CN104598898B - A kind of Aerial Images system for rapidly identifying and its method for quickly identifying based on multitask topology learning - Google Patents

Abstract

The invention discloses a kind of Aerial Images system for rapidly identifying and its method for quickly identifying based on multitask topology learning, its feature includes：Feature construction module, topological characteristic structure module, topological characteristic projection module, multitask topology learning module, feature selection module and Classification and Identification module；Feature construction module is used to obtain Aerial Images and trains the topological characteristic matrix of Aerial Images；Topological characteristic structure module is used for the topological characteristic projection matrix for obtaining training Aerial Images；Multitask topology learning module is used for the topological dictionary for obtaining Aerial Images；Feature selection module is used for the tuple of feature selecting two for obtaining training Aerial Images；Classification and Identification module is used to build topological Boosting learning models and realizes the classification of target Aerial Images.The present invention can quickly identify the image classification of Aerial Images using multi-task learning framework, improve nicety of grading and efficiency.

Description

A kind of Aerial Images system for rapidly identifying based on multitask topology learning and its quick Recognition methods

Technical field

The invention belongs to image procossing, computer vision and mode identification technology, relates generally to a kind of based on more The Aerial Images system for rapidly identifying and its method for quickly identifying of business topology learning.

Background technology

Important component of the Aerial Images as remote sensing images, with its substantial amounts of data message and geometric resolution by Universal attention, therefore it is great that category significance is identified to Aerial Images.The quick identification classification of Aerial Images is a kind of intelligence Energy image analysis method, its purpose are exactly to carry out fast and accurately identifying and classifying to Aerial Images, are computer vision One of field hot research topic.The classification of Aerial Images can apply to scene mark, video monitoring, robot path rule Draw, abnormality detection etc..

With the development of Morden Image Processing Technology, Aerial Images sorting technique has also obtained huge development, but still So face some following problem：

First, in existing method, zone bit distribution map representativeness is not high in the Aerial Images of structure, it is impossible to accurately represents boat Clap the constituent of image.

Such as 2011, the author such as Duchenne is in top international conference IEEE International Conference Delivered on Computer Vision (ICCV)《AGraph-Matching Kernel for Object Categorization》A kind of objective classification method of the graph structure based on core matching proposed in this article, in this method, The node on behalf of figure image lattice, and what geometry connection side represented is the annexation between grid, is based on such one kind The classification of target is realized in the matching of the figure of grid, but because the representativeness of the graph structure of the Aerial Images of structure is not high, it is impossible to The constituent of Aerial Images is accurately represented, reduces the efficiency of Aerial Images classification.

2nd, the topological structure for considering image is lacked in many Aerial Images assorting processes.

Such as 2010, the author such as Yong Jae Lee is in top international conference IEEE Conference on Delivered on Computer Vision and Pattern Recognition (CVPR)《Object-Graphs for Context-Aware Category Discovery》The image classification method proposed in this article, the section of figure in this method Point represents known and unknown target, and the semanteme of each unknown object utilizes the semantic deduction of its k-space arest neighbors known target Out, but only only account for the knowledge of spatially adjacent part, i.e., it is semantic to infer another important clue.And topological structure exists Do not consider during classification, so as to reduce the efficiency of image classification.

3rd, existing many Aerial Images identifications for considering topological relation are classified based on of random walk and tree strategy Match somebody with somebody, limit the matching range of topological relation in Aerial Images.

Such as 2007, the author such as Harchaoui is in international conference IEEE Conference on Computer Delivered on Vision and Pattern Recognition《Image Classification with Segmentation Graph Kernels》A kind of image classification method proposed in this article, this method utilize random walk core and tree kernel function Method use the space topological of random walk that the adjacent domain of finite sequence goes to extract between topography gray scale and tree to close System, finally realize that the identification of image is classified using random walk and the matching of tree strategy, but this method has the phenomenon walked haltingly, So as to introduce noise, the matching of Aerial Images topology is limited, so as to reduce the scope of Aerial Images classification, real life In hardly result in application.

4th, existing many Aerial Images cannot distinguish between image during classification for the much like image of structure Space layout, reduce the accuracy of image classification.

Such as 2006, the author such as Svetlana Lazebnik is in top international conference IEEE Conference on Delivered on Computer Vision and Pattern Recognition (CVPR)《Spatial Pyramid Matching for Recognizing Natural Scene Categories》The one kind proposed in this article is based on sky Between the recognition methods of natural scene classification that matches of pyramid, each image is divided into increasingly thinner grid in this method, Its feature descriptor is calculated by local feature histogram in each grid, but when the space layout very phase of two images Like when, this method can not be distinguished, and so as to reduce the accuracy of image classification, and complexity is high, it is difficult in practice It is applied.

5th, the design of the grader of existing Aerial Images classification is substantially based on SVM classifier, and it is more to lack consideration Information is shared between the grader of individual inter-related task, causes the process that calculates more complicated, the efficiency of classification is low.

Such as 2006, the author such as Svetlana Lazebnik is in top international conference IEEE Conference on Delivered on Computer Vision and Pattern Recognition (CVPR)《Spatial Pyramid Matching for Recognizing Natural Scene Categories》This method is set Aerial Images grader Meter is a kind of non-linear SVM classifier, but due between the multiple inter-related tasks of shortage consideration in assorting process in this method Information of classifier it is shared, it is not high to result in the accuracy of Aerial Images classification, there is certain limitation.

Therefore, up to the present, still without the nicety of grading and efficiency high for a kind of consideration image topology relation occur And can be with the Aerial Images sorting technique of engineer applied.

The content of the invention

In place of the present invention is in order to solve the shortcomings of the prior art, a kind of taking photo by plane based on multitask topology learning is proposed Image system for rapidly identifying and its method for quickly identifying, to which the Fast Classification of Aerial Images can be realized, so as to improve classification Speed and precision.

The present invention is to solve the problems, such as method scheme with the following method：

A kind of the characteristics of Aerial Images system for rapidly identifying based on multitask topology learning of the present invention, includes：Feature construction Module, topological characteristic structure module, topological characteristic projection module, multitask topology learning module, feature selection module and classification Identification module；

The feature construction module obtains some under each image classification respectively according to set N number of image classification Individual tag along sort, and several training Aerial Images for including itself label under each image classification are obtained respectively；

The feature construction module obtains a target Aerial Images to be identified；By target Aerial Images and described Aerial Images are trained to form Aerial Images；

Each Aerial Images in the Aerial Images are divided into using simple linear iteration cluster SLIC dividing methods For multiple zone bits, and form zone bit set；Each area is calculated using probability SVM classifier method to the zone bit set The comentropy of domain member, and whether the comentropy for weighing each zone bit is less than set information entropy threshold；From the zone bit The zone bit less than described information entropy threshold is removed in set, so as to generate the zone bit set after optimization；According to the optimization Positional information of each zone bit in the Aerial Images in zone bit set afterwards, and any two obtained by positional information The connection side that individual adjacent zone bit is generated, build the area distribution of each Aerial Images in the Aerial Images Figure；

Several area adjacency subgraphs are generated by the regional distribution chart of each Aerial Images；By area adjacency The positional information of each zone bit and connection side form the respective topological structure of area adjacency subgraph in figure；By all regions Adjacent subgraph forms the Region adjacency graph of each Aerial Images；

It is adjacent to several regions of each Aerial Images using Sub-Image Feature construction method and core principle component analysis method Connect subgraph to be handled, obtain the characteristic vector of several area adjacency subgraphs of each Aerial Images respectively, each region is adjacent The characteristic vector and the topological structure of respective corresponding area adjacency subgraph for connecing subgraph correspond；By each figure of taking photo by plane The characteristic vector of several area adjacency subgraphs of picture forms the eigenmatrix of respective Aerial Images in the Aerial Images；It is described The eigenmatrix of Aerial Images includes the eigenmatrix of target Aerial Images and the eigenmatrix of training Aerial Images；

The topological characteristic structure module will have identical topological structure in the eigenmatrix of the training Aerial Images Characteristic vector carries out fusion treatment, obtains the topological characteristic matrix of the training Aerial Images；So as to obtain each image The topological characteristic matrix of all training Aerial Images under classification；

The topological characteristic projection module is according to several described tag along sorts and each training Aerial Images Itself label and topological characteristic matrix, obtain the topological characteristic projection of the training Aerial Images under each Aerial Images classification Matrix；

The multitask topology learning module is using multi-task learning framework and principal component analytical method to the topology Projection Character matrix is iterated, so as to obtain the topological encoder matrix of the training Aerial Images；Using F- normal forms to described The row vector of topological encoder matrix is calculated, and the result of calculation of acquisition is ranked up in descending order, and S row vector institute is right before selection The topological structure answered, form the topological dictionary of the Aerial Images；

The feature selection module is according to each topological structure in the topological dictionary and the eigenmatrix of the Aerial Images In topological structure corresponding to each characteristic vector matched, if the match is successful, by the eigenmatrix of the Aerial Images The feature selecting vector of the Aerial Images is formed with successful characteristic vector；The feature selecting vector of the Aerial Images includes The feature selecting vector of the feature selecting vector sum training Aerial Images of target Aerial Images；By the spy of each training Aerial Images Sign selection is vectorial and the training Aerial Images each corresponding image classification, the feature for forming each training Aerial Images are selected Select two tuples；

The Classification and Identification module builds multiple disaggregated models to the topological dictionary using SVM classifier；And according to institute The tuple of feature selecting two for stating training Aerial Images obtains multiple disaggregated models of optimization using probability SVM methods；By described excellent The multiple disaggregated models changed build topological Boosting learning models；By the feature selecting vector input of the target Aerial Images Into the topological Boosting learning models, so as to obtain the classification of the target Aerial Images.

A kind of spy of the method for quickly identifying of the Aerial Images system for rapidly identifying based on multitask topology learning of the present invention Putting is：

It is C={ c to define N number of image classification₁,c₂,…,c_n,…,c_N}；c_nRepresent n-th image classification；Each image point Several tag along sorts are included in class；Remember n-th of classification c_nTag along sort bek_n Represent tag number, 1≤k_n≤K_n；Represent kth_nIndividual tag along sort；K_nRepresent the n-th image classification c_nMiddle contingency table The sum of label；Then the tag along sort collection of N number of image classification C is combined into T={ T₁,T₂,…,T_n,…,T_N}；Remember κ for described point The sum of all tag along sorts in class tag set T, and have

Remember the n-th image classification c_nIn comprising the training Aerial Images of itself label beRepresent n-th image classification c_nIn m_nIndividual training Aerial Images,Represent The n-th image classification c_nIn m_nIndividual training Aerial ImagesItself tag set, 1≤m_n≤M_n；M_nRepresent described N image classification c_nIn it is all training Aerial Images sums；Then the training Aerial Images collection of N number of image classification C is combined into I ={ I₁,I₂,…,I_n,…,I_N}；Remember that target Aerial Images are I₀；

The method for quickly identifying is to carry out as follows：

Step 1, structure the n-th image classification c_nIn m_nIndividual training Aerial ImagesRegional distribution chartSo as to obtain the n-th image classification c_nIn it is all training Aerial Images regional distribution chartsTaken photo by plane figure so as to obtain the training of N number of image classification C Image set closes I area distribution set of graphs RAG={ RAG₁,RAG₂,…,RAG_n,…,RAG_N}；And the target Aerial Images I₀ Regional distribution chart RAG₀；

Step 1.1, using simple linear iteration cluster SLIC dividing methods by the n-th image classify c_nIn m_nIt is individual Train Aerial ImagesIt is divided intoIndividual zone bit, obtain the m_nIndividual training Aerial ImagesZone bit setRepresent the n-th image classification c_nIn m_nIndividual training boat Clap image Individual zone bit；The m_nIndividual training Aerial ImagesIn include described theIt is individual Zone bitPositional information；

Step 1.2, utilize the probability SVM classifier method calculating m_nIndividual zone bit setInIndividual region MemberBelong to n-th image classification c_nProbabilitySo as to obtain describedIndividual zone bitCategory The probability each classified in described image classifies C；

Step 1.3, using formula (1) calculate described theIndividual zone bitComentropySo as to obtain State m_nIndividual zone bit setIn each zone bit comentropy：

Step 1.4, judge described information entropyWhether information entropy threshold δ is less than, if being less than, from the m_nIt is individual Zone bit setThe middle zone bit removed less than described information entropy threshold δ；So as to obtain the m_nIndividual training Aerial ImagesOptimization after zone bit set

Step 1.5, with the m_nIndividual training Aerial ImagesOptimization after zone bit setIn each zone bit Place-centric point as particle, the particle of the adjacent zone bit of any two in each zone bit is carried out using adjacent side Connection, so as to obtain the m_nIndividual training Aerial ImagesRegional distribution chart

Step 2, structure the n-th image classification c_nIn m_nIndividual training Aerial ImagesRegion adjacency graph

From the m_nIndividual training Aerial ImagesRegional distribution chartMiddle exhaustion includes several of connection side Particle, the m is formed by several described particles and corresponding connection side_nIndividual training Aerial ImagesArea adjacency subgraphRepresent theIndividual area adjacency subgraph, by the area adjacency subgraph In each zone bit positional information and connection side form the topological structure of the area adjacency subgraph；Represent the m_nIt is individual Train Aerial ImagesArea adjacency subgraph sum；By the m_nIndividual training Aerial Images's Area adjacency subgraphForm the n-th image classification c_nIn m_nIndividual training is taken photo by plane ImageRegion adjacency graphSo as to obtain the n-th image classification c_nIn all training Aerial Images region it is adjacent Map interlinkingSo as to obtain the training Aerial Images set I's of N number of image classification C Area adjacency set of graphs G={ G₁,G₂,…,G_n,…,G_N}；And the target Aerial Images I₀Regional distribution chart G₀；

Step 3, obtain the n-th image classification c_nIn m_nIndividual training Aerial ImagesEigenmatrixFrom And obtain the n-th image classification c_nTraining Aerial Images eigenmatrix So as to obtain the training Aerial Images set I of N number of image classification C eigenmatrix set X={ X₁,X₂,…,X_n,…, X_N}；And the target Aerial Images I₀Eigenmatrix X₀；

Step 3.1, to the m_nIndividual Aerial ImagesRegion adjacency graphMiddle particle number identical region is adjacent Connect subgraphWithSub-Image Feature matrix is respectively obtained using Sub-Image Feature construction methodWith

Step 3.2, utilize formula (2) the acquisition Sub-Image Feature matrixWithGolub-Werman distances

In formula (2),WithThe Sub-Image Feature matrix is represented respectivelyWithIt is orthogonal Base；||·||₂Represent 2- normal forms；

Step 3.3, utilize formula (3) the calculating area adjacency subgraphWithGaussian kernel

In formula (3),Represent Gauss nuclear parameter；

Step 3.4, repeat step 3.1 arrive step 3.3, so as to obtain the m_nIndividual Region adjacency graphIn all tools There is the Gaussian kernel of the area adjacency subgraph of phase homologous pints；

Step 3.5, the m obtained using core principle component analysis method_nIndividual Aerial ImagesEigenmatrixFor describedIndividual area adjacency subgraphFeature to Amount,D isIndividual characteristic vectorDimension；DescribedIndividual area adjacency subgraphFeature VectorWith describedIndividual area adjacency subgraphTopological structure correspond；

Step 4, obtain the n-th image classification c_nIn m_nIndividual training Aerial ImagesTopological characteristic vector So as to obtain the n-th image classification c_nThe topological characteristic matrix of middle training Aerial ImagesSo as to obtain the training Aerial Images set I of N number of image classification C topology Eigenmatrix set X '={ X '₁,X′₂,…,X′_n,…,X′_N}；

Step 4.1, retain the n-th image classification c_nIn m_nIndividual training Aerial ImagesEigenmatrixIn In topological structure corresponding to all characteristic vectors topological structure differ and particle number no more than upper limit ν topology knot Structure, so as to obtain the n-th image classification c_nIn m_nIndividual training Aerial ImagesTopological structure setSo as to obtain Obtain the n-th image classification c_nTraining Aerial Images topological structure set Handled using unionSo as to obtain the training Aerial Images set I of N number of image classification C topological structure Set tp；

Step 4.2, training Aerial Images set I to N number of image classification C topological structure set tp in it is each Topological structure is numbered, and obtains σ topological structure and is designated as { tp '₁,tp′₂,…,tp′_ο,…,tp′_σ}；tp′_οRepresent that ο are opened up Flutter structure；1≤ο≤σ；

Step 4.3, by the m_nIndividual Aerial ImagesEigenmatrixIn topology corresponding to each characteristic vector Structure and the ο topological structure tp '_οIt is compared, selects the characteristic vector with identical topological structure and summed simultaneously Take average processing；Obtain the m_nIndividual Aerial ImagesIn the ο topological structure tp '_οOn topology fusion vector So as to obtain the m_nIndividual Aerial ImagesIn the σ topological structure { tp '₁,tp′₂,…,tp ′_ο,…,tp′_σOn topology fusion vector

Step 4.3, by the m_nIndividual Aerial ImagesIn the σ topological structure { tp '₁,tp′₂,…,tp′_ο,…, tp′_σOn topology fusion vectorEnd to end fusion treatment is carried out, is obtained The m_nIndividual training Aerial ImagesTopological characteristic vectorAnd haveD σ=d × σ；

Step 5, the topological projection matrix set M=for training Aerial Images set I for obtaining N number of image classification C {M₁,M₂,…,M_n,…,M_N}：

Step 5.1, classified c using formula (4) the acquisition n-th image_nMiddle training Aerial Images I_nTopological transitions matrix

In formula (4),For unit matrix,Represent M_nIndividual element is all 1 One dimensional vector,

Step 5.2, classified c using formula (5) the acquisition n-th image_nIn m_nIndividual training Aerial ImagesKth_n Individual tag elementSo as to obtain the n-th image classification c_nIn m_nIndividual training Aerial ImagesLabel vectorSo as to obtain the n-th image classification c_nInstruction Practice Aerial Images I_nLabel matrixAnd have

Step 5.3, classified c using formula (6) the acquisition n-th image_nMiddle training Aerial Images I_nLabel measurement instruction Matrix F_n,

F_n=Y_n[(Y_n)^TY_n]^-1/2 (6)

Step 5.4, classified c using formula (7) the acquisition n-th image_nMiddle training Aerial Images I_nTopological projection matrix M_n,So as to obtain the training Aerial Images set I of N number of image classification C topological projection matrix set M ={ M₁,M₂,…,M_n,…,M_N}；

Step 6, multitask topology learning feature coding：

Step 6.1, definition iterations are t；Initialize the iterations t=1；

And the topological encoder matrix of the t times iteration is described in when initializing the t times iterationAnd n-th of topological coded sub-matrices when having the t times iterationIn each member Element is obtained by random function；

Step 6.2, the topological encoder matrix for obtaining using formula (8) the t times iteration are W^(t)Code weight matrix D^(t)V row v column elementsCode weight matrix D during so as to obtain the t times iteration^(t)：

In formula (8),The topological encoder matrix for representing the t times iteration is W^(t)V every trades vector, the coding Weight matrix D^(t)For diagonal matrix,

Step 6.3, definition tasking learning number are l；1≤l≤N；And initialize the tasking learning number l=1；

Step 6.4, the weight selection matrix that the t times iteration l subtask learns is obtained using formula (9)In v rows V column elementsSo as to obtain the weight selection matrix of the t times iteration l subtasks study

In formula (9),Represent l-th of topological coded sub-matrices during the t times iterationV row vectors, the power Reselection procedure matrixFor diagonal matrix,

Step 6.5, from the topological characteristic projection matrix set M={ M₁,M₂,…,M_n,…,M_NIn extract and open up for l-th Flutter Projection Character matrix M_l；The code optimization matrix of the t times iteration l subtasks study is obtained using formula (10)

In formula (10), λ presentation code Optimal Parameters；μ presentation code internal coordination parameters；

Step 6.6, note K_lFor l-th of image classification c in N number of image classification C_lThe sum of middle tag along sort；Using Principal component analytical method obtains the code optimization matrix of the t times iteration l subtasks studyMultiple characteristic values and with it is multiple The corresponding characteristic vector of characteristic value, the multiple characteristic value is sorted from small to large, and selects preceding K_lCorresponding to individual characteristic value Characteristic vector as the t times iteration when n-th of topological coded sub-matrices

Step 6.7, judge whether l=N sets up, if so, n times tasking learning is then completed, so as to obtain the t+1 times iteration Topological encoder matrixAnd perform step 6.8；Otherwise l+1 value is assigned to l, weight Step 6.4- steps 6.7 are performed again；

Step 6.8, judge whether t=Γ set up, Γ is iteration total degree；If so, perform step 6.9；Otherwise by t+1 Value be assigned to t, return to step 6.2 sequentially performs；

Step 6.9, using F- normal forms to the t+1 times iteration topology encoder matrix W^(t+1)Row vector calculated, The result of calculation of acquisition is ranked up in descending order, selects the t+1 times iteration topology encoder matrix W^(t+1)S row before middle sequence Vector, the preceding S row vector is obtained in the t+1 times iteration topology encoder matrix W^(t+1)In corresponding line number { h₁, h₂,…,h_s,…,h_S, h_sRepresent that s-th of row vector is compiled in the t+1 times iteration topology in S row vector before the sequence Code matrix W^(t+1)In line number；1≤s≤S；

Step 6.10, from the topological structure { tp '₁,tp′₂,…,tp′_ο,…,tp′_σIn selection theIndividual topology Structure, so as to obtain the topological structure before the sequence in S row vector corresponding to s-th of row vector；So as to obtain the row Topological structure before sequence corresponding to S row vector, and form the topological dictionary Tp={ tp " of the Aerial Images₁,tp″₂,…, tp″_s′,…,tp″_S′}；tp″_s′Represent the individual topological structures of s ' in the topological dictionary Tp；1≤s′≤S′；

Step 7, obtain the n-th image classification c_nIn m_nIndividual training Aerial ImagesFeature selecting vector And the target Aerial Images I₀Feature selecting vector x "₀；So as to obtain the n-th image classification c_nIn m_nIndividual instruction Practice Aerial ImagesThe tuple of feature selecting twoSo as to obtain the n-th image classification c_nIt is middle to train figure of taking photo by plane As I_nThe tuple of feature selecting twoSo as to obtain State the N number of image classification C training Aerial Images set I tuple-set of feature selecting twoNote Represent the training Aerial Images set I of the N number of image classification C tuple-set X of feature selecting two_FSE-th The tuple of feature selecting two, 1≤e≤E；

Step 7.1, definition selection number are ρ, and initialize ρ=1；1≤ρ≤S′；

Define the described the ρ times m_nIndividual training Aerial ImagesFeature Choice Matrix beAnd initialize institute State the ρ times m_nIndividual training Aerial ImagesFeature Choice Matrix beFor sky；Define the ρ times m_nIt is individual Train Aerial ImagesFeature extraction matrix beAnd initialize

Step 7.2, choose the ρ topological structure tp " from topological dictionary Tp_ρ；

Step 7.3, judge the ρ topological structure tp "_ρIn particle number whether be more than the m_nIndividual training boat Clap imageRegional distribution chartParticle number, if being more than, ρ+1 value is assigned to perform step 7.2 after ρ；It is no Then, step 7.4 is performed；

Step 7.4, by the ρ topological structure tp "_ρWith the ρ times m_nIndividual training Aerial ImagesFeature MatrixIn topological structure corresponding to each characteristic vector matched；If the match is successful, from the ρ times m_n Individual training Aerial ImagesFeature extraction matrixMiddle removal and the ρ topological structure tp "_ρWhat the match is successful opens up The characteristic vector corresponding to structure is flutterred, and is put into the ρ times m_nIndividual training Aerial ImagesFeature Choice Matrix beIn；

Step 7.5, ρ+1 value is assigned to ρ, judges whether ρ=S ' sets up, if so, using maximum pond strategy to described M_nIndividual training Aerial ImagesFeature Choice MatrixHandled, obtain the m_nIndividual training Aerial Images Feature selecting vectorOtherwise, step 7.2- steps 7.5 are repeated；

Step 8, the topological Boosting learning models of structure：

Step 8.1, utilize the individual topological structures of s ' in the one-against-rest structure topological dictionary Tp of SVM classifier tp″_s′Disaggregated model ψ_s′；So as to obtain the individual disaggregated model { ψ of S ' of the topological dictionary Tp₁,ψ₂,...,ψ_s′,...,ψ_S′}；

Step 8.2, the iterations of definition Boosting learning models are τ, and initialize iterations τ=1；

DefinitionRepresent the training Aerial Images set I of N number of image classification C during the τ times iteration feature choosing Select two tuple-set X_FSThe tuple of e-th of feature selecting twoClass weight；Then institute during the τ times iteration State the N number of image classification C training Aerial Images set I tuple-set X of feature selecting two_FSClass weight beAnd initialize

Step 8.3, using formula (11) obtain the τ times iteration optimization disaggregated model ψ^(τ)；

In formula (11),Represent using probability SVM methods to the disaggregated model ψ^(τ)Carry out processing acquisition The tuple of e-th of feature selecting twoFeature selecting vectorBelong to image classificationProbability；

Step 8.4, utilize identification mistake err during formula (12) the τ times iteration of calculating_τ：

Step 8.5, utilize identification weight α during formula (13) the τ times iteration of calculating_τ：

Step 8.6, using formula (14) calculate the τ times iteration when described in e-th of class weightCorresponding classification transition power WeightSo as to obtain E classification transition weight

Step 8.7, to the E classification transition weightIt is normalized, obtains Class weight during+1 iteration of τ

Step 8.8, τ+1 value is assigned to τ, repeats step 8.3- steps 8.8；Until τ is equal to iterations A Only, weight set { α is identified so as to obtain₁,α₂,…,α_a,…,α_AAnd the A iteration optimization disaggregated model { ψ⁽¹⁾, ψ⁽²⁾,…,ψ^(a),…,ψ^(A)}；α_aRepresent a-th of identification weight in the identification weight set；ψ^(a)Represent and described a-th knowledge The disaggregated model of the corresponding optimization of other weight；

Step 8.9, utilize the topological Boosting learning models acquisition target Aerial Images I shown in formula (15)₀'s Image classification Δ；So as to realize the quick identification of target Aerial Images：

In formula (15), p_a(x″₀→ Δ) represent the disaggregated model ψ to described a-th optimization using probability SVM methods^(a)Enter The target Aerial Images I that row processing is obtained₀Feature selecting vector x "₀Belong to the probability of image classification Δ.

Compared with the prior art, the present invention has the beneficial effect that：

1st, the present invention considers the topological relation of Aerial Images, employs topological structure Classification and Identification, multi-task learning frame The thought of frame and Boosting learning models：First, SLIC dividing methods and information entropy optimization are clustered using simple linear iteration Method, the regional distribution chart and Region adjacency graph of each Aerial Images are built, remains the topology information in Aerial Images；Then, The eigenmatrix of Aerial Images is obtained using Sub-Image Feature construction method, and then obtains all instructions under each image classification Practice the topological characteristic matrix and topological characteristic projection matrix of Aerial Images；Then, realized and trained using multi-task learning framework The topology coding square of Aerial Images, obtains topological dictionary；Feature selecting is carried out according to topological dictionary, obtains target Aerial Images Feature selecting vector sum trains the tuple of feature selecting two of Aerial Images；Finally, the feature for training Aerial Images according to selection Two tuples are selected to build topological Boosting learning models；The feature selecting vector of the target Aerial Images is input to described In topological Boosting learning models, so as to obtain the classification of the target Aerial Images；Classify with conventional Aerial Images Method is compared, and the present invention has been effectively kept the topology information in Aerial Images；More are realized using multi-task learning framework Information sharing between business；Improve the accuracy of classification；Topological Boosting learning models are once built up, and are once calculated The Classification and Identification of target Aerial Images quickly is obtained, the efficiency of Aerial Images classification is improved, can be obtained in real life It is widely applied.

2nd, the present invention screens according to the principle of comentropy to the zone bit of Aerial Images segmentation gained through having gone, and removes region First noise, the zone bit after being optimized, so as to improve the efficiency of final classification.

3rd, the present invention replaces conventional grid chart using the regional distribution chart for constructing Aerial Images so that each in Aerial Images The topological structure of zone bit can be effectively maintained, the shortcomings that overcoming conventional grid chart, improve Aerial Images topology letter The accuracy of breath.

4th, the eigenmatrix of subgraph is obtained using the method for Sub-Image Feature structure in the present invention, it is more representative, contain The information of abundant Aerial Images, can include the visual signatures such as color, texture, and the geometric space that and can includes topology distribution closes System, it is possible to increase the precision of final classification.

5th, the present invention extends Aerial Images topological characteristic and is identified using based on the characteristic matching between subgraph and subgraph Scope, overcome in existing method because using random walk and set core method, caused by Aerial Images topological relation exist Whether can not accurately be described, the problem of so as to cause nicety of grading low if matching.

6 compare with the Aerial Images sorting technique studied in the past, the present invention so that obtain its topological characteristic vector, fusion Feature description of each Sub-Image Feature matrix of same Aerial Images in different topology structure, improves topological features Precision.

7th, topological characteristic is realized in the classificatory projective measurement of different images according to topological characteristic vector in the present invention, and The computational methods of the topological characteristic projection matrix proposed in the present invention are simple, and the mathematical formulae utilized is simply easily asked, so as to reduce The complexity calculated.

8th, the present invention realizes the coding of final topological characteristic according to topological characteristic projection matrix, solves existing method Aerial Images cannot distinguish between its space layout during classification for the much like image of structure, so as to reduce image point The problem of accuracy of class.

9 compare with the Aerial Images sorting technique studied in the past, and the present invention utilizes multi-task learning framework, to multiple phases The shared of information is realized between the image classification task of pass, realizes that topological characteristic encodes so that there is identity in topological characteristic, gram Taken in previous methods due to a lack of considering the shared of information between inter-related task, caused by the efficiency classified it is low the problem of.

10th, the present invention obtains topological dictionary by using based on multitask topology learning feature coding, utilizes topological dictionary To extract the Sub-Image Feature with distinctive, topological dictionary can assisting sifting go out the subgraph with identification in Aerial Images Characteristic vector, effective classification information is extracted for disaggregated model of establishing below.

11st, the topological Boosting learning models built in the present invention, multiple Weak Classifiers are integrated into point learnt by force Class device model, there is the characteristics of high complexity of accuracy is low, improve the precision and efficiency of learning model.

12nd, the topological Boosting learning models built in the present invention only need every time once being built up according to training image Arbitrary target Aerial Images are inputted, is once calculated using the calculation formula of learning model, can be achieved with target Aerial Images Quick identification, amount of calculation are small and quick.

Brief description of the drawings

Fig. 1 is present system function structure chart；

Fig. 2 is the inventive method flow chart；

Fig. 3 a are example images figures of being taken photo by plane in the inventive method；

Fig. 3 b are the regional distribution chart exemplary plots of Aerial Images in the inventive method；

Fig. 3 c are multiple area adjacency subgraph exemplary plots of Aerial Images in the inventive method；

Fig. 4 a are the topological structure exemplary plots of Aerial Images in the inventive method；

Fig. 4 b are the area adjacency subgraph exemplary plots 1 for having in the inventive method identical topological structure；

Fig. 4 c are the area adjacency subgraph exemplary plots 2 for having in the inventive method identical topological structure；

Fig. 4 d are the area adjacency subgraph exemplary plots 3 for having in the inventive method identical topological structure.

Embodiment

In the present embodiment, a kind of Aerial Images system for rapidly identifying based on multitask topology learning and method are to be used for mesh Aerial Images class categories identification mission is marked, as shown in figure 1, a kind of Aerial Images based on multitask topology learning quickly identify System, including：Feature construction module, topological characteristic structure module, topological characteristic projection module, multitask topology learning module, Feature selection module and Classification and Identification module；

Feature construction module obtains several points under each image classification respectively according to set N number of image classification Class label, and several training Aerial Images for including itself label under each image classification are obtained respectively；

Feature construction module obtains a target Aerial Images to be identified；By target Aerial Images and training Aerial Images Form Aerial Images；Training Aerial Images are used to model；

Each Aerial Images in Aerial Images are partitioned into using simple linear iteration cluster SLIC dividing methods more Individual zone bit, and form zone bit set；The letter of each zone bit is calculated zone bit set using probability SVM classifier method Entropy is ceased, and whether the comentropy for weighing each zone bit is less than set information entropy threshold；Removed from zone bit set small In the zone bit of information entropy threshold, so as to generate the zone bit set after optimization；According to the zone bit set Zhong Ge areas after optimization Positional information of the domain member in Aerial Images, and one generated by the adjacent zone bit of any two of positional information acquisition Bar connects side, builds the regional distribution chart of each Aerial Images in Aerial Images；

Several area adjacency subgraphs are generated by the regional distribution chart of each Aerial Images；By area adjacency subgraph Zhong Ge areas The positional information of domain member and connection side form the respective topological structure of area adjacency subgraph；It is made up of all area adjacency subgraphs The Region adjacency graph of each Aerial Images；

It is sub to several area adjacencies of each Aerial Images using Sub-Image Feature construction method and core principle component analysis method Figure is handled, and obtains the characteristic vector of several area adjacency subgraphs of each Aerial Images, each area adjacency respectively The characteristic vector of figure and the topological structure of respective corresponding area adjacency subgraph correspond；By some of each Aerial Images The characteristic vector of individual area adjacency subgraph forms the eigenmatrix of respective Aerial Images in Aerial Images；The feature square of Aerial Images Battle array includes the eigenmatrix of target Aerial Images and the eigenmatrix of training Aerial Images；

Topological characteristic builds module will have identical topological structure characteristic vector in the eigenmatrix for training Aerial Images Fusion treatment is carried out, obtains the topological characteristic matrix of training Aerial Images；So as to obtain all training under each image classification The topological characteristic matrix of Aerial Images；

Topological characteristic projection module is trained itself label of Aerial Images according to several tag along sorts and each and opened up Eigenmatrix is flutterred, obtains the topological characteristic projection matrix of the training Aerial Images under each Aerial Images classification；

Multitask topology learning module is projected using multi-task learning framework and principal component analytical method to topological characteristic Matrix is iterated, so as to obtain the topological encoder matrix of training Aerial Images；Row using F- normal forms to topological encoder matrix Vector is calculated, and the result of calculation of acquisition is ranked up in descending order, the topological structure before selection corresponding to S row vector, structure Into the topological dictionary of Aerial Images；

Feature selection module is according to each characteristic vector in the eigenmatrix of each topological structure and Aerial Images in topological dictionary Corresponding topological structure is matched, if the match is successful, by the characteristic vector that the match is successful in the eigenmatrix of Aerial Images Form the feature selecting vector of Aerial Images；Aerial Images feature selecting vector include target Aerial Images feature selecting to The feature selecting vector of amount and training Aerial Images；By the feature selecting vector and training Aerial Images of each training Aerial Images Each corresponding image classification, form the tuple of feature selecting two of each training Aerial Images；

Classification and Identification module builds multiple disaggregated models to topological dictionary using SVM classifier；And taken photo by plane figure according to training The tuple of feature selecting two of picture obtains multiple disaggregated models of optimization using probability SVM methods；By the multiple disaggregated models optimized Build topological Boosting learning models；The feature selecting vector of target Aerial Images is input to topological Boosting study mould In type, so as to obtain the classification of target Aerial Images.

In the present embodiment, it is C={ c to define N number of image classification₁,c₂,…,c_n,…,c_N}；c_nRepresent n-th image classification； Several tag along sorts are included in each image classification；N-th of classification c of note_nTag along sort bek_nRepresent tag number, 1≤k_n≤K_n；Represent kth_nIndividual tag along sort；K_nRepresent n-th Individual image classification c_nThe sum of middle tag along sort, K_n∈ [10,50], in specific implementation, K_n=10；Then N number of image classification C point Class tag set is T={ T₁,T₂,…,T_n,…,T_N}；Remember that κ is the sum of all tag along sorts in tag along sort set T, and have

Remember n-th image classification c_nIn comprising the training Aerial Images of itself label be Represent n-th image classification c_nIn m_nIndividual training Aerial Images,Represent N-th image classification c_nIn m_nIndividual training Aerial ImagesItself tag set, 1≤m_n≤M_n；M_nRepresent n-th image Classify c_nIn it is all training Aerial Images sums, for example, M_n∈[5000,10000]；Then N number of image classification C training is taken photo by plane Image collection is I={ I₁,I₂,…,I_n,…,I_N, i.e., as shown in Fig. 2 havingRemember that target Aerial Images are I₀；

As shown in Fig. 2 a kind of recognition methods of the Aerial Images system for rapidly identifying based on multitask topology learning be by Following steps are carried out：

Step 1, structure n-th image classification c_nIn m_nIndividual training Aerial ImagesRegional distribution chart So as to obtain n-th image classification c_nIn it is all training Aerial Images regional distribution chartsSo as to obtain N number of image classification C training Aerial Images collection Close I area distribution set of graphs RAG={ RAG₁,RAG₂,…,RAG_n,…,RAG_N, i.e., as shown in Fig. 2 havingAnd target Aerial Images I₀Regional distribution chart RAG₀；

Step 1.1, cluster SLIC dividing methods using simple linear iteration n-th image is classified c_nIn m_nIndividual training Aerial ImagesIt is divided intoIndividual zone bit, obtain m_nIndividual training Aerial ImagesZone bit setRepresent n-th image classification c_nIn m_nIt is individual to train figure of taking photo by plane Picture Individual zone bit；M_nIndividual training Aerial ImagesIn include theIndividual zone bit's Positional information；Specific simple linear iteration cluster SLIC dividing methods refer to document《SLIC Superpixels Compared to State-of-the-art Superpixel Methods》In introduction；Simple linear iteration clusters SLIC Dividing method is a kind of super-pixel segmentation based on clustering algorithm, and by LAB spaces and x, y pixel coordinate table, totally 5 dimension spaces come Calculate；The method not only can with Segmentation of Color Images, can also compatible segmentation gray-scale map, but also can taking human as setting need The quantity in the region of segmentation；The advantages of simple linear iteration cluster SLIC dividing methods, is：On the one hand, efficiency is improved, and The subregion of division has been effectively maintained the information of source images, and the degree of accuracy is high；On the other hand, amount of calculation is small, and time cost is with super Number of pixels linear increase；In this implementation, by set different simple linear iteration cluster SLIC serious forgiveness parameter 0.1, 0.2,0.3,0.4,0.5 }, repeated segmentation is carried out to Aerial Images；

Step 1.2, utilize probability SVM classifier method calculating m_nIndividual zone bit setInIndividual zone bitBelong to n-th image classification c_nProbabilitySo as to obtainIndividual zone bitBelong to image point The probability each classified in class C；Specific probability SVM classifier method refers to document《Probabilistic Outputs for Support Vector Machines and Comparisons to Regularized Likelihood Method》 In introduction；Probability SVM classifier method is the support vector method of posterior probability, is the extension of standard SVM classifier and pushes away Extensively；All it is that Nonlinear Mapping is first selected according to priori in advance, input vector is mapped to a high dimensional feature sky using it Between, construct optimal hyperlane in this space；In SVM, training sample always has clear and definite classification information, and for one A little uncertain problems are simultaneously incorrect；Probability SVM is then inspired by Bayes decision rule, using the posterior probability of sample come table Show this uncertainty, Bayes decision rule is combined with SVM, establishes posterior probability support vector machine method, to solve Certain problem by no means；The present invention can improve to obtain nicety of grading using the method；

Step 1.3, using formula (1) calculate theIndividual zone bitComentropySo as to obtain m_nIt is individual Zone bit setIn each zone bit comentropy：

Step 1.4, judge comentropyWhether information entropy threshold δ is less than, if being less than, from m_nIndividual region metaset CloseThe middle zone bit removed less than information entropy threshold δ；So as to obtain m_nIndividual training Aerial ImagesOptimization after area Domain member setδ∈[0.1,0.5]；δ can be adjusted using step-length as 0.05；

Step 1.5, with m_nIndividual training Aerial ImagesOptimization after zone bit setIn each zone bit position Central point is put as particle, the particle of the adjacent zone bit of any two in each zone bit is attached using adjacent side, from And obtain m_nIndividual training Aerial ImagesRegional distribution chart

For example, Fig. 3 a are an Aerial Images example；After shown in Fig. 3 b being then the optimization in Fig. 3 a corresponding to Aerial Images The regional distribution chart that is obtained of zone bit set；

Step 2, structure n-th image classification c_nIn m_nIndividual training Aerial ImagesRegion adjacency graph

From m_nIndividual training Aerial ImagesRegional distribution chartMiddle exhaustive several particles for including connection side, M is formed by several particles and corresponding connection side_nIndividual training Aerial ImagesArea adjacency subgraphRepresent theIndividual area adjacency subgraph, by each in area adjacency subgraph The positional information of zone bit and connection side form the topological structure of area adjacency subgraph；Represent m_nIndividual training Aerial ImagesArea adjacency subgraph sum；By m_nIndividual training Aerial ImagesArea adjacency subgraphForm n-th image classification c_nIn m_nIndividual training Aerial ImagesRegion it is adjacent Map interlinkingAs shown in Figure 3 c, the multiple area adjacency subgraphs obtained by regional distribution chart in Fig. 3 b；It is only portion shown in Fig. 3 c Subregion abuts subgraph example；So as to obtain n-th image classification c_nIn it is all training Aerial Images Region adjacency graphsSo as to obtain N number of image classification C training Aerial Images set I area adjacency Set of graphs G={ G₁,G₂,…,G_n,…,G_N, i.e., as shown in Fig. 2 havingAnd target Aerial Images I₀ Regional distribution chart G₀；

Step 3, obtain n-th image classification c_nIn m_nIndividual training Aerial ImagesEigenmatrixSo as to obtain N-th image classification c_nTraining Aerial Images eigenmatrixSo as to obtain N Individual image classification C training Aerial Images set I eigenmatrix set X={ X₁,X₂,…,X_n,…,X_N, i.e., such as Fig. 2 institutes Show haveAnd target Aerial Images I₀Eigenmatrix X₀；

Step 3.1, to m_nIndividual Aerial ImagesRegion adjacency graphMiddle particle number identical area adjacency FigureWithSub-Image Feature matrix is respectively obtained using Sub-Image Feature construction methodWith Specific Sub-Image Feature construction method bibliography《Probabilistic Graphlet Transfer for Photo Cropping》；Sub-Image Feature construction method utilizes classical color of image feature extracting method and textural characteristics Extracting method obtains the color and texture visual signature of each area adjacency subgraph, and extracts its geometry feature, to each Area adjacency subgraph, these three features are sequentially connected constitutive characteristic matrix；That is Sub-Image Feature matrix；

Step 3.2, utilize formula (2) acquisition Sub-Image Feature matrixWithGolub-Werman distances

In formula (2),WithSub-Image Feature matrix is represented respectivelyWithOrthogonal basis；| |·||₂Represent 2- normal forms；

Step 3.3, utilize formula (3) zoning adjoining subgraphWithGaussian kernel

In formula (3),Gauss nuclear parameter is represented,

Step 3.4, repeat step 3.1 arrive step 3.3, so as to obtain m_nIndividual Region adjacency graphIn all there is phase The Gaussian kernel of the area adjacency subgraph of homologous pints；

Step 3.5, using core principle component analysis method (Kernel PCA methods) obtain m_nIndividual Aerial ImagesFeature MatrixForIndividual area adjacency subgraphFeature to Amount,D isIndividual characteristic vectorDimension；TheIndividual area adjacency subgraphCharacteristic vectorWithIndividual area adjacency subgraphTopological structure correspond；Specific core principle component analysis method (Kernel PCA methods) refer to document《Kernel PCA for Feature Extraction and De-Noising in Nonlinear Regression》；A kind of core principle component analysis method (Kernel PCA methods) nonlinear pca method (PCA methods), its main thought are that input vector is mapped into a higher-dimension by the Nonlinear Mapping that certain is selected in advance Among linear character space, pivot composition then is calculated using PCA in High-dimensional Linear feature space, it has The advantages that nonlinear characteristic of data can effectively be caught, do not required the distribution situations of the data in luv space；It is based on This advantage, the present invention can utilize the method by the very high variables transformations of script correlation into being mutually independent or incoherent Variable, the characteristic vector mutual independence for obtaining each Region adjacency graph are stronger；

Step 4, obtain n-th image classification c_nIn m_nIndividual training Aerial ImagesTopological characteristic vectorSo as to Obtain n-th image classification c_nThe topological characteristic matrix of middle training Aerial ImagesFrom And obtain N number of image classification C training Aerial Images set I topological characteristic set of matrices X '={ X₁′,X′₂,…,X′_n,…, X′_N, i.e., as shown in Fig. 2 having

Step 4.1, retain n-th image classification c_nIn m_nIndividual training Aerial ImagesEigenmatrixIn own In topological structure corresponding to characteristic vector topological structure differ and particle number be no more than upper limit ν topological structure, from And obtain n-th image classification c_nIn m_nIndividual training Aerial ImagesTopological structure setSo as to obtain n-th of figure As classification c_nTraining Aerial Images topological structure setHandled using unionI.e. to topological structure set { tp₁,tp₂,…,tp_n,…,tp_NIn all topological structure do union operation： Retain different topological structures, one is only retained to identical topological structure；The different Aerial Images productions as shown in Fig. 4 b, 4c and 4d Raw multiple area adjacency subgraphs have identical topological structure, and identical topological structure is as shown in institute Fig. 4 a；To such topology When structure does union processing, it is only necessary to retain a topological structure, it is not necessary to repeat to retain；So as to obtain N number of image point Class C training Aerial Images set I topological structure set tp；

Step 4.2, training Aerial Images set I to N number of image classification C topological structure set tp in each topology Structure is numbered, and obtains σ topological structure and is designated as { tp '₁,tp′₂,…,tp′_ο,…,tp′_σ}；tp′_οRepresent the ο topology knot Structure；1≤ο≤σ；In the present embodiment, as ν=7, σ values are 160 or so；

Step 4.3, by m_nIndividual Aerial ImagesEigenmatrixIn topological structure corresponding to each characteristic vector With the ο topological structure tp '_οIt is compared, selects the characteristic vector with identical topological structure and summed and take average place Reason；Obtain m_nIndividual Aerial ImagesIn the ο topological structure tp '_οOn topology fusion vector So as to obtain m_nIndividual Aerial ImagesIn σ topological structure { tp₁′,tp₂′,…,tp′_ο,…,tp′_σOn topology merge to Amount

Step 4.3, by m_nIndividual Aerial ImagesIn σ topological structure { tp '₁,tp′₂,…,tp′_ο,…,tp′_σOn Topology fusion vectorEnd to end fusion treatment is carried out, obtains m_nIndividual instruction Practice Aerial ImagesTopological characteristic vectorAnd haveD σ=d × σ；

Step 5, the topological projection matrix set M={ M for training Aerial Images set I for obtaining N number of image classification C₁, M₂,…,M_n,…,M_N}：

Step 5.1, n-th image classification c is obtained using formula (4)_nMiddle training Aerial Images I_nTopological transitions matrix

In formula (4),For unit matrix,Represent M_nIndividual element is all 1 One dimensional vector,

Step 5.2, n-th image classification c is obtained using formula (5)_nIn m_nIndividual training Aerial ImagesKth_nIndividual mark Sign elementSo as to obtain n-th image classification c_nIn m_nIndividual training Aerial ImagesLabel vectorSo as to obtain n-th image classification c_nTraining boat Clap image I_nLabel matrixAnd have

Step 5.3, n-th image classification c is obtained using formula (6)_nMiddle training Aerial Images I_nLabel measurement oriental matrix F_n,

F_n=Y_n[(Y_n)^TY_n]^-1/2 (6)

Step 5.4, n-th image classification c is obtained using formula (7)_nMiddle training Aerial Images I_nTopological projection matrix M_n,So as to obtain N number of image classification C training Aerial Images set I topological projection matrix set M={ M₁, M₂,…,M_n,…,M_N, i.e., as shown in Fig. 2 having

Step 6, multitask topology learning feature coding：

Step 6.1, definition iterations are t；Initialize iterations t=1；

And the topological encoder matrix of the t times iteration is when initializing the t times iterationAnd n-th of topological coded sub-matrices when having the t times iterationIn each member Element is obtained by random function；

Step 6.2, using formula (8) obtain the t times iteration topological encoder matrix be W^(t)Code weight matrix D^(t)'s V row v column elementsCode weight matrix D during so as to obtain the t times iteration^(t)：

In formula (8),The topological encoder matrix for representing the t times iteration is W^(t)V every trades vector, code weight matrix D^(t)For diagonal matrix,

Step 6.3, definition tasking learning number are l；1≤l≤N；And initialization task study number l=1；

Step 6.4, the weight selection matrix that the t times iteration l subtask learns is obtained using formula (9)In v rows V column elementsSo as to obtain the weight selection matrix of the t times iteration l subtasks study

In formula (9),Represent l-th of topological coded sub-matrices during the t times iterationV row vectors, weight choosing Select matrixFor diagonal matrix,

Step 6.5, from topological characteristic projection matrix set M={ M₁,M₂,…,M_n,…,M_NIn to extract l-th of topology special Levy projection matrix M_l；The code optimization matrix of the t times iteration l subtasks study is obtained using formula (10)

In formula (10), λ presentation code Optimal Parameters, λ ∈ [0,0.1]；μ presentation code internal coordination parameters, μ ∈ [0, 0.5]；For example, λ=0.01, μ=0.3；

Step 6.6, note K_lFor l-th of image classification c in N number of image classification C_lThe sum of middle tag along sort；Using it is main into Analysis method (PCA methods) is divided to obtain the code optimization matrix of the t times iteration l subtasks studyMultiple characteristic values and The characteristic vector corresponding with multiple characteristic values, multiple characteristic values are sorted from small to large, and select preceding K_lIndividual characteristic value institute is right N-th of topological coded sub-matrices when the characteristic vector answered is as the t times iterationPCA (PCA methods) is one Data analysing method of the kind based on second-order statisticses, one group of given correlated variables is changed into another group of not phase by it by linear transformation The variable of pass, the order that these new variables successively decrease successively according to variance are arranged, and the population variance of variable is kept in mathematic(al) manipulation It is constant；The main purpose of PCA is the most of variable gone to explain in original data with less variable；We Selected using the method than code optimization matrix in faceMiddle characteristic vector number is few but can explain most of code optimization matrixSeveral new vectors of middle characteristic vector；

Step 6.7, judge whether l=N sets up, if so, n times tasking learning is then completed, so as to obtain the t+1 times iteration Topological encoder matrixAnd perform step 6.8；Otherwise l+1 value is assigned to l, weight Step 6.4- steps 6.7 are performed again；

Step 6.8, judge whether t=Γ set up, Γ is iteration total degree, for example, Γ=1000；If so, perform step Rapid 6.9；Otherwise t+1 value is assigned to t, return to step 6.2 sequentially performs；

Step 6.9, using F- normal forms to the t+1 times iteration topology encoder matrix W^(t+1)Row vector calculated, obtain Result of calculation be ranked up in descending order, select the t+1 times iteration topology encoder matrix W^(t+1)S row vector, is obtained before middle sequence S row vector is in the t+1 times iteration topology encoder matrix W before obtaining^(t+1)In corresponding line number { h₁,h₂,…,h_s,…,h_S, h_sRepresent before sequence that s-th of row vector is in the t+1 times iteration topology encoder matrix W in S row vector^(t+1)In line number；1≤ s≤S；The t+1 times iteration topology encoder matrix W^(t+1)In line number started counting up with 1；

Step 6.10, from topological structure { tp '₁,tp′₂,…,tp′_ο,…,tp′_σIn selection theIndividual topology knot Structure, so as to the topological structure corresponding to s-th of row vector in S row vector before being sorted；So as to before being sorted S row to The corresponding topological structure of amount, due to there may be identical topological structure in the topological structure corresponding to preceding S row vector, this Only retain different topological structures in implementation, and form the topological dictionary Tp={ tp " of Aerial Images₁,tp″₂,…,tp″_s′,…, tp″_S′}；tp″_s′Represent the individual topological structures of s ' in topological dictionary Tp；1≤s′≤S′；

Step 7, obtain n-th image classification c_nIn m_nIndividual training Aerial ImagesFeature selecting vectorWith And target Aerial Images I₀Feature selecting vector x "₀；So as to obtain n-th image classification c_nIn m_nIndividual training Aerial ImagesThe tuple of feature selecting twoSo as to obtain n-th image classification c_nMiddle training Aerial Images I_nFeature selecting Two tuplesSo as to obtain N number of image classification C instruction Practice the Aerial Images set I tuple-set of feature selecting two

NoteI.e. as shown in Fig. 2 havingRepresent N number of image classification C training Aerial Images set I feature selecting binary Group set X_FSThe tuple of e-th of feature selecting two, 1≤e≤E；

Step 7.1, definition selection number are ρ, and initialize ρ=1；1≤ρ≤S′；

Define the ρ times m_nIndividual training Aerial ImagesFeature Choice Matrix beAnd initialize the ρ times m_n Individual training Aerial ImagesFeature Choice Matrix beFor sky；Define the ρ times m_nIndividual training Aerial ImagesSpy Sign extracts matrixAnd initialize

Step 7.2, choose the ρ topological structure tp " from topological dictionary Tp_ρ；

Step 7.3, judge the ρ topological structure tp '_ρ' in particle number whether be more than m_nIndividual training Aerial ImagesRegional distribution chartParticle number, if being more than, ρ+1 value is assigned to perform step 7.2 after ρ；Otherwise, hold Row step 7.4；

Step 7.4, by the ρ topological structure tp "_ρWith the ρ times m_nIndividual training Aerial ImagesEigenmatrix In topological structure corresponding to each characteristic vector matched；If the match is successful, from the ρ times m_nIndividual training Aerial ImagesFeature extraction matrixMiddle removal and the ρ topological structure tp "_ρThe feature corresponding to topological structure that the match is successful Vector, and it is put into the ρ times m_nIndividual training Aerial ImagesFeature Choice Matrix beIn；

Step 7.5, ρ+1 value is assigned to ρ, judges whether ρ=S ' sets up, if so, using maximum pond strategy to m_n Individual training Aerial ImagesFeature Choice MatrixHandled, obtain m_nIndividual training Aerial ImagesFeature Selection vectorOtherwise, step 7.2- steps 7.5 are repeated；Using maximum pond strategy i.e. by often row retains maximum in matrix It is worth element and forms a new vector；

Step 8, the topological Boosting learning models of structure：

Boosting methods are a kind of methods for improving the weak typing algorithm degree of accuracy, by constructing an anticipation function Series, they are then combined into an anticipation function in some way；The discrimination of each single grader is not necessarily It is very high, but the result after their joints has very high discrimination, and this implementation uses the method, it is possible to increase the identification of disaggregated model Rate；

Step 8.1, the is built in topological dictionary Tp using the one-against-rest (One-Against-All) of SVM classifier The individual topological structure tp " of s '_s′Disaggregated model ψ_s′；So as to obtain the individual disaggregated model { ψ of topological dictionary Tp S '₁,ψ₂,..., ψ_s′,...,ψ_S′}；The one-to-many methods of one-against-rest (one-against-all) of SVM classifier, successively some during training The sample of classification is classified as one kind, and other remaining samples are classified as another kind of, and the sample of such k classification has just constructed k SVM.Unknown sample is categorized as to that class with maximum classification function value during classification；Realized and taken photo by plane using the method in the present invention The classification of image, it is possible to achieve the multi-class identification of Aerial Images classification；

Step 8.2, the iterations of definition Boosting learning models are τ, and initialize iterations τ=1；

DefinitionRepresent the training Aerial Images set I of the N number of image classification C during the τ times iteration tuple of feature selecting two Set X_FSThe tuple of e-th of feature selecting twoClass weight；Then during the τ times iteration N number of image classification C training The Aerial Images set I tuple-set X of feature selecting two_FSClass weight beAnd initialize

Step 8.3, using formula (11) obtain the τ times iteration optimization disaggregated model ψ^(τ)；

In formula (11),Represent using probability SVM methods to disaggregated model ψ^(τ)Carry out handling the of acquisition The e tuple of feature selecting twoFeature selecting vectorBelong to image classificationProbability；With general in step 1.2 Rate SVM classifier explains that equally specific probability SVM classifier method refers to document《Probabilistic Outputs for Support Vector Machines and Comparisons to Regularized Likelihood Method》 In introduction, the present invention can improve to obtain nicety of grading using the method；

Step 8.4, utilize identification mistake err during formula (12) the τ times iteration of calculating_τ：

Step 8.5, utilize identification weight α during formula (13) the τ times iteration of calculating_τ：

Step 8.6, utilize e-th of class weight during formula (14) the τ times iteration of calculatingCorresponding classification transition weight；So as to obtain E classification transition weight

Step 8.7, to E classification transition weightIt is normalized, acquisition τ+ Class weight during 1 iteration

Step 8.8, τ+1 value is assigned to τ, repeats step 8.3- steps 8.8；Until τ is equal to iterations A Only, for example, A=1000, weight set { α is identified so as to obtain₁,α₂,…,α_a,…,α_AAnd A iteration optimization classification mould Type { ψ⁽¹⁾,ψ⁽²⁾,…,ψ^(a),…,ψ^(A)}；α_aRepresent a-th of identification weight in identification weight set；ψ^(a)Represent and a-th of knowledge The disaggregated model of the corresponding optimization of other weight；

Step 8.9, utilize the topological Boosting learning models acquisition target Aerial Images I shown in formula (15)₀Image Classification Δ；So as to realize the quick identification of target Aerial Images：

In formula (15), p_a(x″₀→ Δ) represent the disaggregated model ψ to a-th of optimization using probability SVM methods^(a)Located The obtained target Aerial Images I of reason₀Feature selecting vector x "₀Belong to the probability of image classification Δ；With probability in step 1.2 SVM classifier explains that equally specific probability SVM classifier method sees bibliography《Probabilistic Outputs for Support VectorMachines and Comparisons to Regularized LikelihoodMethod》In Jie Continue, the present invention can improve to obtain nicety of grading using the method.

More than, it is only a kind of present invention preferably embodiment, other researchers, completely can be at other according to above In association area, such as image-region target, image scene, picture position Classification and Identification etc., obtain identical effect.It must say It is bright, any one skilled in the art the invention discloses technical scope in, the technique according to the invention Scheme and its inventive concept are subject to equivalent substitution or relevant parameter changes, and should all be included within the scope of the present invention.

Claims (2)

1. a kind of Aerial Images system for rapidly identifying based on multitask topology learning, it is characterized in that including：Feature construction module, Topological characteristic structure module, topological characteristic projection module, multitask topology learning module, feature selection module and Classification and Identification mould Block；

The feature construction module obtains several points under each image classification respectively according to set N number of image classification Class label, and several training Aerial Images for including itself label under each image classification are obtained respectively；

The feature construction module obtains a target Aerial Images to be identified；By the target Aerial Images and the training Aerial Images form Aerial Images；

Each Aerial Images in the Aerial Images are partitioned into using simple linear iteration cluster SLIC dividing methods more Individual zone bit, and form zone bit set；Each zone bit is calculated using probability SVM classifier method to the zone bit set Comentropy, and whether the comentropy for weighing each zone bit is less than set information entropy threshold；From the zone bit set The middle zone bit removed less than described information entropy threshold, so as to generate the zone bit set after optimization；After the optimization Positional information of each zone bit in the Aerial Images in zone bit set, and any two phase obtained by positional information The connection side that adjacent zone bit is generated, build the regional distribution chart of each Aerial Images in the Aerial Images；

Several area adjacency subgraphs are generated by the regional distribution chart of each Aerial Images；By in the area adjacency subgraph The positional information of each zone bit and connection side form the respective topological structure of area adjacency subgraph；By all area adjacencies Subgraph forms the Region adjacency graph of each Aerial Images；

It is sub to several area adjacencies of each Aerial Images using Sub-Image Feature construction method and core principle component analysis method Figure is handled, and obtains the characteristic vector of several area adjacency subgraphs of each Aerial Images, each area adjacency respectively The characteristic vector of figure and the topological structure of respective corresponding area adjacency subgraph correspond；By each Aerial Images The characteristic vector of several area adjacency subgraphs forms the eigenmatrix of respective Aerial Images in the Aerial Images；It is described to take photo by plane The eigenmatrix of image includes the eigenmatrix of target Aerial Images and the eigenmatrix of training Aerial Images；

The topological characteristic structure module will have identical topological structure feature in the eigenmatrix of the training Aerial Images Vector carries out fusion treatment, obtains the topological characteristic matrix of the training Aerial Images；So as to obtain each image classification Under all training Aerial Images topological characteristic matrix；

The topological characteristic projection module is according to itself of several described tag along sorts and each training Aerial Images Label and topological characteristic matrix, obtain the topological characteristic projection square of the training Aerial Images under each Aerial Images classification Battle array；

The multitask topology learning module is using multi-task learning framework and principal component analytical method to the topological characteristic Projection matrix is iterated, so as to obtain the topological encoder matrix of the training Aerial Images；Using F- normal forms to the topology The row vector of encoder matrix is calculated, and the result of calculation of acquisition is ranked up in descending order, before selection corresponding to S row vector Topological structure, form the topological dictionary of the Aerial Images；

The feature selection module is according to each in the eigenmatrix of each topological structure in the topological dictionary and the Aerial Images Topological structure is matched corresponding to characteristic vector, if the match is successful, by the eigenmatrix of the Aerial Images match into The characteristic vector of work(forms the feature selecting vector of the Aerial Images；The feature selecting vector of the Aerial Images includes target The feature selecting vector of the feature selecting vector sum training Aerial Images of Aerial Images；Selected by the feature of each training Aerial Images Vectorial and described training Aerial Images each corresponding image classification is selected, forms the feature selecting two for each training Aerial Images Tuple；

The Classification and Identification module builds multiple disaggregated models to the topological dictionary using SVM classifier；And according to the instruction The tuple of feature selecting two for practicing Aerial Images obtains multiple disaggregated models of optimization using probability SVM methods；By the optimization Multiple disaggregated models build topological Boosting learning models；The feature selecting vector of the target Aerial Images is input to institute State in topological Boosting learning models, so as to obtain the classification of the target Aerial Images.

A kind of 2. quick knowledge of Aerial Images system for rapidly identifying based on multitask topology learning using described in claim 1 Other method, it is characterized in that,

It is C={ c to define N number of image classification₁,c₂,…,c_n,…,c_N}；c_nRepresent n-th image classification；In each image classification Include several tag along sorts；Remember n-th of classification c_nTag along sort bek_nRepresent Tag number, 1≤k_n≤K_n；Represent kth_nIndividual tag along sort；K_nRepresent the n-th image classification c_nMiddle tag along sort Sum；Then the tag along sort collection of N number of image classification C is combined into T={ T₁,T₂,…,T_n,…,T_N}；Note κ is the contingency table The sum of all tag along sorts in set T is signed, and is had

Remember the n-th image classification c_nIn comprising the training Aerial Images of itself label be Represent n-th image classification c_nIn m_nIndividual training Aerial Images,Represent The n-th image classification c_nIn m_nIndividual training Aerial ImagesItself tag set, 1≤m_n≤M_n；M_nRepresent described N image classification c_nIn it is all training Aerial Images sums；Then the training Aerial Images collection of N number of image classification C is combined into I ={ I₁,I₂,…,I_n,…,I_N}；Remember that target Aerial Images are I₀；

The method for quickly identifying is to carry out as follows：

Step 1, structure the n-th image classification c_nIn m_nIndividual training Aerial ImagesRegional distribution chart So as to obtain the n-th image classification c_nIn it is all training Aerial Images regional distribution chartsTaken photo by plane figure so as to obtain the training of N number of image classification C Image set closes I area distribution set of graphs RAG={ RAG₁,RAG₂,…,RAG_n,…,RAG_N}；And the target Aerial Images I₀ Regional distribution chart RAG₀；

Step 1.1, using simple linear iteration cluster SLIC dividing methods by the n-th image classify c_nIn m_nIndividual training Aerial ImagesIt is divided intoIndividual zone bit, obtain the m_nIndividual training Aerial ImagesZone bit set Represent the n-th image classification c_nIn m_nIndividual training is taken photo by plane Image Individual zone bit；The m_nIndividual training Aerial ImagesIn include described theIndividual area Domain memberPositional information；

Step 1.2, utilize the probability SVM classifier method calculating m_nIndividual zone bit setInIndividual zone bitBelong to n-th image classification c_nProbabilitySo as to obtain describedIndividual zone bitBelong to institute State the probability each classified in image classification C；

Step 1.3, using formula (1) calculate described theIndividual zone bitComentropySo as to obtain the m_n Individual zone bit setIn each zone bit comentropy：

<mrow> <mi>&amp;epsiv;</mi> <mrow> <mo>(</mo> <msubsup> <mi>r</mi> <msub> <mi>h</mi> <msub> <mi>m</mi> <mi>n</mi> </msub> </msub> <mrow> <mo>(</mo> <mi>n</mi> <mo>,</mo> <msub> <mi>m</mi> <mi>n</mi> </msub> <mo>)</mo> </mrow> </msubsup> <mo>)</mo> </mrow> <mo>=</mo> <mo>-</mo> <msub> <mi>&amp;Sigma;</mi> <mrow> <msub> <mi>c</mi> <mi>n</mi> </msub> <mo>&amp;Element;</mo> <mi>C</mi> </mrow> </msub> <mi>p</mi> <mrow> <mo>(</mo> <msubsup> <mi>r</mi> <msub> <mi>h</mi> <msub> <mi>m</mi> <mi>n</mi> </msub> </msub> <mrow> <mo>(</mo> <mi>n</mi> <mo>,</mo> <msub> <mi>m</mi> <mi>n</mi> </msub> <mo>)</mo> </mrow> </msubsup> <mo>&amp;RightArrow;</mo> <mi>c</mi> <mo>)</mo> </mrow> <mo>&amp;times;</mo> <mi>log</mi> <mi> </mi> <mi>p</mi> <mrow> <mo>(</mo> <msubsup> <mi>r</mi> <msub> <mi>h</mi> <msub> <mi>m</mi> <mi>n</mi> </msub> </msub> <mrow> <mo>(</mo> <mi>n</mi> <mo>,</mo> <msub> <mi>m</mi> <mi>n</mi> </msub> <mo>)</mo> </mrow> </msubsup> <mo>&amp;RightArrow;</mo> <msub> <mi>c</mi> <mi>n</mi> </msub> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>1</mn> <mo>)</mo> </mrow> </mrow>

Step 1.4, judge described information entropyWhether information entropy threshold δ is less than, if being less than, from the m_nIndividual zone bit SetThe middle zone bit removed less than described information entropy threshold δ；So as to obtain the m_nIndividual training Aerial Images's Zone bit set after optimization

Step 1.5, with the m_nIndividual training Aerial ImagesOptimization after zone bit setIn each zone bit position Central point is put as particle, the particle of the adjacent zone bit of any two in each zone bit is connected using adjacent side Connect, so as to obtain the m_nIndividual training Aerial ImagesRegional distribution chart

Step 2, structure the n-th image classification c_nIn m_nIndividual training Aerial ImagesRegion adjacency graph

From the m_nIndividual training Aerial ImagesRegional distribution chartMiddle exhaustive several particles for including connection side, The m is formed by several described particles and corresponding connection side_nIndividual training Aerial ImagesArea adjacency subgraphRepresent theIndividual area adjacency subgraph, by the area adjacency subgraph In each zone bit positional information and connection side form the topological structure of the area adjacency subgraph；Represent the m_nIt is individual Train Aerial ImagesArea adjacency subgraph sum；By the m_nIndividual training Aerial ImagesArea Domain abuts subgraphForm the n-th image classification c_nIn m_nIt is individual to train figure of taking photo by plane PictureRegion adjacency graphSo as to obtain the n-th image classification c_nIn it is all training Aerial Images area adjacencies FigureSo as to obtain the training Aerial Images set I of N number of image classification C area Domain adjoining set of graphs G={ G₁,G₂,…,G_n,…,G_N}；And the target Aerial Images I₀Regional distribution chart G₀；

Step 3, obtain the n-th image classification c_nIn m_nIndividual training Aerial ImagesEigenmatrixSo as to obtain The n-th image classification c_nTraining Aerial Images eigenmatrixSo as to obtain Obtain the training Aerial Images set I of N number of image classification C eigenmatrix set X={ X₁,X₂,…,X_n,…,X_N}；And The target Aerial Images I₀Eigenmatrix X₀；

Step 3.1, to the m_nIndividual Aerial ImagesRegion adjacency graphMiddle particle number identical area adjacency subgraphWithSub-Image Feature matrix is respectively obtained using Sub-Image Feature construction methodWith

Step 3.2, utilize formula (2) the acquisition Sub-Image Feature matrixWithGolub-Werman distances

<mrow> <msub> <mi>d</mi> <mrow> <mi>G</mi> <mi>W</mi> </mrow> </msub> <mrow> <mo>(</mo> <msubsup> <mi>B</mi> <msub> <mi>i</mi> <msub> <mi>m</mi> <mi>n</mi> </msub> </msub> <mrow> <mo>(</mo> <mi>n</mi> <mo>,</mo> <msub> <mi>m</mi> <mi>n</mi> </msub> <mo>)</mo> </mrow> </msubsup> <mo>,</mo> <msubsup> <mi>B</mi> <msub> <mi>j</mi> <msub> <mi>m</mi> <mi>n</mi> </msub> </msub> <mrow> <mo>(</mo> <mi>n</mi> <mo>,</mo> <msub> <mi>m</mi> <mi>n</mi> </msub> <mo>)</mo> </mrow> </msubsup> <mo>)</mo> </mrow> <mo>=</mo> <mo>|</mo> <mo>|</mo> <msup> <mrow> <mo>(</mo> <msubsup> <mi>B</mi> <msub> <mi>i</mi> <msub> <mi>m</mi> <mi>n</mi> </msub> </msub> <mrow> <mo>(</mo> <mi>n</mi> <mo>,</mo> <msub> <mi>m</mi> <mi>n</mi> </msub> <mo>)</mo> </mrow> </msubsup> <mo>)</mo> </mrow> <mi>U</mi> </msup> <mo>-</mo> <msup> <mrow> <mo>(</mo> <msubsup> <mi>B</mi> <msub> <mi>j</mi> <msub> <mi>m</mi> <mi>n</mi> </msub> </msub> <mrow> <mo>(</mo> <mi>n</mi> <mo>,</mo> <msub> <mi>m</mi> <mi>n</mi> </msub> <mo>)</mo> </mrow> </msubsup> <mo>)</mo> </mrow> <mi>V</mi> </msup> <mo>|</mo> <msub> <mo>|</mo> <mn>2</mn> </msub> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>2</mn> <mo>)</mo> </mrow> </mrow>

In formula (2),WithThe Sub-Image Feature matrix is represented respectivelyWithOrthogonal basis；| |·||₂Represent 2- normal forms；

Step 3.3, utilize formula (3) the calculating area adjacency subgraphWithGaussian kernel

In formula (3),Represent Gauss nuclear parameter；

Step 3.4, repeat step 3.1 arrive step 3.3, so as to obtain the m_nIndividual Region adjacency graphIn it is all have it is identical The Gaussian kernel of the area adjacency subgraph of particle；

Step 3.5, the m obtained using core principle component analysis method_nIndividual Aerial ImagesEigenmatrix For describedIndividual area adjacency subgraphFeature to Amount,D isIndividual characteristic vectorDimension；DescribedIndividual area adjacency subgraphSpy Sign vectorWith describedIndividual area adjacency subgraphTopological structure correspond；

Step 4, obtain the n-th image classification c_nIn m_nIndividual training Aerial ImagesTopological characteristic vector So as to obtain the n-th image classification c_nThe topological characteristic matrix of middle training Aerial ImagesSo as to obtain the training Aerial Images set I of N number of image classification C topology Eigenmatrix set X '={ X₁′,X′₂,…,X′_n,…,X′_N}；

Step 4.1, retain the n-th image classification c_nIn m_nIndividual training Aerial ImagesEigenmatrixIn own In topological structure corresponding to characteristic vector topological structure differ and particle number be no more than upper limit ν topological structure, from And obtain the n-th image classification c_nIn m_nIndividual training Aerial ImagesTopological structure setSo as to obtain State n-th image classification c_nTraining Aerial Images topological structure setUtilize Union processingSo as to obtain the training Aerial Images set I of N number of image classification C topological structure set tp；

Step 4.2, training Aerial Images set I to N number of image classification C topological structure set tp in each topology Structure is numbered, and obtains σ topological structure and is designated as { tp₁′,tp₂′,…,tp′_ο,…,tp′_σ}；tp′_οRepresent the ο topology knot Structure；1≤ο≤σ；

Step 4.3, by the m_nIndividual Aerial ImagesEigenmatrixIn topological structure corresponding to each characteristic vector With the ο topological structure tp '_οIt is compared, selects the characteristic vector with identical topological structure and summed and made even Handle；Obtain the m_nIndividual Aerial ImagesIn the ο topological structure tp '_οOn topology fusion vector So as to obtain the m_nIndividual Aerial ImagesIn the σ topological structure { tp₁′,tp₂′,…,tp ′_ο,…,tp′_σOn topology fusion vector

Step 4.3, by the m_nIndividual Aerial ImagesIn the σ topological structure { tp₁′,tp₂′,…,tp′_ο,…,tp′_σ} On topology fusion vectorEnd to end fusion treatment is carried out, described in acquisition M_nIndividual training Aerial ImagesTopological characteristic vectorAnd haveD σ=d × σ；

Step 5, the topological projection matrix set M={ M for training Aerial Images set I for obtaining N number of image classification C₁, M₂,…,M_n,…,M_N}：

Step 5.1, classified c using formula (4) the acquisition n-th image_nMiddle training Aerial Images I_nTopological transitions matrix

<mrow> <msub> <mover> <mi>X</mi> <mo>&amp;OverBar;</mo> </mover> <mi>n</mi> </msub> <mo>=</mo> <msubsup> <mi>X</mi> <mi>n</mi> <mo>&amp;prime;</mo> </msubsup> <mrow> <mo>(</mo> <msub> <mi>I</mi> <msub> <mi>M</mi> <mi>n</mi> </msub> </msub> <mo>-</mo> <mfrac> <mn>1</mn> <msub> <mi>M</mi> <mi>n</mi> </msub> </mfrac> <msub> <mn>1</mn> <msub> <mi>M</mi> <mi>n</mi> </msub> </msub> <msup> <mrow> <mo>(</mo> <msub> <mn>1</mn> <msub> <mi>M</mi> <mi>n</mi> </msub> </msub> <mo>)</mo> </mrow> <mi>T</mi> </msup> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>4</mn> <mo>)</mo> </mrow> </mrow>

In formula (4), For unit matrix, Represent M_nIndividual element all for 1 one-dimensional row to Amount,

Step 5.2, classified c using formula (5) the acquisition n-th image_nIn m_nIndividual training Aerial ImagesKth_nIndividual label ElementSo as to obtain the n-th image classification c_nIn m_nIndividual training Aerial ImagesLabel vectorSo as to obtain the n-th image classification c_nInstruction Practice Aerial Images I_nLabel matrixAnd have

<mrow> <msubsup> <mi>y</mi> <msub> <mi>k</mi> <mi>n</mi> </msub> <mrow> <mo>(</mo> <mi>n</mi> <mo>,</mo> <msub> <mi>m</mi> <mi>n</mi> </msub> <mo>)</mo> </mrow> </msubsup> <mo>=</mo> <mfenced open = "{" close = ""> <mtable> <mtr> <mtd> <mn>1</mn> </mtd> <mtd> <mrow> <mi>i</mi> <mi>f</mi> <mrow> <mo>(</mo> <mrow> <msubsup> <mi>t</mi> <msub> <mi>k</mi> <mi>n</mi> </msub> <mrow> <mo>(</mo> <mi>n</mi> <mo>)</mo> </mrow> </msubsup> <mo>&amp;Element;</mo> <msubsup> <mi>Tag</mi> <msub> <mi>m</mi> <mi>n</mi> </msub> <mrow> <mo>(</mo> <mi>n</mi> <mo>)</mo> </mrow> </msubsup> </mrow> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> <mtr> <mtd> <mn>0</mn> </mtd> <mtd> <mrow> <mi>e</mi> <mi>l</mi> <mi>s</mi> <mi>e</mi> </mrow> </mtd> </mtr> </mtable> </mfenced> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>5</mn> <mo>)</mo> </mrow> </mrow>

Step 5.3, classified c using formula (6) the acquisition n-th image_nMiddle training Aerial Images I_nLabel measurement oriental matrix F_n,

F_n=Y_n[(Y_n)^TY_n]^-1/2 (6)

Step 5.4, classified c using formula (7) the acquisition n-th image_nMiddle training Aerial Images I_nTopological projection matrix M_n,So as to obtain the training Aerial Images set I of N number of image classification C topological projection matrix set M= {M₁,M₂,…,M_n,…,M_N}；

<mrow> <msub> <mi>M</mi> <mi>n</mi> </msub> <mo>=</mo> <msup> <mrow> <mo>&amp;lsqb;</mo> <msub> <mover> <mi>X</mi> <mo>&amp;OverBar;</mo> </mover> <mi>n</mi> </msub> <mrow> <mo>(</mo> <msub> <mi>I</mi> <msub> <mi>M</mi> <mi>n</mi> </msub> </msub> <mo>-</mo> <msup> <mrow> <mo>(</mo> <msub> <mi>F</mi> <mi>n</mi> </msub> <mo>)</mo> </mrow> <mi>T</mi> </msup> <msub> <mi>F</mi> <mi>n</mi> </msub> <mo>)</mo> </mrow> <msup> <mrow> <mo>(</mo> <msub> <mover> <mi>X</mi> <mo>&amp;OverBar;</mo> </mover> <mi>n</mi> </msub> <mo>)</mo> </mrow> <mi>T</mi> </msup> <mo>-</mo> <mi>&amp;gamma;</mi> <msup> <mrow> <mo>(</mo> <msub> <mover> <mi>X</mi> <mo>&amp;OverBar;</mo> </mover> <mi>n</mi> </msub> <mo>)</mo> </mrow> <mi>T</mi> </msup> <msub> <mover> <mi>X</mi> <mo>&amp;OverBar;</mo> </mover> <mi>n</mi> </msub> <mo>&amp;rsqb;</mo> </mrow> <mi>T</mi> </msup> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>7</mn> <mo>)</mo> </mrow> </mrow>

Step 6, multitask topology learning feature coding：

Step 6.1, definition iterations are t；Initialize the iterations t=1；

And the topological encoder matrix of the t times iteration is described in when initializing the t times iterationAnd n-th of topological coded sub-matrices when having the t times iterationIn each member Element is obtained by random function；

Step 6.2, the topological encoder matrix for obtaining using formula (8) the t times iteration are W^(t)Code weight matrix D^(t)'s V row v column elementsCode weight matrix D during so as to obtain the t times iteration^(t)：

<mrow> <msubsup> <mi>d</mi> <mrow> <mi>v</mi> <mo>,</mo> <mi>v</mi> </mrow> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> </msubsup> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <mn>2</mn> <mo>|</mo> <mo>|</mo> <msubsup> <mi>w</mi> <mi>v</mi> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> </msubsup> <mo>|</mo> <msub> <mo>|</mo> <mn>2</mn> </msub> </mrow> </mfrac> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>8</mn> <mo>)</mo> </mrow> </mrow>

In formula (8),The topological encoder matrix for representing the t times iteration is W^(t)V every trades vector, the code weight Matrix D^(t)For diagonal matrix,

Step 6.3, definition tasking learning number are l；1≤l≤N；And initialize the tasking learning number l=1；

Step 6.4, the weight selection matrix that the t times iteration l subtask learns is obtained using formula (9)In v rows v arrange ElementSo as to obtain the weight selection matrix of the t times iteration l subtasks study

<mrow> <msubsup> <mi>d</mi> <mrow> <mi>l</mi> <mo>,</mo> <mi>v</mi> <mo>,</mo> <mi>v</mi> </mrow> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> </msubsup> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <mn>2</mn> <mo>|</mo> <mo>|</mo> <msubsup> <mi>w</mi> <mrow> <mi>l</mi> <mo>,</mo> <mi>v</mi> </mrow> <mrow> <mo>(</mo> <mi>t</mi> <mo>-</mo> <mn>1</mn> <mo>)</mo> </mrow> </msubsup> <mo>|</mo> <msub> <mo>|</mo> <mn>2</mn> </msub> </mrow> </mfrac> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>9</mn> <mo>)</mo> </mrow> </mrow>

In formula (9),Represent l-th of topological coded sub-matrices during the t times iterationV row vectors, weight choosing Select matrixFor diagonal matrix,

Step 6.5, from the topological characteristic projection matrix set M={ M₁,M₂,…,M_n,…,M_NL-th of topological characteristic of middle extraction Projection matrix M_l；The code optimization matrix of the t times iteration l subtasks study is obtained using formula (10)

<mrow> <msubsup> <mi>U</mi> <mi>l</mi> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> </msubsup> <mo>=</mo> <msub> <mi>M</mi> <mi>l</mi> </msub> <mo>+</mo> <mi>&amp;lambda;</mi> <mrow> <mo>(</mo> <msubsup> <mi>D</mi> <mi>l</mi> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> </msubsup> <mo>+</mo> <msup> <mi>&amp;mu;D</mi> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> </msup> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>10</mn> <mo>)</mo> </mrow> </mrow>

In formula (10), λ presentation code Optimal Parameters；μ presentation code internal coordination parameters；

Step 6.6, note K_lFor l-th of image classification c in N number of image classification C_lThe sum of middle tag along sort；Using it is main into Analysis method is divided to obtain the code optimization matrix of the t times iteration l subtasks studyMultiple characteristic values and with multiple features It is worth corresponding characteristic vector, the multiple characteristic value is sorted from small to large, and selects preceding K_lSpy corresponding to individual characteristic value N-th of topological coded sub-matrices when sign vector is as the t times iteration

Step 6.7, judge whether l=N sets up, if so, n times tasking learning is then completed, so as to obtain opening up for the t+1 times iteration Flutter encoder matrixAnd perform step 6.8；Otherwise l+1 value is assigned to l, repetition is held Row step 6.4- steps 6.7；

Step 6.8, judge whether t=Γ set up, Γ is iteration total degree；If so, perform step 6.9；Otherwise by t+1 value T is assigned to, return to step 6.2 sequentially performs；

Step 6.9, using F- normal forms to the t+1 times iteration topology encoder matrix W^(t+1)Row vector calculated, obtain Result of calculation be ranked up in descending order, select the t+1 times iteration topology encoder matrix W^(t+1)Before middle sequence S row to Amount, the preceding S row vector is obtained in the t+1 times iteration topology encoder matrix W^(t+1)In corresponding line number { h₁, h₂,…,h_s,…,h_S, h_sRepresent that s-th of row vector is compiled in the t+1 times iteration topology in S row vector before the sequence Code matrix W^(t+1)In line number；1≤s≤S；

Step 6.10, from the topological structure { tp₁′,tp₂′,…,tp′_ο,…,tp′_σIn selection theIndividual topological structure, So as to obtain the topological structure before the sequence in S row vector corresponding to s-th of row vector；So as to obtain S before the sequence Topological structure corresponding to individual row vector, and form the topological dictionary Tp={ tp of the Aerial Images₁″,tp₂″,…,tp ″_s′,…,tp″_S′}；tp″_s′Represent the individual topological structures of s ' in the topological dictionary Tp；1≤s′≤S′；

Step 7, obtain the n-th image classification c_nIn m_nIndividual training Aerial ImagesFeature selecting vectorAnd The target Aerial Images I₀Feature selecting vector x₀″；So as to obtain the n-th image classification c_nIn m_nIndividual training Aerial ImagesThe tuple of feature selecting twoSo as to obtain the n-th image classification c_nMiddle training is taken photo by plane Image I_nThe tuple of feature selecting twoSo as to obtain The training Aerial Images set I of the N number of image classification C tuple-set of feature selecting twoNote Represent the training Aerial Images set I of the N number of image classification C tuple-set X of feature selecting two_FSE-th The tuple of feature selecting two, 1≤e≤E；

Step 7.1, definition selection number are ρ, and initialize ρ=1；1≤ρ≤S′；

Define the described the ρ times m_nIndividual training Aerial ImagesFeature Choice Matrix beAnd initialize described The ρ m_nIndividual training Aerial ImagesFeature Choice Matrix beFor sky；Define the ρ times m_nIndividual training Aerial ImagesFeature extraction matrix beAnd initialize

Step 7.2, choose the ρ topological structure tp " from topological dictionary Tp_ρ；

Step 7.3, judge the ρ topological structure tp "_ρIn particle number whether be more than the m_nIt is individual to train figure of taking photo by plane PictureRegional distribution chartParticle number, if being more than, ρ+1 value is assigned to perform step 7.2 after ρ；Otherwise, Perform step 7.4；

Step 7.4, by the ρ topological structure tp "_ρWith the ρ times m_nIndividual training Aerial ImagesEigenmatrixIn topological structure corresponding to each characteristic vector matched；If the match is successful, from the ρ times m_nIndividual instruction Practice Aerial ImagesFeature extraction matrixMiddle removal and the ρ topological structure tp "_ρThe topology knot that the match is successful Characteristic vector corresponding to structure, and it is put into the ρ times m_nIndividual training Aerial ImagesFeature Choice Matrix be In；

Step 7.5, ρ+1 value is assigned to ρ, judges whether ρ=S ' sets up, if so, using maximum pond strategy to the m_nIt is individual Train Aerial ImagesFeature Choice MatrixHandled, obtain the m_nIndividual training Aerial ImagesFeature Selection vectorOtherwise, step 7.2- steps 7.5 are repeated；

Step 8, the topological Boosting learning models of structure：

Step 8.1, utilize the individual topological structure tp " of s ' in the one-against-rest structure topological dictionary Tp of SVM classifier_s′'s Disaggregated model ψ_s′；So as to obtain the individual disaggregated model { ψ of S ' of the topological dictionary Tp₁,ψ₂,...,ψ_s′,...,ψ_S′}；

Step 8.2, the iterations of definition Boosting learning models are τ, and initialize iterations τ=1；

DefinitionRepresent the training Aerial Images set I of the N number of image classification C during the τ times iteration tuple of feature selecting two Set X_FSThe tuple of e-th of feature selecting twoClass weight；Then N number of image classification C during the τ times iteration Train the Aerial Images set I tuple-set X of feature selecting two_FSClass weight beAnd just Beginningization

Step 8.3, using formula (11) obtain the τ times iteration optimization disaggregated model ψ^(τ)；

<mrow> <msup> <mi>&amp;psi;</mi> <mrow> <mo>(</mo> <mi>&amp;tau;</mi> <mo>)</mo> </mrow> </msup> <mo>=</mo> <mi>arg</mi> <mi> </mi> <msub> <mi>min</mi> <mrow> <msup> <mi>&amp;psi;</mi> <mrow> <mo>(</mo> <mi>&amp;tau;</mi> <mo>)</mo> </mrow> </msup> <mo>&amp;Element;</mo> <mo>{</mo> <msub> <mi>&amp;psi;</mi> <mn>1</mn> </msub> <mo>,</mo> <msub> <mi>&amp;psi;</mi> <mn>2</mn> </msub> <mo>,</mo> <mo>...</mo> <mo>,</mo> <msub> <mi>&amp;psi;</mi> <msup> <mi>s</mi> <mo>&amp;prime;</mo> </msup> </msub> <mo>,</mo> <mo>...</mo> <mo>,</mo> <msub> <mi>&amp;psi;</mi> <msup> <mi>S</mi> <mo>&amp;prime;</mo> </msup> </msub> <mo>}</mo> </mrow> </msub> <msubsup> <mi>&amp;Sigma;</mi> <mrow> <mi>e</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>E</mi> </msubsup> <msubsup> <mi>&amp;omega;</mi> <mi>e</mi> <mrow> <mo>(</mo> <mi>&amp;tau;</mi> <mo>)</mo> </mrow> </msubsup> <mo>&amp;times;</mo> <mo>&amp;lsqb;</mo> <mn>1</mn> <mo>-</mo> <msub> <mi>p</mi> <mi>&amp;tau;</mi> </msub> <mrow> <mo>(</mo> <msubsup> <mi>x</mi> <mi>e</mi> <mrow> <mi>F</mi> <mi>S</mi> </mrow> </msubsup> <mo>&amp;RightArrow;</mo> <msubsup> <mi>c</mi> <mi>e</mi> <mrow> <mi>F</mi> <mi>S</mi> </mrow> </msubsup> <mo>)</mo> </mrow> <mo>&amp;rsqb;</mo> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>11</mn> <mo>)</mo> </mrow> </mrow>

In formula (11),Represent using probability SVM methods to the disaggregated model ψ^(τ)Handle the institute of acquisition State e-th of tuple of feature selecting twoFeature selecting vectorBelong to image classificationProbability；

Step 8.4, utilize identification mistake err during formula (12) the τ times iteration of calculating_τ：

<mrow> <msub> <mi>err</mi> <mi>&amp;tau;</mi> </msub> <mo>=</mo> <msubsup> <mi>&amp;Sigma;</mi> <mrow> <mi>e</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>E</mi> </msubsup> <msubsup> <mi>&amp;omega;</mi> <mi>e</mi> <mrow> <mo>(</mo> <mi>&amp;tau;</mi> <mo>)</mo> </mrow> </msubsup> <mo>&amp;times;</mo> <mo>&amp;lsqb;</mo> <mn>1</mn> <mo>-</mo> <msub> <mi>p</mi> <mi>&amp;tau;</mi> </msub> <mrow> <mo>(</mo> <msubsup> <mi>x</mi> <mi>e</mi> <mrow> <mi>F</mi> <mi>S</mi> </mrow> </msubsup> <mo>&amp;RightArrow;</mo> <msubsup> <mi>c</mi> <mi>e</mi> <mrow> <mi>F</mi> <mi>S</mi> </mrow> </msubsup> <mo>)</mo> </mrow> <mo>&amp;rsqb;</mo> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>12</mn> <mo>)</mo> </mrow> </mrow>

Step 8.5, utilize identification weight α during formula (13) the τ times iteration of calculating_τ：

<mrow> <msub> <mi>&amp;alpha;</mi> <mi>&amp;tau;</mi> </msub> <mo>=</mo> <mi>l</mi> <mi>o</mi> <mi>g</mi> <mfrac> <mrow> <mo>(</mo> <mn>1</mn> <mo>-</mo> <msub> <mi>err</mi> <mi>&amp;tau;</mi> </msub> <mo>)</mo> </mrow> <mrow> <msub> <mi>err</mi> <mi>&amp;tau;</mi> </msub> </mrow> </mfrac> <mo>+</mo> <mi>l</mi> <mi>o</mi> <mi>g</mi> <mrow> <mo>(</mo> <mi>N</mi> <mo>-</mo> <mn>1</mn> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>13</mn> <mo>)</mo> </mrow> </mrow>

Step 8.6, using formula (14) calculate the τ times iteration when described in e-th of class weightCorresponding classification transition weightSo as to obtain E classification transition weight

<mrow> <msubsup> <mi>&amp;omega;</mi> <mi>e</mi> <mrow> <mo>&amp;prime;</mo> <mrow> <mo>(</mo> <mi>&amp;tau;</mi> <mo>)</mo> </mrow> </mrow> </msubsup> <mo>=</mo> <msubsup> <mi>&amp;omega;</mi> <mi>e</mi> <mrow> <mo>(</mo> <mi>&amp;tau;</mi> <mo>)</mo> </mrow> </msubsup> <mo>&amp;times;</mo> <mi>exp</mi> <mo>{</mo> <msub> <mi>&amp;alpha;</mi> <mi>&amp;tau;</mi> </msub> <mo>&amp;times;</mo> <mo>&amp;lsqb;</mo> <mn>1</mn> <mo>-</mo> <msub> <mi>p</mi> <mi>&amp;tau;</mi> </msub> <mrow> <mo>(</mo> <msubsup> <mi>X</mi> <mi>e</mi> <mrow> <mi>F</mi> <mi>S</mi> </mrow> </msubsup> <mo>&amp;RightArrow;</mo> <msubsup> <mi>c</mi> <mi>e</mi> <mrow> <mi>F</mi> <mi>S</mi> </mrow> </msubsup> <mo>)</mo> </mrow> <mo>&amp;rsqb;</mo> <mo>}</mo> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>14</mn> <mo>)</mo> </mrow> </mrow>

Step 8.7, to the E classification transition weightIt is normalized, obtains τ+1 Class weight during secondary iteration

Step 8.8, τ+1 value is assigned to τ, repeats step 8.3- steps 8.8；Untill τ is equal to iterations A, from And obtain identification weight set { α₁,α₂,…,α_a,…,α_AAnd the A iteration optimization disaggregated model { ψ⁽¹⁾,ψ⁽²⁾,…, ψ^(a),…,ψ^(A)}；α_aRepresent a-th of identification weight in the identification weight set；ψ^(a)Represent and described a-th identification weight The disaggregated model of corresponding optimization；

Step 8.9, utilize the topological Boosting learning models acquisition target Aerial Images I shown in formula (15)₀Image point Class Δ；So as to realize the quick identification of target Aerial Images：

<mrow> <mi>&amp;Delta;</mi> <mo>=</mo> <mi>arg</mi> <mi> </mi> <msub> <mi>min</mi> <mrow> <mi>&amp;Delta;</mi> <mo>&amp;Element;</mo> <mi>C</mi> <mo>=</mo> <mo>{</mo> <msub> <mi>c</mi> <mn>1</mn> </msub> <mo>,</mo> <msub> <mi>c</mi> <mn>2</mn> </msub> <mo>,</mo> <mo>...</mo> <mo>,</mo> <msub> <mi>c</mi> <mi>n</mi> </msub> <mo>,</mo> <mo>...</mo> <mo>,</mo> <msub> <mi>c</mi> <mi>N</mi> </msub> <mo>}</mo> </mrow> </msub> <msubsup> <mi>&amp;Sigma;</mi> <mrow> <mi>a</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>A</mi> </msubsup> <msub> <mi>&amp;alpha;</mi> <mi>a</mi> </msub> <mo>&amp;times;</mo> <msub> <mi>p</mi> <mi>a</mi> </msub> <mrow> <mo>(</mo> <msubsup> <mi>x</mi> <mn>0</mn> <mrow> <mo>&amp;prime;</mo> <mo>&amp;prime;</mo> </mrow> </msubsup> <mo>&amp;RightArrow;</mo> <mi>&amp;Delta;</mi> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>15</mn> <mo>)</mo> </mrow> </mrow>

In formula (15), p_a(x₀" → Δ) represent the disaggregated model ψ to described a-th optimization using probability SVM methods^(a)Located Manage the target Aerial Images I obtained₀Feature selecting vector x₀" belong to the probability of image classification Δ.