawing 2) in immune system, memory set MC mark, mc ∈ MC.In classification of remote-sensing images, memory set MC has determined the recognition capability of whole AB, and in accompanying drawing 2, σ represents the identification range of MC.In the σ scope, AB can discern antigen ag.The embodiment of the invention adopts antibody array ABArray and memory antibody array MCArray to deposit the corresponding antibodies data.
The present invention has designed computer program and has realized unsupervised classification task of the present invention, program adopts not supervised classification process of artificial immunity that image antigen I mageAg is trained, if arrive end condition then classify and finish, otherwise loop iteration training image antigen satisfies up to end condition.In order to make program succinct, and convenient the combination with existing remote sensing image process software and the program realization, the function call thinking adopted in large quantities.
Describe technical solution of the present invention in detail below in conjunction with the concrete implementation step of embodiment:
(1) utilize remote sensing image processing system, by input image width, highly, wave band number and data type open input remote sensing image to be classified, and saves as image antigen I mageAg array.This process belongs to the image input process, and the realization program is not introduced in detail.Can be provided with by program during concrete enforcement the algorithm parameter input frame is provided, referring to accompanying drawing 3.After ejecting the algorithm parameter input frame, the required parameter of input algorithm, mainly comprise: class categories is counted nc, and maximum iteration time nIte selects antibody to count n, cloning efficiency Clonal_rate, replacement rate β, distance proportion threshold value DTS.Set up the executive routine that activates this algorithm behind the algorithm parameter, the principal function program flow chart is seen accompanying drawing 4.Certainly, some algorithm parameter also can be imported immediately or revise in the computation process afterwards.At present, the computer utility in remote sensing field is very general, it is conventional means that the digital picture that adopts the remote sensing image handling procedure that remote sensing is obtained is handled, and has possessed the remote sensing image handling procedure of opening image function and sampling instrument substantially and all can use for the invention process.During concrete enforcement, can be combined into one other step required functions of the basic function of remote sensing image handling procedure and the inventive method, can realize by computer programming.
(2) adopt the minimax distance to select heart method to choose unsupervised classification required all kinds of initial antibodies populations and memory antibody, deposit antibody array ABArray and memory antibody array MCArray in, the minimax distance is selected heart method program to adopt and is called initialization function Initialization () realization, and FB(flow block) is seen accompanying drawing 5.The specific embodiment step is: obtain the sampling collection of pixels X=X1, X2 ... Xn}; The unique point of getting arbitrary pixel in the sampling collection of pixels calculates Z1 sample point farthest as second initial category center Z2 as first initial category center Z1; Calculate the distance at existing center for remaining each sample point, getting minor increment is the representative distance; The maximal value of relatively representing distance is next classification center Zi; Judge at last whether the classification number meets the demands, then call non-supervision immunoevolution function U IEvolution () if satisfy, do not satisfy then return continuation by sample point calculate existing centre distance represent apart from, establish next classification center, count nc up to satisfying classification.
(3) image antigen is carried out the artificial immune system training, all finished by training up to the view picture image, concrete enforcement is to call non-supervision immunoevolution function U IEvolution () to realize.Training to each image antigen comprises 4 steps, so following 5 steps are carried out in the circulation that shows as substantially of specific procedure:
(3.1) certain the image antigen ag among the image antigen array ImageAg is input in the algorithm, calculates the affinity f of antigen to every class memory antibody k, computing formula be labeled as affinity (ag, mc).And antigen adjudicated under the memory antibody with maximum affinity in the classification, be assumed to be the k class, write down this antibody for mating most antibody C JoinCorrespondingly, the set of k class memory antibody is labeled as antibody population AB k
C Join=arg max Mc ∈ MCAffinity (ag, mc)
Suppose that wherein two pixels are x = { x 1 , x 2 , · · · , x N b } With y = { y 1 , y 2 , · · · , y N b } , N bExpression wave band sum, then affinity computing formula between the two is as follows: (x y) adopts spectrum angle method to calculate to its middle distance dis.
affinity ( x , y ) = exp ( - dis ( x , y ) 2 σ i 2 )
dis ( x , y ) = α = cos - 1 [ Σ i = 1 N b x i y i [ Σ i = 1 N b ( x i ) 2 ] 1 2 [ Σ i = 1 N b ( y i ) 2 ] 1 2 ]
(3.2) calculate certain image antigen to corresponding certain antibody-like set A B kIn the affinity of all antibody, concentrate from antibody and to select and n the highest antibody of antigen affinity, produce a new antibody set A B { n} k
(3.3) the selected antibody that goes out of n is carried out clone operations by certain cloning efficiency, produce clonal antibody set C kWherein affinity is high more, and clone's number is high more.All selected antibody clonings that go out add up to:
NumClones = Σ i = 1 n round ( Clonal _ rate · affinity ( ag , ab i )
Wherein NumClones represents clone's sum.The cloning efficiency parameters C lonal_rate decision clone's who is provided with multiple, typical parameter value is 10, function round () expression rounds.
To clonal antibody set C kCarry out mutation operation, produce the antibody set MU after making a variation k, wherein and the affinity between the antigen big more, the antibody variation chance is more little.Aberration rate determines in the following manner:
Aberration rate mutate_rate=1-affinity (ag, ab i)
Clone's set C kThe mutation operation formula is as follows:
MU i k = C i k + mutation _ rate · N ( 0,1 ) , i ∈ [ 1 , NumClones ]
Wherein N (0,1) expression average is 0, and contrast is 1 Gaussian random variable.
The calculation procedure FB(flow block) is seen accompanying drawing 6, and the mode that variation function mutate () is called in the employing of specific embodiment realizes that the functional operation program step is as follows: enter variation function mutate () inlet, input clonal antibody set C k, adopt variable i labelled antibody, j labelled antibody wave band, by being i, j assignment and after operation, adding 1 mode certainly, be clonal antibody set C one by one kIn certain wave band of certain antibody B assignment that makes a variation; Variable nrandom=N (0,1) applies mechanically certain wave band variation value that the mutation operation formula can be calculated certain antibody B Mab i j = ab i k + mutation _ rate · nrandom ; Variable j value＞when total wave band is counted is for antagonist B variation assignment finishes, and during the total NumClones of variable i value＞clone clonal antibody gathered C kVariation finishes.The antibody of output variation at last set MU k
(3.4) calculate image antigen ag and variation antibody collection MU kIn all the variation antibody affinity f k *, therefrom select have the highest affinity antibody as candidate's memory antibody C Wait
In order to increase the diversity of antibody, calculate affinity f in step (3.4) k *After, from variation antibody collection MU kIn select d the highest variation antibody of affinity and replace original antibody collection AB kThe d of a middle affinity minimum original antibody.Wherein, d=β * N, β are replacement rate, and N is an antibody sum in the antibody population.If establish the β value is 0, has then saved replacement operation.
(3.5) judge that image antigen ag is to C WaitAnd C JoinIrritation level, if affinity affinity (ag, C Wait) greater than affinity (ag, C Join) then with C WaitAdd among the data base MCArray.And then calculating C WaitAnd C JoinBetween distance, the present invention has set apart from allowed band, (distance threshold, DT) (distance threshold scalar, product DTS) obtains this scope, if apart from dis (C with the distance proportion threshold value by distance threshold Wait, C Join) less than DT*DTS, then with C JoinFrom data base MCArray, remove, to guarantee that memory cell in the data base is within certain quantitative range.Calculation procedure realizes by calling data base evolution function DevelopMCPop (), and FB(flow block) is referring to accompanying drawing 7, affinity (ag, C Wait), affinity (ag, C Join), dis (C Wait, C Join) respectively assignment give variable CandAff, MatchAff, CellDis so that computing machine carries out arithmetic operation.Wherein distance threshold DT is obtained by following formula:
DT = Σ j = 1 N b ( MAX j - MIN j )
N wherein bBe remote sensing image wave band number, MAX jBe the maximal value of j wave band of remote sensing image, MIN jMinimum value for j wave band of remote sensing image.
After training is finished to image antigen, then next image antigen is carried out the artificial immune system training, up to the view picture image all training finish, then finish iterative process one time.
(5) judge whether to satisfy stop condition, if do not satisfy then begin to carry out next iteration from step (3), otherwise the artificial immunity unsupervised classification is finished the output category result.Whether the stop condition of the embodiment of the invention is for reaching maximum iteration time nIte, and this parameter can be set in the step (1) that program run begins in advance.
After adopting function call thought to be technical scheme programming of the present invention, whole program is implemented structure and is: import remote sensing image to be classified; Definition image antigen array ImageAg; Algorithm parameter is set; Calling classification is counted UAIclassifier (); Enter function U AIclassifier () inlet; Call initialization function Initialization (), choose unsupervised classification required all kinds of initial antibodies populations and memory antibody; G classifies to the image antigen A; Call non-supervision immunoevolution function U IEvolution (); Call data base evolution function DevelopMCPop (); Judge whether to satisfy end condition, if do not satisfy then begin to carry out next iteration from step (3), otherwise the artificial immunity unsupervised classification is finished the output category result, obtains the classification remote sensing image.

Claims (4)

1. the not supervised classification process of artificial immunity of a remote sensing image is characterized in that:
Step 1 is opened remote sensing image to be classified by the remote sensing image handling procedure, and the input algorithm parameter;
Step 2 according to the class categories number of setting, is chosen all kinds of initial antibodies populations and the memory antibody of unsupervised classification, deposits each antibody-like array and memory antibody array in;
Step 3 is carried out artificial immune system training to image antigen, is all finished by training up to the view picture image, and the training of each image antigen is comprised following 5 steps,
Step 3.1 is calculated the affinity of image antigen to every class memory antibody, this image antigen is adjudicated under the memory antibody with maximum affinity in the classification, and write down this memory antibody for mating most antibody;
Step 3.2 is calculated the affinity of all antibody in the set of this image antigen and similar memory antibody, and the quantity with the selection antibody number set from the memory antibody set is selected and several the highest antibody of antigen affinity, produces a new antibody and gathers;
Step 3.3 is carried out clone operations to the antibody that chooses with certain cloning efficiency, produces the clonal antibody set, and mutation operation is carried out in clone's set, the antibody that obtains making a variation set;
Step 3.4 is calculated the affinity of all variation back antibody in this image antigen and the variation antibody set, therefrom select have maximum affinity antibody as candidate's memory antibody;
Step 3.5, judge the affinity size of mating antibody most that obtains in candidate's memory antibody and the step 3.1, if greater than coupling memory antibody affinity level, then candidate's memory antibody enters in the antibody memory bank, deposit the memory antibody array in, and then calculate between the two distance, if distance within the allowed band of setting then mate memory antibody and from the memory antibody array, remove, this scope is obtained by the product of distance threshold and distance proportion threshold value;
Step 4 judges whether to satisfy the training stop condition, and is if do not satisfy then begin to carry out next iteration from step 3, satisfied then the artificial immunity unsupervised classification is finished the output category result.
2. not supervised classification process of artificial immunity as claimed in claim 1, it is characterized in that: algorithm parameter includes class categories number, maximum iteration time, selection antibody number, cloning efficiency and distance proportion threshold value, the Rule of judgment that adopts maximum iteration time to stop as training carries out clone operations according to cloning efficiency to the antibody that chooses.
3. not supervised classification process of artificial immunity as claimed in claim 1, it is characterized in that: carrying out after step 3.4 calculates the affinity of all variation antibody in the set of variation antibody, from the set of variation antibody, select d the highest variation antibody of affinity to replace d original antibody that affinity is minimum in the original antibody array, wherein d value value is d=β * N, β is a replacement rate, and N is an antibody sum in the original antibody population.
4. as claim 1 or 2 or 3 described not supervised classification process of artificial immunity, it is characterized in that: adopt the minimax distance to select heart method to choose all kinds of initial antibodies populations and the memory antibody of unsupervised classification.
CNB2006100195081A 2006-06-29 2006-06-29 Not supervised classification process of artificial immunity in remote sensing images Expired - Fee Related CN100416599C (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CNB2006100195081A CN100416599C (en) 2006-06-29 2006-06-29 Not supervised classification process of artificial immunity in remote sensing images

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CNB2006100195081A CN100416599C (en) 2006-06-29 2006-06-29 Not supervised classification process of artificial immunity in remote sensing images

Publications (2)

Publication Number Publication Date
CN1873661A CN1873661A (en) 2006-12-06
CN100416599C true CN100416599C (en) 2008-09-03

Family

ID=37484132

Family Applications (1)

Application Number Title Priority Date Filing Date
CNB2006100195081A Expired - Fee Related CN100416599C (en) 2006-06-29 2006-06-29 Not supervised classification process of artificial immunity in remote sensing images

Country Status (1)

Country Link
CN (1) CN100416599C (en)

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102073882A (en) * 2011-01-27 2011-05-25 武汉大学 Method for matching and classifying spectrums of hyperspectral remote sensing image by DNA computing
CN103310224A (en) * 2013-06-08 2013-09-18 上海电机学院 Unsupervised artificial immune classification method
CN104615679A (en) * 2015-01-21 2015-05-13 华侨大学 Multi-agent data mining method based on artificial immunity network
TWI672637B (en) * 2018-05-03 2019-09-21 長庚醫療財團法人林口長庚紀念醫院 Patern recognition method of autoantibody immunofluorescence image
CN109376779B (en) * 2018-10-19 2020-10-27 西安交通大学 Complex electromechanical system service mode automatic identification method
CN109948657A (en) * 2019-02-25 2019-06-28 湖北同诚通用航空有限公司 Withered trees recognition methods and equipment based on visible images
CN114092795A (en) * 2020-07-31 2022-02-25 中国矿业大学(北京) Crop disease grade evaluation method based on vegetation index normalization
CN112907484B (en) * 2021-03-18 2022-08-12 国家海洋信息中心 Remote sensing image color cloning method based on artificial immune algorithm

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO1999001843A1 (en) * 1997-07-01 1999-01-14 Applied Spectral Imaging Ltd. Method for remote sensing analysis by decorrelation statistical analysis and hardware therefor
WO2002025574A2 (en) * 2000-09-22 2002-03-28 Http Insights Limited Data clustering methods and applications
US20030026484A1 (en) * 2001-04-27 2003-02-06 O'neill Mark Automated image identification system
CN1790379A (en) * 2004-12-17 2006-06-21 中国林业科学研究院资源信息研究所 Remote sensing image decision tree classification method and system

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO1999001843A1 (en) * 1997-07-01 1999-01-14 Applied Spectral Imaging Ltd. Method for remote sensing analysis by decorrelation statistical analysis and hardware therefor
WO2002025574A2 (en) * 2000-09-22 2002-03-28 Http Insights Limited Data clustering methods and applications
US20030026484A1 (en) * 2001-04-27 2003-02-06 O'neill Mark Automated image identification system
CN1790379A (en) * 2004-12-17 2006-06-21 中国林业科学研究院资源信息研究所 Remote sensing image decision tree classification method and system

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
The immune system as a model for pattern recognition andclassification. Jerome H. Carter, MD.Journal of the American Mddical Informatic Association,Vol.7 No.1. 2000
The immune system as a model for pattern recognition andclassification. Jerome H. Carter, MD.Journal of the American Mddical Informatic Association,Vol.7 No.1. 2000 *
基于克隆选择的多光谱遥感影像分类算法. 钟燕飞，张良培，龚健雅，李平湘.中国图象图形学报，第10卷第1期. 2005
基于克隆选择的多光谱遥感影像分类算法. 钟燕飞，张良培，龚健雅，李平湘.中国图象图形学报，第10卷第1期. 2005 *

Also Published As

Publication number Publication date
CN1873661A (en) 2006-12-06

Similar Documents

Publication Publication Date Title
He et al. An end-to-end steel surface defect detection approach via fusing multiple hierarchical features
CN100416599C (en) Not supervised classification process of artificial immunity in remote sensing images
CN106779087B (en) A kind of general-purpose machinery learning data analysis platform
Zhou et al. Extracting symbolic rules from trained neural network ensembles
Zhang et al. A novel cascade ensemble classifier system with a high recognition performance on handwritten digits
Razmjooy et al. A study on metaheuristic-based neural networks for image segmentation purposes
CN110414554A (en) One kind being based on the improved Stacking integrated study fish identification method of multi-model
CN106919951A (en) A kind of Weakly supervised bilinearity deep learning method merged with vision based on click
CN103473786B (en) Gray level image segmentation method based on multi-objective fuzzy clustering
Pal Ensemble learning with decision tree for remote sensing classification
Cloppet et al. ICFHR2016 competition on the classification of medieval handwritings in latin script
CN111144496A (en) Garbage classification method based on hybrid convolutional neural network
Kaur et al. A survey on machine learning algorithms
CN103116766A (en) Increment neural network and sub-graph code based image classification method
CN105320967A (en) Multi-label AdaBoost integration method based on label correlation
CN109815920A (en) Gesture identification method based on convolutional neural networks and confrontation convolutional neural networks
CN105046323B (en) Regularization-based RBF network multi-label classification method
CN110288028A (en) ECG detecting method, system, equipment and computer readable storage medium
CN103631753A (en) Progressively-decreased subspace ensemble learning algorithm
Untoro et al. Evaluation of decision tree, k-NN, Naive Bayes and SVM with MWMOTE on UCI dataset
CN114398485B (en) Expert portrait construction method and device based on multi-view fusion
Yanmin et al. An artificial immune network clustering algorithm for mangroves remote sensing image
Gillala et al. An efficient chaotic salp swarm optimization approach based on ensemble algorithm for class imbalance problems
Patidar et al. Decision tree C4. 5 algorithm and its enhanced approach for educational data mining
Dorobanţiu et al. A novel contextual memory algorithm for edge detection

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: 20080903

Termination date: 20110629