CN104598898A - Aerially photographed image quick recognizing system and aerially photographed image quick recognizing method based on multi-task topology learning - Google Patents

Abstract

The invention discloses an aerially photographed image quick recognizing system and an aerially photographed image quick recognizing method based on multi-task topology learning. The aerially photographed image quick recognizing system is characterized by comprising a feature establishing module, a topology feature establishing module, a topology feature projection module, a multi-task topology learning module, a feature selecting module and a classifying and recognizing module, wherein the feature establishing module is used for acquiring a topology feature matrix of aerially photographed images and training aerially photographed images; the topology feature establishing module is used for acquiring a topology feature projection matrix of the training aerially photographed images; the multi-task topology learning module is used for acquiring a topology dictionary of the aerially photographed images; the feature selecting module is used for acquiring feature selecting two-tuples of the training aerially photographed images; and the classifying and recognizing module is used for establishing a topology Boosting learning module and classifying objective aerially photographed images. Classification of the aerially photographed images can be recognized quickly by a multi-task learning frame, and the classifying precision and the classifying efficiency are improved.

Description

A kind of Aerial Images system for rapidly identifying based on multitask topology learning and method for quickly identifying thereof

Technical field

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

Background technology

Aerial Images is as the important component part of remote sensing images, and the data message a large amount of with it and geometric resolution receive widespread attention, and therefore carries out discriminator to Aerial Images significant.The quick discriminator of Aerial Images is a kind of intelligent image analytical approach, and its object is exactly to identify fast and accurately Aerial Images and classifying, and is one of popular research topic of computer vision field.The classification of Aerial Images can be applied to the aspects such as scene mark, video monitoring, robot path planning, abnormality detection.

Along with the development of Morden Image Processing Technology, Aerial Images sorting technique have also been obtained huge development, but still faces some problem following:

One, in existing method, in the Aerial Images of structure, zone bit distribution plan representativeness is not high, can not accurately represent the constituent of Aerial Images.

Such as 2011, the objective classification method of a kind of graph structure based on core coupling proposed in " A Graph-Matching Kernel for Object Categorization " this section of article that the authors such as Duchenne deliver on top international conference IEEE International Conference on ComputerVision (ICCV), in the method, the node on behalf of figure image lattice, the annexation that what geometry fillet represented is between grid, with the classification of the coupling realize target of so a kind of figure based on grid, but because the representativeness of the graph structure of the Aerial Images of structure is not high, can accurately not represent the constituent of Aerial Images, reduce the efficiency of Aerial Images classification.

Two, the topological structure considering image is lacked in a lot of Aerial Images assorting processes.

Such as 2010, the image classification method proposed in " Object-Graphs for Context-Aware Category Discovery " this section of article that the authors such as Yong Jae Lee deliver on top international conference IEEE Conference on Computer Visionand Pattern Recognition (CVPR), in the method, the node of figure represents known and unknown target, the semanteme of each unknown object utilizes the semanteme of its k-space arest neighbors known target to infer, but only consider the knowledge of spatially adjacent part, i.e. semantic another important clue of deduction.And topological structure is not considered in the process of classification, thus reduce the efficiency of Images Classification.

Three, the Aerial Images discriminator of existing a lot of consideration topological relation is based on random walk and the coupling setting strategy, limits the matching range of topological relation in Aerial Images.

Such as 2007, a kind of image classification method proposed in " Image Classification with Segmentation Graph Kernels " this section of article that the authors such as Harchaoui deliver on international conference IEEE Conference on Computer Vision andPattern Recognition, the method utilizes the method for random walk core and tree kernel function to use the adjacent domain of finite sequence to remove to extract the spatial topotaxy of random walk between topography's gray scale and tree, the coupling of random walk and tree strategy is finally utilized to realize the discriminator of image, but the method has the phenomenon of walking haltingly, thus introduce noise, limit the coupling of Aerial Images topology, thus reduce the scope of Aerial Images classification, be difficult in real life be applied.

Four, existing a lot of Aerial Images can not the space layout of differentiate between images for the very similar image of structure in the process of classification, reduces the degree of accuracy of Images Classification.

Such as 2006, the recognition methods of a kind of natural scene classification based on spatial pyramid coupling proposed in " Spatial Pyramid Matching for RecognizingNatural Scene Categories " this section of article that the authors such as Svetlana Lazebnik deliver on top international conference IEEE Conference on ComputerVision and Pattern Recognition (CVPR), in the method, each image is divided into more and more thinner grid, each grid calculates its feature descriptor by local feature histogram, but when the space layout of two images is very similar, the method can not be distinguished, thus reduce the degree of accuracy of Images Classification, and complexity is high, be difficult to be applied in practice.

Five, the design of the sorter of existing Aerial Images classification is substantially based on SVM classifier, and between the sorter of the multiple inter-related task of shortage consideration, information is shared, and cause the process more complicated calculated, the efficiency of classification is low.

Such as 2006, " Spatial Pyramid Matching for RecognizingNatural Scene Categories " the method that the authors such as Svetlana Lazebnik deliver on top international conference IEEE Conference on ComputerVision and Pattern Recognition (CVPR) is a kind of non-linear SVM classifier in the design of Aerial Images sorter, but sharing due to the information of classifier between the multiple inter-related task of shortage consideration in assorting process in the method, the degree of accuracy that result in Aerial Images classification is not high, there is certain limitation.

Therefore, up to the present, still do not occur a kind of consider image topology relation nicety of grading and efficiency high and can the Aerial Images sorting technique of engineer applied.

Summary of the invention

The weak point that the present invention exists to solve prior art, proposes a kind of Aerial Images system for rapidly identifying based on multitask topology learning and method for quickly identifying thereof, to realizing the Fast Classification of Aerial Images, thus improves speed and the precision of classification.

The present invention is that solution problem adopts scheme with the following method:

The feature of a kind of Aerial Images system for rapidly identifying based on multitask topology learning of the present invention comprises: feature construction module, topological characteristic build module, topological characteristic projection module, multitask topology learning module, feature selection module and Classification and Identification module;

Described feature construction module, according to set N number of Images Classification, obtains several tag along sorts under each Images Classification respectively, and several under obtaining each Images Classification respectively comprise the training Aerial Images of self label;

Described feature construction module obtains a target Aerial Images to be identified; Aerial Images is formed by described target Aerial Images and described training Aerial Images;

Utilize SLIC dividing method to be partitioned into multiple zone bit to each Aerial Images in described Aerial Images, and form zone bit set; Utilize probability SVM classifier method to calculate the information entropy of each zone bit to described zone bit set, and whether the information entropy weighing each zone bit is less than set information entropy threshold value; From described zone bit set, remove the zone bit being less than described information entropy threshold value, thus generate the zone bit set after optimizing; According to the positional information of each zone bit in described Aerial Images in the zone bit set after described optimization, and the fillet that any two the adjacent zone bits to be obtained by positional information generate, build the regional distribution chart of each Aerial Images in described Aerial Images;

Several area adjacency subgraphs are generated by the regional distribution chart of described each Aerial Images; Described area adjacency subgraph topological structure is separately formed by the positional information of each zone bit in described area adjacency subgraph and fillet; The Region adjacency graph of each Aerial Images is made up of all area adjacency subgraphs;

Sub-Image Feature construction method and core principle component analysis method several area adjacency subgraphs to described each Aerial Images are utilized to process, obtain the proper vector of several area adjacency subgraphs of each Aerial Images respectively, the proper vector of each area adjacency subgraph and the topological structure one_to_one corresponding of area adjacency subgraph corresponding separately; The eigenmatrix of respective Aerial Images in described Aerial Images is made up of the proper vector of several area adjacency subgraphs of described each Aerial Images; The eigenmatrix of described Aerial Images comprises the eigenmatrix of target Aerial Images and the eigenmatrix of training Aerial Images;

Described topological characteristic builds module and the proper vector in the eigenmatrix of described training Aerial Images with identical topological structure is carried out fusion treatment, obtains the topological characteristic matrix of described training Aerial Images; Thus the topological characteristic matrix of all training Aerial Images under obtaining described each Images Classification;

Described topological characteristic projection module, according to self label of several tag along sorts described and described each training Aerial Images and topological characteristic matrix, obtains the topological characteristic projection matrix of the training Aerial Images under described each Aerial Images classification;

Described multitask topology learning module utilizes multi-task learning framework and principal component analytical method to carry out iteration to described topological characteristic projection matrix, thus obtains the topological encode matrix of described training Aerial Images; Utilize the row vector of F-normal form to described topological encode matrix to calculate, the result of calculation of acquisition sorts by descending, and the topological structure before selecting corresponding to S row vector, forms the topological dictionary of described Aerial Images;

The described feature selection module topological structure corresponding with each proper vector in the eigenmatrix of described Aerial Images according to each topological structure in described topological dictionary mates, if the match is successful, be then made up of the feature selecting vector of described Aerial Images the proper vector that in the eigenmatrix of described Aerial Images, the match is successful; The feature selecting vector of described Aerial Images comprises the feature selecting vector of the feature selecting vector sum training Aerial Images of target Aerial Images; By the Images Classification that feature selecting is vectorial and described training Aerial Images is corresponding separately of each training Aerial Images, form feature selecting two tuple of each training Aerial Images;

Described Classification and Identification module adopts SVM classifier to build multiple disaggregated model to described topological dictionary; And adopt probability SVM method to obtain the multiple disaggregated models optimized according to feature selecting two tuple of described training Aerial Images; Topological Boosting learning model is built by multiple disaggregated models of described optimization; The feature selecting vector of described target Aerial Images is input in described topological Boosting learning model, thus obtains the classification of described target Aerial Images.

The feature of the method for quickly identifying of a kind of Aerial Images system for rapidly identifying based on multitask topology learning of the present invention is:

Defining N number of Images Classification is C={c ₁, c ₂..., c _n..., c _n; c _nrepresent the n-th Images Classification; Several tag along sorts are comprised in each Images Classification; Remember described n-th classification c _ntag along sort be k _nrepresent tag number, 1≤k _n≤ K _n; represent kth _nindividual tag along sort; K _nrepresent described n-th Images Classification c _nthe sum of middle tag along sort; Then the tag along sort set of described N number of Images Classification C is T={T ₁, T ₂..., T _n..., T _n; Note κ is the sum of all tag along sorts in described tag along sort set T, and has

Remember described n-th Images Classification c _nin comprise self label training Aerial Images be represent the n-th Images Classification c _nin m _nindividual training Aerial Images, represent described n-th Images Classification c _nin m _nindividual training Aerial Images self tag set, 1≤m _n≤ M _n; M _nrepresent described n-th Images Classification c _nin the sum of all training Aerial Images; Then the training Aerial Images set of described N number of Images Classification C is I={I ₁, I ₂..., I _n..., I _n; Note target Aerial Images is I ₀;

Described method for quickly identifying carries out as follows:

Step 1, build described n-th Images Classification c _nin m _nindividual training Aerial Images regional distribution chart thus obtain described n-th Images Classification c _nin the regional distribution chart of all training Aerial Images ${\mathrm{RAG}}_n=\{{\mathrm{RAG}}_1^{\left(n\right)},{\mathrm{RAG}}_2^{\left(n\right)},...,{\mathrm{RAG}}_{m_n}^{\left(n\right)},...,{\mathrm{RAG}}_{M_n}^{\left(n\right)}\};$ Thus obtain the areal distribution set of graphs RAG={RAG of the training Aerial Images set I of described N number of Images Classification C ₁, RAG ₂..., RAG _n..., RAG _n; And described target Aerial Images I ₀regional distribution chart RAG ₀;

Step 1.1, utilize SLIC dividing method by described n-th Images Classification c _nin m _nindividual training Aerial Images be divided into individual zone bit, obtains described m _nindividual training Aerial Images zone bit set represent described n-th Images Classification c _nin m _nindividual training Aerial Images ? individual zone bit; described m _nindividual training Aerial Images in comprise described individual zone bit positional information;

Step 1.2, utilize probability SVM classifier method calculate described m _nindividual zone bit set in individual zone bit belong to the n-th Images Classification c _nprobability thus obtain described individual zone bit belong to the probability of each classification in described Images Classification C;

Step 1.3, formula (1) is utilized to calculate described individual zone bit information entropy thus obtain described m _nindividual zone bit set in the information entropy of each zone bit:

$\mathrm{\ϵ}\left(r_{h_{m_n}}^{(n,m_n)}\right)=-{\mathrm{\Σ}}_{c_n\∈C}p(r_{h_{m_n}}^{(n,m_n)}\→c_n)\×\log p(r_{h_{m_n}}^{(n,m_n)}\→c_n)---\left(1\right)$

Step 1.4, judge described information entropy whether be less than information entropy threshold value δ, if be less than, from described m _nindividual zone bit set middle removal is less than the zone bit of described information entropy threshold value δ; Thus obtain described m _nindividual training Aerial Images optimization after zone bit set

Step 1.5, with described m _nindividual training Aerial Images optimization after zone bit set in the place-centric point of each zone bit as particle, adopt adjacent side to connect the particle of two adjacent zone bits any in described each zone bit, thus obtain described m _nindividual training Aerial Images regional distribution chart

Step 2, build described n-th Images Classification c _nin m _nindividual training Aerial Images region adjacency graph

From described m _nindividual training Aerial Images regional distribution chart in exhaustive several particles comprising fillet, form described m by several particles described and corresponding fillet _nindividual training Aerial Images area adjacency subgraph represent the individual area adjacency subgraph, is made up of the topological structure of described area adjacency subgraph the positional information of each zone bit in described area adjacency subgraph and fillet; represent described m _nindividual training Aerial Images the sum of area adjacency subgraph; by described m _nindividual training Aerial Images area adjacency subgraph form described n-th Images Classification c _nin m _nindividual training Aerial Images region adjacency graph thus obtain described n-th Images Classification c _nin the Region adjacency graph of all training Aerial Images thus obtain the area adjacency set of graphs G={G of the training Aerial Images set I of described N number of Images Classification C ₁, G ₂..., G _n..., G _n; And described target Aerial Images I ₀regional distribution chart G ₀;

Step 3, obtain described n-th Images Classification c _nin m _nindividual training Aerial Images eigenmatrix thus obtain described n-th Images Classification c _nthe eigenmatrix of training Aerial Images thus obtain the eigenmatrix set X={X of the training Aerial Images set I of described N number of Images Classification C ₁, X ₂..., X _n..., X _n; And described target Aerial Images I ₀eigenmatrix X ₀;

Step 3.1, to described m _nindividual Aerial Images region adjacency graph the area adjacency subgraph that middle particle number is identical with sub-Image Feature construction method is adopted to obtain Sub-Image Feature matrix respectively with $1\≤i_{m_n},j_{m_n}\≤Q_{m_n};$

Step 3.2, formula (2) is utilized to obtain described Sub-Image Feature matrix with golub-Werman distance $d_{\mathrm{GW}}(B_{i_{m_n}}^{(n,m_n)},B_{j_{m_n}}^{(n,m_n)}):$

$d_{\mathrm{GW}}(B_{i_{m_n}}^{(n,m_n)},B_{j_{m_n}}^{(n,m_n)})={\left|\right|{\left(B_{i_{m_n}}^{(n,m_n)}\right)}^U-{\left(B_{j_{m_n}}^{(n,m_n)}\right)}^V\left|\right|}_2---\left(2\right)$

In formula (2), with represent described Sub-Image Feature matrix respectively with orthogonal basis; || || ₂represent 2-normal form;

Step 3.3, formula (3) is utilized to calculate described area adjacency subgraph with gaussian kernel

In formula (3), represent gaussian kernel parameter;

Step 3.4, repetition step 3.1 to step 3.3, thus obtain described m _nindividual Region adjacency graph in all gaussian kernel with the area adjacency subgraph of phase homologous pints;

Step 3.5, employing core principle component analysis method obtain described m _nindividual Aerial Images eigenmatrix $X_{m_n}^{\left(n\right)}=[x_1^{(n,m_n)},x_2^{(n,m_n)},...,x_{q_{m_n}}^{(n,m_n)},...,x_{Q_{m_n}}^{(n,m_n)}];$ for described individual area adjacency subgraph proper vector, d is individual proper vector dimension; Described individual area adjacency subgraph proper vector with described individual area adjacency subgraph topological structure one_to_one corresponding;

Step 4, obtain described n-th Images Classification c _nin m _nindividual training Aerial Images topological characteristic vector thus obtain described n-th Images Classification c _nthe topological characteristic matrix of middle training Aerial Images thus obtain the topological characteristic set of matrices X ' of the training Aerial Images set I of described N number of Images Classification C=X ' ₁, X ' ₂..., X ' _n..., X ' _n;

Step 4.1, retain described n-th Images Classification c _nin m _nindividual training Aerial Images eigenmatrix in in topological structure corresponding to all proper vectors topological structure not identical and particle number is no more than the topological structure of upper limit ν, thus obtain described n-th Images Classification c _nin m _nindividual training Aerial Images topological structure set thus obtain described n-th Images Classification c _nthe topological structure set of training Aerial Images utilize union process thus obtain the topological structure set tp of the training Aerial Images set I of described N number of Images Classification C;

Step 4.2, each topological structure in the topological structure set tp of the training Aerial Images set I of described N number of Images Classification C to be numbered, obtain σ topological structure be designated as tp ' ₁, tp ' ₂..., tp ' _ο..., tp ' _σ; Tp ' _οrepresent ο topological structure; 1≤ο≤σ;

Step 4.3, by described m _nindividual Aerial Images eigenmatrix in topological structure corresponding to each proper vector and described ο topological structure tp ' _οcompare, select the proper vector with identical topological structure and carry out suing for peace and being averaged process; Obtain described m _nindividual Aerial Images at ο topological structure tp ' _οon topology merge vector thus obtain described m _nindividual Aerial Images a described σ topological structure tp ' ₁, tp ' ₂..., tp ' _ο..., tp ' _σon topology merge vector

Step 4.3, by described m _nindividual Aerial Images a described σ topological structure tp ' ₁, tp ' ₂..., tp ' _ο..., tp ' _σon topology merge vector carry out end to end fusion treatment, obtain described m _nindividual training Aerial Images topological characteristic vector and have d σ=d × σ;

Step 5, obtain the topological projection matrix set M={M of the training Aerial Images set I of described N number of Images Classification C ₁, M ₂..., M _n..., M _n}:

Step 5.1, formula (4) is utilized to obtain described n-th Images Classification c _nmiddle training Aerial Images I _ntopological transitions matrix

${\stackrel{\‾}{X}}_n=X_n^{\′}(I_{M_n}-\frac{1}{M_n}{1}_{M_n}{\left(1_{M_n}\right)}^T)---\left(4\right)$

In formula (4), for unit matrix, represent M _nindividual element is all a dimensional vector of 1,

Step 5.2, formula (5) is utilized to obtain described n-th Images Classification c _nin m _nindividual training Aerial Images kth _nindividual tag element thus obtain described n-th Images Classification c _nin m _nindividual training Aerial Images label vector $y_{m_n}^{\left(n\right)}=[y_1^{(n,m_n)},y_2^{(n,m_n)},...,y_{k_n}^{(n,m_n)},...,y_{K_n}^{(n,m_n)}],$ thus obtain described n-th Images Classification c _ntraining Aerial Images I _nlabel matrix $Y_n=[y_1^{\left(n\right)},y_2^{\left(n\right)},...,y_{m_n}^{\left(n\right)},...,y_{M_n}^{\left(n\right)}];$ And have

$y_{k_n}^{(n,m_n)}=\left\{\begin{array}{cc}1& \mathrm{if}(t_{k_n}^{\left(n\right)}\∈{\mathrm{Tag}}_{m_n}^{\left(n\right)})\\ 0& \mathrm{else}\end{array}\right.---\left(5\right)$

Step 5.3, formula (6) is utilized to obtain described n-th Images Classification c _nmiddle training Aerial Images I _nlabel measure oriental matrix F _n,

F _n＝Y _n[(Y _n) ^TY _n] ^-1/2(6)

Step 5.4, formula (7) is utilized to obtain described n-th Images Classification c _nmiddle training Aerial Images I _ntopological projection matrix M _n, thus obtain the topological projection matrix set M={M of the training Aerial Images set I of described N number of Images Classification C ₁, M ₂..., M _n..., M _n;

$M_n={[{\stackrel{\‾}{X}}_n(I_{M_n}-{\left(F_n\right)}^T{F}_n){\left({\stackrel{\‾}{X}}_n\right)}^T-\mathrm{\γ}{\left({\stackrel{\‾}{X}}_n\right)}^T{\stackrel{\‾}{X}}_n]}^T---\left(7\right)$

Step 6, multitask topology learning feature coding:

Step 6.1, definition iterations are t; Iterations t=1 described in initialization;

And the topological encode matrix of the t time iteration is described in during initialization the t time iteration and the n-th topological encode submatrix when having the t time iteration in each element obtained by random function;

Step 6.2, the topological encode matrix utilizing formula (8) to obtain described the t time iteration are W ^(t)code weight matrix D ^(t)the capable v column element of v thus code weight matrix D when obtaining the t time iteration ^(t):

$d_{v,v}^t=\frac{1}{2{\left|\right|w_v^{\left(t\right)}\left|\right|}_2}---\left(8\right)$

In formula (4), represent that the topological encode matrix of described the t time iteration is W ^(t)v every trade vector, described code weight matrix D ^(t)for diagonal matrix,

Step 6.3, definition tasking learning number of times are l; 1≤l≤N; And tasking learning number of times l=1 described in initialization;

Step 6.4, formula (9) is utilized to obtain the weight selection matrix of the t time iteration l subtask study in the capable v column element of v thus obtain the weight selection matrix of the t time iteration l subtask study

$d_{l,v,v}^{\left(t\right)}=\frac{1}{2{\left|\right|w_{l,v}^{(t-1)}\left|\right|}_2}---\left(9\right)$

In formula (4), l topological encode submatrix when representing the t time iteration v row vector, described weight selection matrix for diagonal matrix,

Step 6.5, from described topological characteristic projection matrix set M={M ₁, M ₂..., M _n..., M _nmiddle extraction l topological characteristic projection matrix M _l; Formula (10) is utilized to obtain the code optimization matrix of the t time iteration l subtask study

$U_l^{\left(t\right)}=M_l+\mathrm{\λ}(D_l^{\left(t\right)}+\mathrm{\μ}{D}^{\left(t\right)})---\left(10\right)$

In formula (8), λ presentation code Optimal Parameters; μ presentation code internal coordination parameter;

Step 6.6, note K _lfor l Images Classification c in described N number of Images Classification C _lthe sum of middle tag along sort; Principal component analytical method is adopted to obtain the code optimization matrix of the t time iteration l subtask study multiple eigenwert and the proper vector corresponding with multiple eigenwert, described multiple eigenwert is sorted from small to large, and selects front K _lproper vector corresponding to individual eigenwert is as the n-th topological encode submatrix during described the t time iteration

Step 6.7, judge whether l=N sets up, if set up, then complete the study of N subtask, thus obtain the topological encode matrix of the t+1 time iteration and perform step 6.8; Otherwise the value of l+1 is assigned to l, repeated execution of steps 6.4-step 6.7;

Step 6.8, judge whether t=Γ sets up, Γ is iteration total degree; If set up, perform step 6.9; Otherwise the value of t+1 is assigned to t, returns step 6.2 order and perform;

Step 6.9, utilize F-normal form to described the t+1 time iteration topological encode matrix W ^(t+1)row vector calculate, the result of calculation of acquisition sorts by descending, selects described the t+1 time iteration topological encode matrix W ^(t+1)s row vector before middle sequence, obtains described front S row vector in described the t+1 time iteration topological encode matrix W ^(t+1)in corresponding line number { h ₁, h ₂..., h _s..., h _s, h _sto represent before described sequence that in S row vector, s row vector is in described the t+1 time iteration topological encode matrix W ^(t+1)in line number; 1≤s≤S;

Step 6.10, from described topological structure tp ' ₁, tp ' ₂..., tp ' _ο..., tp ' _σin select the individual topological structure, thus the topological structure before obtaining described sequence in S row vector corresponding to s row vector; Thus the topological structure before obtaining described sequence corresponding to S row vector, and form the topological dictionary Tp={tp of described Aerial Images " ₁, tp " ₂..., tp " _{s '}..., tp " _{s '}; Tp " _{s '}represent the individual topological structure of s ' in described topological dictionary Tp; 1≤s '≤S ';

Step 7, obtain described n-th Images Classification c _nin m _nindividual training Aerial Images feature selecting vector and described target Aerial Images I ₀feature selecting vector x " ₀; Thus obtain described n-th Images Classification c _nin m _nindividual training Aerial Images feature selecting two tuple thus obtain described n-th Images Classification c _nmiddle training Aerial Images I _nfeature selecting two tuple $X_{\mathrm{FS}}^{\left(n\right)}=\{<x_1^{\&prime;\&prime;\left(n\right)},c_n>,<x_2^{\&prime;\&prime;\left(n\right)},c_n>,...,<x_{m_n}^{\&prime;\&prime;\left(n\right)},c_n>,...,<x_{M_n}^{\&prime;\&prime;\left(n\right)},c_n>\};$ Thus obtain feature selecting two tuple-set of the training Aerial Images set I of described N number of Images Classification C $X_{\mathrm{FS}}=\{X_{\mathrm{FS}}^{\left(1\right)},X_{\mathrm{FS}}^{\left(2\right)},...,X_{\mathrm{FS}}^{\left(m_n\right)},...,X_{\mathrm{FS}}^{\left(M_n\right)}\};$ Note $X_{\mathrm{FS}}=\{<x_1^{\mathrm{FS}},c_1^{\mathrm{FS}}>,<x_2^{\mathrm{FS}},c_2^{\mathrm{FS}}>,...,<x_e^{\mathrm{FS}},c_e^{\mathrm{FS}}>,...,<x_E^{\mathrm{FS}},c_E^{\mathrm{FS}}>\};$ represent the feature selecting two tuple-set X of the training Aerial Images set I of described N number of Images Classification C _fSe feature selecting two tuple, 1≤e≤E;

Step 7.1, definition select number of times to be ρ, and initialization ρ=1; 1≤ρ≤S ';

Define described the ρ time described m _nindividual training Aerial Images feature Choice Matrix be and the ρ time described m described in initialization _nindividual training Aerial Images feature Choice Matrix be for sky; Define the ρ time described m _nindividual training Aerial Images feature extraction matrix be and initialization

Step 7.2, from topological dictionary Tp, choose ρ topological structure tp " _ρ;

Step 7.3, judge described ρ topological structure tp " _ρin particle number whether be greater than described m _nindividual training Aerial Images regional distribution chart particle number, if be greater than, then perform step 7.2 after the value of ρ+1 being assigned to ρ; Otherwise, perform step 7.4;

Step 7.4, by described ρ topological structure tp " _ρwith the ρ time described m _nindividual training Aerial Images eigenmatrix in topological structure corresponding to each proper vector mate; If the match is successful, then from the ρ time described m _nindividual training Aerial Images feature extraction matrix in remove and described ρ topological structure tp " _ρproper vector corresponding to the topological structure that the match is successful, and put into the ρ time described m _nindividual training Aerial Images feature Choice Matrix be in;

Step 7.5, the value of ρ+1 is assigned to ρ, judges whether ρ=S ' sets up, if set up, adopt maximum pond strategy to described m _nindividual training Aerial Images feature Choice Matrix process, obtain described m _nindividual training Aerial Images feature selecting vector otherwise, repeated execution of steps 7.2-step 7.5;

Step 8, build topological Boosting learning model:

Step 8.1, the one-against-rest of SVM classifier is utilized to build the individual topological structure tp of s ' in described topological dictionary Tp " _{s '}disaggregated model ψ _{s '}; Thus obtain the individual disaggregated model { ψ of S ' of described topological dictionary Tp ₁, ψ ₂..., ψ _{s '}..., ψ _{s '};

The iterations of step 8.2, definition Boosting learning model is τ, and iterations τ=1 described in initialization;

Definition represent the feature selecting two tuple-set X of the training Aerial Images set I of described N number of Images Classification C during the τ time iteration _fSe feature selecting two tuple class weight; The then feature selecting two tuple-set X of the training Aerial Images set I of described N number of Images Classification C during the τ time iteration _fSclass weight be and initialization ${\mathrm{\&omega;}}_e^{\left(\mathrm{\&tau;}\right)}=\frac{1}{E};$

Step 8.3, formula (11) is utilized to obtain the disaggregated model ψ of the optimization of the τ time iteration ^(τ);

${\mathrm{\&psi;}}^{\left(\mathrm{\&tau;}\right)}=\arg {\min }_{{\mathrm{\&psi;}}^{\left(\mathrm{\&tau;}\right)}\&Element;\{{\mathrm{\&psi;}}_1,{\mathrm{\&psi;}}_2,...,{\mathrm{\&psi;}}_{s^{\&prime;}},...,{\mathrm{\&psi;}}_{S^{\&prime;}}\}}{\mathrm{\&Sigma;}}_{e=1}^E{\mathrm{\&omega;}}_e^{\left(\mathrm{\&tau;}\right)}\&times;[1-p_{\mathrm{\&tau;}}(x_e^{\mathrm{FS}}\&RightArrow;c_e^{\mathrm{FS}})]---\left(11\right)$

In formula (11), represent and adopt probability SVM method to described disaggregated model ψ ^(τ)carry out processing described e feature selecting two tuple obtained feature selecting vector belong to Images Classification probability;

Step 8.4, identification error err when utilizing formula (12) to calculate the τ time iteration _τ:

${\mathrm{err}}_{\mathrm{\&tau;}}={\mathrm{\&Sigma;}}_{e=1}^E{\mathrm{\&omega;}}_e^{\left(\mathrm{\&tau;}\right)}\&times;[1-p_{\mathrm{\&tau;}}(x_e^{\mathrm{FS}}\&RightArrow;c_e^{\mathrm{FS}})]---\left(12\right)$

Step 8.5, identification weight α when utilizing formula (13) to calculate the τ time iteration _τ:

${\mathrm{\&alpha;}}_{\mathrm{\&tau;}}=\log \frac{(1-{\mathrm{err}}_{\mathrm{\&tau;}})}{{\mathrm{err}}_{\mathrm{\&tau;}}}+\log (N-1)---\left(13\right)$

Step 8.6, e class weight described in when utilizing formula (14) to calculate the τ time iteration corresponding classification transition weight thus obtain E classification transition weight

${\mathrm{\&omega;}}_e^{\&prime;\left(\mathrm{\&tau;}\right)}={\mathrm{\&omega;}}_e^{\left(\mathrm{\&tau;}\right)}\&times;\exp \{{\mathrm{\&alpha;}}_{\mathrm{\&tau;}}\&times;[1-p_{\mathrm{\&tau;}}(x_e^{\mathrm{FS}}\&RightArrow;c_e^{\mathrm{FS}})\left]\right\}---\left(14\right)$

Step 8.7, to described E classification transition weight be normalized, obtain class weight during τ+1 iteration

Step 8.8, the value of τ+1 is assigned to τ, repeated execution of steps 8.3-step 8.8; Till τ equals iterations A, thus obtain identification weight set { α ₁, α ₂..., α _a..., α _aand the disaggregated model { ψ of optimization of described A iteration ⁽¹⁾, ψ ⁽²⁾..., ψ ^(a)..., ψ ^(A); α _arepresent that in the set of described identification weight, a identifies weight; ψ ^(a)represent the disaggregated model identifying the optimization that weight is corresponding with described a;

Step 8.9, the topological Boosting learning model shown in formula (15) is utilized to obtain described target Aerial Images I ₀images Classification Δ; Thus the quick identification of realize target Aerial Images:

$\mathrm{\&Delta;}=\arg {\min }_{\mathrm{\&Delta;}\&Element;C=\{c_1,c_2,...,c_n,...,c_N\}}{\mathrm{\&Sigma;}}_{a=1}^A{\mathrm{\&alpha;}}_a\&times;p_a(x_0^{\&prime;\&prime;}\&RightArrow;\mathrm{\&Delta;})---\left(15\right)$

In formula (15), p _a(x " ₀→ Δ) represent the disaggregated model ψ adopting probability SVM method to described a optimization ^(a)carry out processing obtained described target Aerial Images I ₀feature selecting vector x " ₀belong to the probability of Images Classification Δ.

Compared with the prior art, beneficial effect of the present invention is embodied in:

1, contemplated by the invention the topological relation of Aerial Images, have employed the thought of topological structure Classification and Identification, multi-task learning framework and Boosting learning model: first, utilize SLIC dividing method and information entropy optimization method, build regional distribution chart and the Region adjacency graph of each Aerial Images, remain the topology information in Aerial Images; Then, utilize Sub-Image Feature construction method to obtain the eigenmatrix of Aerial Images, and then obtain topological characteristic matrix and the topological characteristic projection matrix of all training Aerial Images under described each Images Classification; Subsequently, utilize multi-task learning framework to realize the topological encode square of training Aerial Images, obtain topological dictionary; Carry out feature selecting according to topological dictionary, obtain feature selecting two tuple of the feature selecting vector sum training Aerial Images of target Aerial Images; Finally, feature selecting two tuple according to the training Aerial Images selected builds topological Boosting learning model; The feature selecting vector of described target Aerial Images is input in described topological Boosting learning model, thus obtains the classification of described target Aerial Images; Compared with the method for classifying with Aerial Images in the past, the present invention effectively remains the topology information in Aerial Images; Multi-task learning framework is utilized to achieve information sharing between multitask; Improve the degree of accuracy of classification; Topology Boosting learning model, once build up, once calculates the Classification and Identification that can obtain target Aerial Images fast, improves the efficiency of Aerial Images classification, can be widely used in real life.

2, the present invention splits the zone bit of gained through having gone screening according to the principle of information entropy to Aerial Images, removes zone bit noise, the zone bit after being optimized, thus improves the efficiency of final classification.

3, the present invention utilizes the regional distribution chart replacement grid chart in the past of structure Aerial Images, the topological structure of each zone bit in Aerial Images well can be remained, overcome the shortcoming of grid chart in the past, improve the degree of accuracy of Aerial Images topology information.

4, the method utilizing Sub-Image Feature to build in the present invention obtains the eigenmatrix of subgraph, more representative, contain the information of abundant Aerial Images, the visual signature such as color, texture can be comprised, the geometric space relation of topology distribution can be comprised again, the precision of final classification can be improved.

5, the present invention utilizes based on the characteristic matching between subgraph and subgraph, extend the scope that Aerial Images topological characteristic is identified, overcome in existing method because adopting the method for random walk and tree core, and whether the Aerial Images topological relation caused can not accurately be described matching, thus the problem causing nicety of grading low.

6 compare with the Aerial Images sorting technique studied in the past, the present invention and then obtain its topological characteristic vector, and each Sub-Image Feature matrix having merged same Aerial Images, at the structural feature interpretation of different topology, improves the precision of topological features.

7, topological characteristic is realized in the classificatory projective measurement of different images according to topological characteristic vector in the present invention, and the computing method of the topological characteristic projection matrix proposed in the present invention are simple, the mathematical formulae utilized is simple and easy to ask, thus reduces the complexity of calculating.

8, the present invention realizes the coding of final topological characteristic according to topological characteristic projection matrix, solve existing method Aerial Images and can not distinguish its space layout for the image that structure is very similar in the process of classification, thus reduce the problem of the degree of accuracy of Images Classification.

9 compare with the Aerial Images sorting technique studied in the past, the present invention utilizes multi-task learning framework, to realizing sharing of information between multiple associated picture classification task, realize topological characteristic coding, make, in topological characteristic, there is identity, overcome in previous methods and for want of consider sharing of information between inter-related task, and the inefficient problem of the classification caused.

10, the present invention obtains topological dictionary by utilizing based on multitask topology learning feature coding, utilize topological dictionary to extract and there is distinctive Sub-Image Feature, topology dictionary can go out in Aerial Images to have the proper vector of the subgraph of identification, for disaggregated model of setting up below extracts effective classified information by assisting sifting.

11, the topological Boosting learning model built in the present invention, by the sorter model that integrated for multiple Weak Classifier one-tenth learns by force, has the advantages that degree of accuracy high complexity is low, improves precision and the efficiency of learning model.

12, the topological Boosting learning model built in the present invention is once build up according to training image, namely arbitrary target Aerial Images need input by each, adopt the computing formula of learning model once to calculate, just the quick identification of energy realize target Aerial Images, calculated amount is little and quick.

Accompanying drawing explanation

Fig. 1 is present system function structure chart;

Fig. 2 is the inventive method process flow diagram;

Fig. 3 a is the example images figure that takes photo by plane in the inventive method;

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

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

Fig. 4 a is the topological structure exemplary plot of Aerial Images in the inventive method;

Fig. 4 b is the area adjacency subgraph exemplary plot 1 in the inventive method with identical topological structure;

Fig. 4 c is the area adjacency subgraph exemplary plot 2 in the inventive method with identical topological structure;

Fig. 4 d is the area adjacency subgraph exemplary plot 3 in the inventive method with 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 for target Aerial Images class categories identification mission, as shown in Figure 1, based on an Aerial Images system for rapidly identifying for multitask topology learning, comprising: feature construction module, topological characteristic build module, topological characteristic projection module, multitask topology learning module, feature selection module and Classification and Identification module;

Feature construction module, according to set N number of Images Classification, obtains several tag along sorts under each Images Classification respectively, and several under obtaining each Images Classification respectively comprise the training Aerial Images of self label;

Feature construction module obtains a target Aerial Images to be identified; Aerial Images is formed by target Aerial Images and training Aerial Images; Training Aerial Images is used for modeling;

Utilize SLIC dividing method to be partitioned into multiple zone bit to each Aerial Images in Aerial Images, and form zone bit set; Utilize probability SVM classifier method to calculate the information entropy of each zone bit to zone bit set, and whether the information entropy weighing each zone bit is less than set information entropy threshold value; From zone bit set, remove the zone bit being less than information entropy threshold value, thus generate the zone bit set after optimizing; According to the positional information of each zone bit in Aerial Images in the zone bit set after optimizing, and the fillet that any two the adjacent zone bits obtained by positional information generate, build 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; Area adjacency subgraph topological structure is separately formed by the positional information of zone bit each in area adjacency subgraph and fillet; The Region adjacency graph of each Aerial Images is made up of all area adjacency subgraphs;

Sub-Image Feature construction method and core principle component analysis method several area adjacency subgraphs to each Aerial Images are utilized to process, obtain the proper vector of several area adjacency subgraphs of each Aerial Images respectively, the proper vector of each area adjacency subgraph and the topological structure one_to_one corresponding of area adjacency subgraph corresponding separately; The eigenmatrix of respective Aerial Images in Aerial Images is made up of the proper vector of several area adjacency subgraphs of each Aerial Images; The eigenmatrix of Aerial Images comprises the eigenmatrix of target Aerial Images and the eigenmatrix of training Aerial Images;

Topological characteristic builds module and the proper vector in the eigenmatrix of training Aerial Images with identical topological structure is carried out fusion treatment, obtains the topological characteristic matrix of training Aerial Images; Thus the topological characteristic matrix of all training Aerial Images under obtaining each Images Classification;

Topological characteristic projection module, according to self label of several tag along sorts and each training Aerial Images and topological characteristic matrix, obtains the topological characteristic projection matrix of the training Aerial Images under the classification of each Aerial Images;

Multitask topology learning module utilizes multi-task learning framework and principal component analytical method to carry out iteration to topological characteristic projection matrix, thus obtains the topological encode matrix of training Aerial Images; Utilize the row vector of F-normal form to topological encode matrix to calculate, the result of calculation of acquisition sorts by descending, the topological structure before selecting corresponding to S row vector, forms the topological dictionary of Aerial Images;

The feature selection module topological structure corresponding with each proper vector in the eigenmatrix of Aerial Images according to topological structure each in topological dictionary mates, if the match is successful, be then made up of the feature selecting vector of Aerial Images the proper vector that in the eigenmatrix of Aerial Images, the match is successful; The feature selecting vector of Aerial Images comprises the feature selecting vector of the feature selecting vector sum training Aerial Images of target Aerial Images; The Images Classification vectorial by the feature selecting of each training Aerial Images and training Aerial Images is corresponding separately, forms feature selecting two tuple of each training Aerial Images;

Classification and Identification module adopts SVM classifier to build multiple disaggregated model to topological dictionary; And adopt probability SVM method to obtain the multiple disaggregated models optimized according to feature selecting two tuple of training Aerial Images; Topological Boosting learning model is built by the multiple disaggregated models optimized; The feature selecting vector of target Aerial Images is input in topological Boosting learning model, thus obtains the classification of target Aerial Images.

In the present embodiment, defining N number of Images Classification is C={c ₁, c ₂..., c _n..., c _n; c _nrepresent the n-th Images Classification; Several tag along sorts are comprised in each Images Classification; Note n-th classification c _ntag along sort be k _nrepresent tag number, 1≤k _n≤ K _n; represent kth _nindividual tag along sort; K _nrepresent the n-th Images Classification c _nthe sum of middle tag along sort, K _n∈ [10,50], in concrete enforcement, K _n=10; Then the tag along sort set of N number of Images Classification C is T={T ₁, T ₂..., T _n..., T _n; Note κ is the sum of all tag along sorts in tag along sort set T, and has $\mathrm{\&kappa;}={\mathrm{\&Sigma;}}_{n=1}^N{K}_n;$

Remember the n-th Images Classification c _nin comprise self label training Aerial Images be represent the n-th Images Classification c _nin m _nindividual training Aerial Images, represent the n-th Images Classification c _nin m _nindividual training Aerial Images self tag set, 1≤m _n≤ M _n; M _nrepresent the n-th Images Classification c _nin the sum of all training Aerial Images, such as, M _n∈ [5000,10000]; Then the training Aerial Images set of N number of Images Classification C is I={I ₁, I ₂..., I _n..., I _n, namely as shown in Figure 2, have note target Aerial Images is I ₀;

As shown in Figure 2, a kind of recognition methods of the Aerial Images system for rapidly identifying based on multitask topology learning carries out as follows:

Step 1, build the n-th Images Classification c _nin m _nindividual training Aerial Images regional distribution chart thus obtain the n-th Images Classification c _nin the regional distribution chart of all training Aerial Images ${\mathrm{RAG}}_n=\{{\mathrm{RAG}}_1^{\left(n\right)},{\mathrm{RAG}}_2^{\left(n\right)},...,{\mathrm{RAG}}_{m_n}^{\left(n\right)},...,{\mathrm{RAG}}_{M_n}^{\left(n\right)}\};$ Thus obtain the areal distribution set of graphs RAG={RAG of the training Aerial Images set I of N number of Images Classification C ₁, RAG ₂..., RAG _n..., RAG _n, namely as shown in Figure 2, have and target Aerial Images I ₀regional distribution chart RAG ₀;

Step 1.1, utilize SLIC dividing method by the n-th Images Classification c _nin m _nindividual training Aerial Images be divided into individual zone bit, obtains m _nindividual training Aerial Images zone bit set represent the n-th Images Classification c _nin m _nindividual training Aerial Images ? individual zone bit; m _nindividual training Aerial Images in comprise individual zone bit positional information; Concrete SLIC dividing method can introduction in list of references " SLIC Superpixels Compared to State-of-the-artSuperpixel Methods "; SLIC dividing method is that a kind of super-pixel based on clustering algorithm is split, by LAB space and x, y pixel coordinate table totally 5 dimension spaces calculate; The method not only can Segmentation of Color Images, also can compatible segmentation gray-scale map, but also the quantity that the region needing segmentation is set that can be artificial; The advantage of SLIC dividing method is: on the one hand, raises the efficiency, and the subregion divided well remains the information of source images, and accuracy is high; On the other hand, calculated amount is little, and time cost is with super-pixel number linear increase; In this enforcement, by arranging different SLIC serious forgiveness parameters, { 0.1,0.2,0.3,0.4,0.5}, carries out repeated segmentation to Aerial Images;

Step 1.2, probability SVM classifier method is utilized to calculate m _nindividual zone bit set in individual zone bit belong to the n-th Images Classification c _nprobability thus obtain the individual zone bit belong to the probability of each classification in Images Classification C; Concrete probability SVM classifier method can introduction in list of references " Probabilistic Outputs for SupportVector Machines and Comparisons to Regularized Likelihood Method "; The support vector method of probability SVM classifier method and posterior probability is expansion and the popularization of standard SVM classifier; Be all first according to priori selected Nonlinear Mapping in advance, utilize it input vector to be mapped to a high-dimensional feature space, in this space, construct optimal hyperlane; In SVM, training sample always has clear and definite classification information, and incorrect for some uncertain problems; Probability SVM then by the inspiration of Bayes decision rule, utilizes the posterior probability of sample to represent this uncertainty, is combined by Bayes decision rule with SVM, set up posterior probability support vector machine method, in order to solve uncertain problems; The present invention utilizes the method to improve and obtains nicety of grading;

Step 1.3, formula (1) is utilized to calculate the individual zone bit information entropy thus obtain m _nindividual zone bit set in the information entropy of each zone bit:

$\mathrm{\&epsiv;}\left(r_{h_{m_n}}^{(n,m_n)}\right)=-{\mathrm{\&Sigma;}}_{c_n\&Element;C}p(r_{h_{m_n}}^{(n,m_n)}\&RightArrow;c_n)\&times;\log p(r_{h_{m_n}}^{(n,m_n)}\&RightArrow;c_n)---\left(1\right)$

Step 1.4, judge information entropy whether be less than information entropy threshold value δ, if be less than, from m _nindividual zone bit set middle removal is less than the zone bit of information entropy threshold value δ; Thus obtain m _nindividual training Aerial Images optimization after zone bit set δ ∈ [0.1,0.5]; δ can be 0.05 to regulate with step-length;

Step 1.5, with m _nindividual training Aerial Images optimization after zone bit set in the place-centric point of each zone bit as particle, adopt adjacent side to connect the particle of two adjacent zone bits any in each zone bit, thus obtain m _nindividual training Aerial Images regional distribution chart

Such as, Fig. 3 a is an Aerial Images example; It is then regional distribution chart that the zone bit set after the optimization in Fig. 3 a corresponding to Aerial Images obtains shown in Fig. 3 b;

Step 2, build the n-th Images Classification c _nin m _nindividual training Aerial Images region adjacency graph

From m _nindividual training Aerial Images regional distribution chart in exhaustive several particles comprising fillet, form m by several particles and corresponding fillet _nindividual training Aerial Images area adjacency subgraph represent the individual area adjacency subgraph, is made up of the topological structure of area adjacency subgraph the positional information of zone bit each in area adjacency subgraph and fillet; represent m _nindividual training Aerial Images the sum of area adjacency subgraph; by m _nindividual training Aerial Images area adjacency subgraph form the n-th Images Classification c _nin m _nindividual training Aerial Images region adjacency graph as shown in Figure 3 c, the multiple area adjacency subgraphs obtained by regional distribution chart in Fig. 3 b; Be only subregion shown in Fig. 3 c and adjoin subgraph example; Thus obtain the n-th Images Classification c _nin the Region adjacency graph of all training Aerial Images thus obtain the area adjacency set of graphs G={G of the training Aerial Images set I of N number of Images Classification C ₁, G ₂..., G _n..., G _n, namely as shown in Figure 2, have and target Aerial Images I ₀regional distribution chart G ₀;

Step 3, obtain the n-th Images Classification c _nin m _nindividual training Aerial Images eigenmatrix thus obtain the n-th Images Classification c _nthe eigenmatrix of training Aerial Images thus obtain the eigenmatrix set X={X of the training Aerial Images set I of N number of Images Classification C ₁, X ₂..., X _n..., X _n, namely as shown in Figure 2, have and target Aerial Images I ₀eigenmatrix X ₀;

Step 3.1, to m _nindividual Aerial Images region adjacency graph the area adjacency subgraph that middle particle number is identical with sub-Image Feature construction method is adopted to obtain Sub-Image Feature matrix respectively with concrete Sub-Image Feature construction method list of references " Probabilistic Graphlet Transfer for PhotoCropping "; Sub-Image Feature construction method utilizes classical color of image feature extracting method and texture characteristic extracting method to obtain color and the texture visual signature of each area adjacency subgraph, and extract its geometry feature, to each area adjacency subgraph, be connected these three features constitutive characteristic matrix successively; I.e. Sub-Image Feature matrix;

Step 3.2, formula (2) is utilized to obtain Sub-Image Feature matrix with golub-Werman distance $d_{\mathrm{GW}}(B_{i_{m_n}}^{(n,m_n)},B_{j_{m_n}}^{(n,m_n)}):$

$d_{\mathrm{GW}}(B_{i_{m_n}}^{(n,m_n)},B_{j_{m_n}}^{(n,m_n)})={\left|\right|{\left(B_{i_{m_n}}^{(n,m_n)}\right)}^U-{\left(B_{j_{m_n}}^{(n,m_n)}\right)}^V\left|\right|}_2---\left(2\right)$

In formula (2), with represent Sub-Image Feature matrix respectively with orthogonal basis; || || ₂represent 2-normal form;

Step 3.3, formula (3) zoning is utilized to adjoin subgraph with gaussian kernel

In formula (3), represent gaussian kernel parameter,

Step 3.4, repetition step 3.1 to step 3.3, thus obtain m _nindividual Region adjacency graph in all gaussian kernel with the area adjacency subgraph of phase homologous pints;

Step 3.5, employing core principle component analysis method (Kernel PCA method) obtain m _nindividual Aerial Images eigenmatrix $X_{m_n}^{\left(n\right)}=[x_1^{(n,m_n)},x_2^{(n,m_n)},...,x_{q_{m_n}}^{(n,m_n)},...,x_{Q_{m_n}}^{(n,m_n)}];$ be individual area adjacency subgraph proper vector, d is individual proper vector dimension; The individual area adjacency subgraph proper vector with individual area adjacency subgraph topological structure one_to_one corresponding; Concrete core principle component analysis method (KernelPCA method) can list of references " Kernel PCA for Feature Extraction and De-Noising in NonlinearRegression "; Core principle component analysis method (Kernel PCA method) a kind of nonlinear pca method (PCA method), its main thought is that input vector is mapped among a High-dimensional Linear feature space by the Nonlinear Mapping selected in advance by certain, then in High-dimensional Linear feature space, use principal component analysis (PCA) to calculate pivot composition, it have effectively can catch data nonlinear characteristic, to advantages such as the distribution situation of the data in luv space do not require; Based on this advantage, the present invention can utilize the method to become to be mutually independent or incoherent variable by variables transformations very high for script correlativity, and the proper vector mutual independence obtaining each Region adjacency graph is stronger;

Step 4, obtain the n-th Images Classification c _nin m _nindividual training Aerial Images topological characteristic vector thus obtain the n-th Images Classification c _nthe topological characteristic matrix of middle training Aerial Images thus obtain the topological characteristic set of matrices X ' of the training Aerial Images set I of N number of Images Classification C=X ' ₁, X ' ₂..., X ' _n..., X ' _n, namely as shown in Figure 2, have

Step 4.1, retain the n-th Images Classification c _nin m _nindividual training Aerial Images eigenmatrix in in topological structure corresponding to all proper vectors topological structure not identical and particle number is no more than the topological structure of upper limit ν, thus obtain the n-th Images Classification c _nin m _nindividual training Aerial Images topological structure set thus obtain the n-th Images Classification c _nthe topological structure set of training Aerial Images utilize union process namely to topological structure set { tp ₁, tp ₂..., tp _n..., tp _nin all topological structures do union operation: retain different topological structures, one only retained to identical topological structure; Multiple area adjacency subgraphs that different Aerial Images produces as shown in Fig. 4 b, 4c and 4d have identical topological structure, and identical topological structure is as shown in institute Fig. 4 a; When union process is done to this type of topological structure, only need reservation topological structure, do not need to repeat to retain; Thus obtain the topological structure set tp of the training Aerial Images set I of N number of Images Classification C;

Step 4.2, each topological structure in the topological structure set tp of the training Aerial Images set I of N number of Images Classification C to be numbered, obtain σ topological structure be designated as tp ' ₁, tp ' ₂..., tp ' _ο..., tp ' _σ; Tp ' _οrepresent ο topological structure; 1≤ο≤σ; In the present embodiment, when ν=7, σ value is about 160;

Step 4.3, by m _nindividual Aerial Images eigenmatrix in topological structure corresponding to each proper vector and ο topological structure tp ' _οcompare, select the proper vector with identical topological structure and carry out suing for peace and being averaged process; Obtain m _nindividual Aerial Images at ο topological structure tp ' _οon topology merge vector thus obtain m _nindividual Aerial Images σ topological structure tp ' ₁, tp ' ₂..., tp ' _ο..., tp ' _σon topology merge vector $\{x_{{\mathrm{tp}}_1^{\&prime;}}^{\&prime;(n,m_n)},x_{{\mathrm{tp}}_2^{\&prime;}}^{\&prime;(n,m_n)},...,x_{{\mathrm{tp}}_o^{\&prime;}}^{\&prime;(n,m_n)},...,x_{{\mathrm{tp}}_{\mathrm{\&sigma;}}^{\&prime;}}^{\&prime;(n,m_n)}\};$

Step 4.3, by m _nindividual Aerial Images σ topological structure tp ' ₁, tp ' ₂..., tp ' _ο..., tp ' _σon topology merge vector carry out end to end fusion treatment, obtain m _nindividual training Aerial Images topological characteristic vector and have d σ=d × σ;

Step 5, obtain the topological projection matrix set M={M of the training Aerial Images set I of N number of Images Classification C ₁, M ₂..., M _n..., M _n}:

Step 5.1, formula (4) is utilized to obtain the n-th Images Classification c _nmiddle training Aerial Images I _ntopological transitions matrix

${\stackrel{\&OverBar;}{X}}_n=X_n^{\&prime;}(I_{M_n}-\frac{1}{M_n}{1}_{M_n}{\left(1_{M_n}\right)}^T)---\left(4\right)$

In formula (4), for unit matrix, represent M _nindividual element is all a dimensional vector of 1,

Step 5.2, formula (5) is utilized to obtain the n-th Images Classification c _nin m _nindividual training Aerial Images kth _nindividual tag element thus obtain the n-th Images Classification c _nin m _nindividual training Aerial Images label vector $y_{m_n}^{\left(n\right)}=[y_1^{(n,m_n)},y_2^{(n,m_n)},...,y_{k_n}^{(n,m_n)},...,y_{K_n}^{(n,m_n)}],$ thus obtain the n-th Images Classification c _ntraining Aerial Images I _nlabel matrix $Y_n=[y_1^{\left(n\right)},y_2^{\left(n\right)},...,y_{m_n}^{\left(n\right)},...,y_{M_n}^{\left(n\right)}];$ And have

$y_{k_n}^{(n,m_n)}=\left\{\begin{array}{cc}1& \mathrm{if}(t_{k_n}^{\left(n\right)}\&Element;{\mathrm{Tag}}_{m_n}^{\left(n\right)})\\ 0& \mathrm{else}\end{array}\right.---\left(5\right)$

Step 5.3, formula (6) is utilized to obtain the n-th Images Classification c _nmiddle training Aerial Images I _nlabel measure oriental matrix F _n,

F _n＝Y _n[(Y _n) ^TY _n] ^-1/2(6)

Step 5.4, formula (7) is utilized to obtain the n-th Images Classification c _nmiddle training Aerial Images I _ntopological projection matrix M _n, thus obtain the topological projection matrix set M={M of the training Aerial Images set I of N number of Images Classification C ₁, M ₂..., M _n..., M _n, namely as shown in Figure 2, have

$M_n={[{\stackrel{\&OverBar;}{X}}_n(I_{M_n}-{\left(F_n\right)}^T{F}_n){\left({\stackrel{\&OverBar;}{X}}_n\right)}^T-\mathrm{\&gamma;}{\left({\stackrel{\&OverBar;}{X}}_n\right)}^T{\stackrel{\&OverBar;}{X}}_n]}^T---\left(7\right)$

Step 6, multitask topology learning feature coding:

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

And the topological encode matrix of the t time iteration is during initialization the t time iteration and the n-th topological encode submatrix when having the t time iteration in each element obtained by random function;

Step 6.2, the topological encode matrix utilizing formula (8) to obtain the t time iteration are W ^(t)code weight matrix D ^(t)the capable v column element of v thus code weight matrix D when obtaining the t time iteration ^(t):

$d_{v,v}^t=\frac{1}{2{\left|\right|w_v^{\left(t\right)}\left|\right|}_2}---\left(8\right)$

In formula (4), represent that the topological encode matrix of the t time iteration is W ^(t)v every trade vector, code weight matrix D ^(t)for diagonal matrix,

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

Step 6.4, formula (9) is utilized to obtain the weight selection matrix of the t time iteration l subtask study in the capable v column element of v thus obtain the weight selection matrix of the t time iteration l subtask study

$d_{l,v,v}^{\left(t\right)}=\frac{1}{2{\left|\right|w_{l,v}^{(t-1)}\left|\right|}_2}---\left(9\right)$

In formula (4), l topological encode submatrix when representing the t time iteration v row vector, weight selection matrix for diagonal matrix,

Step 6.5, from topological characteristic projection matrix set M={M ₁, M ₂..., M _n..., M _nmiddle extraction l topological characteristic projection matrix M _l; Formula (10) is utilized to obtain the code optimization matrix of the t time iteration l subtask study

$U_l^{\left(t\right)}=M_l+\mathrm{\&lambda;}(D_l^{\left(t\right)}+\mathrm{\&mu;}{D}^{\left(t\right)})---\left(10\right)$

In formula (8), λ presentation code Optimal Parameters, λ ∈ [0,0.1]; μ presentation code internal coordination parameter, μ ∈ [0,0.5]; Such as, λ=0.01, μ=0.3;

Step 6.6, note K _lfor l Images Classification c in N number of Images Classification C _lthe sum of middle tag along sort; Principal component analytical method (PCA method) is adopted to obtain the code optimization matrix of the t time iteration l subtask study multiple eigenwert and the proper vector corresponding with multiple eigenwert, multiple eigenwert is sorted from small to large, and selects front K _lproper vector corresponding to individual eigenwert is as the n-th topological encode submatrix during the t time iteration principal component analysis (PCA) (PCA method) is a kind of data analysing method based on second-order statistics, it changes into another given one group of correlated variables by linear transformation and organizes incoherent variable, the order arrangement that these new variablees successively decrease successively according to variance, keeps the population variance of variable constant in mathematic(al) manipulation; The main object of principal component analysis (PCA) removes with less variable most of variable of explaining in original data; This aspect utilizes the method to select than code optimization matrix middle proper vector number is few but can explain most of code optimization matrix the several new vector of middle proper vector;

Step 6.7, judge whether l=N sets up, if set up, then complete the study of N subtask, thus obtain the topological encode matrix of the t+1 time iteration and perform step 6.8; Otherwise the value of l+1 is assigned to l, repeated execution of steps 6.4-step 6.7;

Step 6.8, judge whether t=Γ sets up, Γ is iteration total degree, such as, and Γ=1000; If set up, perform step 6.9; Otherwise the value of t+1 is assigned to t, returns step 6.2 order and perform;

Step 6.9, utilize F-normal form to the t+1 time iteration topological encode matrix W ^(t+1)row vector calculate, the result of calculation of acquisition sorts by descending, selects the t+1 time iteration topological encode matrix W ^(t+1)s row vector before middle sequence, before obtaining, S row vector is the t+1 time iteration topological encode matrix W ^(t+1)in corresponding line number { h ₁, h ₂..., h _s..., h _s, h _srepresent that in front S the row vector of sequence, s row vector is the t+1 time iteration topological encode matrix W ^(t+1)in line number; 1≤s≤S; The t+1 time iteration topological encode matrix W ^(t+1)in line number with 1 start counting;

Step 6.10, from topological structure tp ' ₁, tp ' ₂..., tp ' _ο..., tp ' _σin select the individual topological structure, thus the topological structure before obtaining sequence in S row vector corresponding to s row vector; Thus the topological structure before obtaining sequence corresponding to S row vector, identical topological structure may be there is in the topological structure corresponding to a front S row vector, in this enforcement, only retain different topological structures, and form the topological dictionary Tp={tp of Aerial Images " ₁, tp " ₂..., tp " _{s '}..., tp " _{s '}; Tp " _{s '}represent the individual topological structure of s ' in topological dictionary Tp; 1≤s '≤S ';

Step 7, obtain the n-th Images Classification c _nin m _nindividual training Aerial Images feature selecting vector and target Aerial Images I ₀feature selecting vector x " ₀; Thus obtain the n-th Images Classification c _nin m _nindividual training Aerial Images feature selecting two tuple thus obtain the n-th Images Classification c _nmiddle training Aerial Images I _nfeature selecting two tuple $X_{\mathrm{FS}}^{\left(n\right)}=\{<x_1^{\&prime;\&prime;\left(n\right)},c_n>,<x_2^{\&prime;\&prime;\left(n\right)},c_n>,...,<x_{m_n}^{\&prime;\&prime;\left(n\right)},c_n>,...,<x_{M_n}^{\&prime;\&prime;\left(n\right)},c_n>\};$ Thus obtain feature selecting two tuple-set of the training Aerial Images set I of N number of Images Classification C $X_{\mathrm{FS}}=\{X_{\mathrm{FS}}^{\left(1\right)},X_{\mathrm{FS}}^{\left(2\right)},...,X_{\mathrm{FS}}^{\left(m_n\right)},...,X_{\mathrm{FS}}^{\left(M_n\right)}\};$

Note $X_{\mathrm{FS}}=\{<x_1^{\mathrm{FS}},c_1^{\mathrm{FS}}>,<x_2^{\mathrm{FS}},c_2^{\mathrm{FS}}>,...,<x_e^{\mathrm{FS}},c_e^{\mathrm{FS}}>,...,<x_E^{\mathrm{FS}},c_E^{\mathrm{FS}}>\};$ Namely as shown in Figure 2, have represent the feature selecting two tuple-set X of the training Aerial Images set I of N number of Images Classification C _fSe feature selecting two tuple, 1≤e≤E;

Step 7.1, definition select number of times to be ρ, and initialization ρ=1; 1≤ρ≤S ';

Define ρ order m _nindividual training Aerial Images feature Choice Matrix be and initialization ρ order m _nindividual training Aerial Images feature Choice Matrix be for sky; Define ρ order m _nindividual training Aerial Images feature extraction matrix be and initialization $X_{\mathrm{\&rho;}}^{(n,m_n)}=X_{m_n}^{\left(n\right)};$

Step 7.2, from topological dictionary Tp, choose ρ topological structure tp " _ρ;

Step 7.3, judge ρ topological structure tp " _ρin particle number whether be greater than m _nindividual training Aerial Images regional distribution chart particle number, if be greater than, then perform step 7.2 after the value of ρ+1 being assigned to ρ; Otherwise, perform step 7.4;

Step 7.4, by ρ topological structure tp " _ρwith ρ order m _nindividual training Aerial Images eigenmatrix in topological structure corresponding to each proper vector mate; If the match is successful, then from ρ order m _nindividual training Aerial Images feature extraction matrix in remove and ρ topological structure tp " _ρproper vector corresponding to the topological structure that the match is successful, and put into ρ order m _nindividual training Aerial Images feature Choice Matrix be in;

Step 7.5, the value of ρ+1 is assigned to ρ, judges whether ρ=S ' sets up, if set up, adopt maximum pond strategy to m _nindividual training Aerial Images feature Choice Matrix process, obtain m _nindividual training Aerial Images feature selecting vector otherwise, repeated execution of steps 7.2-step 7.5; Adopt maximum pond strategy namely by matrix often row retain maximal value element and form a new vector;

Step 8, build topological Boosting learning model:

Boosting method is a kind of method being used for improving the accuracy of weak typing algorithm, by constructing an anticipation function series, then they is combined into an anticipation function in some way; The discrimination of each single sorter is not necessarily very high, but the result after their associating has very high discrimination, and this enforcement adopts the method, can improve the discrimination of disaggregated model;

Step 8.1, the one-against-rest of SVM classifier (One-Against-All) is utilized to build the individual topological structure tp of s ' in topological dictionary Tp " _{s '}disaggregated model ψ _{s '}; Thus the individual disaggregated model { ψ of the S ' obtaining topological dictionary Tp ₁, ψ ₂..., ψ _{s '}..., ψ _{s '}; The one-against-rest of SVM classifier. one-to-many method (one-against-all), the sample successively certain classification during training is classified as a class, and other remaining samples are classified as another kind of, and the sample of such k classification has just constructed k SVM.During classification, unknown sample is categorized as that class with maximum classification function value; Utilize the method to realize the classification of Aerial Images in the present invention, the multi-class identification of Aerial Images classification can be realized;

The iterations of step 8.2, definition Boosting learning model is τ, and initialization iterations τ=1;

Definition represent the feature selecting two tuple-set X of the training Aerial Images set I of N number of Images Classification C during the τ time iteration _fSe feature selecting two tuple class weight; The then feature selecting two tuple-set X of the training Aerial Images set I of N number of Images Classification C during the τ time iteration _fSclass weight be and initialization ${\mathrm{\&omega;}}_e^{\left(\mathrm{\&tau;}\right)}=\frac{1}{E};$

Step 8.3, formula (11) is utilized to obtain the disaggregated model ψ of the optimization of the τ time iteration ^(τ);

${\mathrm{\&psi;}}^{\left(\mathrm{\&tau;}\right)}=\arg {\min }_{{\mathrm{\&psi;}}^{\left(\mathrm{\&tau;}\right)}\&Element;\{{\mathrm{\&psi;}}_1,{\mathrm{\&psi;}}_2,...,{\mathrm{\&psi;}}_{s^{\&prime;}},...,{\mathrm{\&psi;}}_{S^{\&prime;}}\}}{\mathrm{\&Sigma;}}_{e=1}^E{\mathrm{\&omega;}}_e^{\left(\mathrm{\&tau;}\right)}\&times;[1-p_{\mathrm{\&tau;}}(x_e^{\mathrm{FS}}\&RightArrow;c_e^{\mathrm{FS}})]---\left(11\right)$

In formula (11), represent and adopt probability SVM method to disaggregated model ψ ^(τ)carry out processing e feature selecting two tuple obtained feature selecting vector belong to Images Classification probability; Explaining with probability SVM classifier in step 1.2, concrete probability SVM classifier method can introduction in list of references " Probabilistic Outputs forSupport Vector Machines and Comparisons to Regularized Likelihood Method ", and the present invention utilizes the method to improve and obtains nicety of grading;

Step 8.4, identification error err when utilizing formula (12) to calculate the τ time iteration _τ:

${\mathrm{err}}_{\mathrm{\&tau;}}={\mathrm{\&Sigma;}}_{e=1}^E{\mathrm{\&omega;}}_e^{\left(\mathrm{\&tau;}\right)}\&times;[1-p_{\mathrm{\&tau;}}(x_e^{\mathrm{FS}}\&RightArrow;c_e^{\mathrm{FS}})]---\left(12\right)$

Step 8.5, identification weight α when utilizing formula (13) to calculate the τ time iteration _τ:

${\mathrm{\&alpha;}}_{\mathrm{\&tau;}}=\log \frac{(1-{\mathrm{err}}_{\mathrm{\&tau;}})}{{\mathrm{err}}_{\mathrm{\&tau;}}}+\log (N-1)---\left(13\right)$

Step 8.6, e class weight when utilizing formula (14) to calculate the τ time iteration corresponding classification transition weight thus obtain E classification transition weight

${\mathrm{\&omega;}}_e^{\&prime;\left(\mathrm{\&tau;}\right)}={\mathrm{\&omega;}}_e^{\left(\mathrm{\&tau;}\right)}\&times;\exp \{{\mathrm{\&alpha;}}_{\mathrm{\&tau;}}\&times;[1-p_{\mathrm{\&tau;}}(x_e^{\mathrm{FS}}\&RightArrow;c_e^{\mathrm{FS}})\left]\right\}---\left(14\right)$

Step 8.7, to E classification transition weight be normalized, obtain class weight during τ+1 iteration

Step 8.8, the value of τ+1 is assigned to τ, repeated execution of steps 8.3-step 8.8; Till τ equals iterations A, such as, A=1000, thus obtain identification weight set { α ₁, α ₂..., α _a..., α _aand the disaggregated model { ψ of optimization of A iteration ⁽¹⁾, ψ ⁽²⁾..., ψ ^(a)..., ψ ^(A); α _arepresent and identify that in weight set, a identifies weight; ψ ^(a)represent the disaggregated model identifying the optimization that weight is corresponding with a;

Step 8.9, the topological Boosting learning model shown in formula (15) is utilized to obtain target Aerial Images I ₀images Classification Δ; Thus the quick identification of realize target Aerial Images:

$\mathrm{\&Delta;}=\arg {\min }_{\mathrm{\&Delta;}\&Element;C=\{c_1,c_2,...,c_n,...,c_N\}}{\mathrm{\&Sigma;}}_{a=1}^A{\mathrm{\&alpha;}}_a\&times;p_a(x_0^{\&prime;\&prime;}\&RightArrow;\mathrm{\&Delta;})---\left(15\right)$

In formula (15), p _a(x " ₀→ Δ) represent the disaggregated model ψ adopting probability SVM method to optimize a ^(a)carry out processing obtained target Aerial Images I ₀feature selecting vector x " ₀belong to the probability of Images Classification Δ; Explaining with probability SVM classifier in step 1.2, concrete probability SVM classifier method sees the introduction in list of references " Probabilistic Outputs for SupportVectorMachines and Comparisons to Regularized Likelihood Method ", and the present invention utilizes the method to improve and obtains nicety of grading.

Above, be only the present invention's preferably a kind of embodiment, other researchists, completely can in other association areas according to above, and such as image-region target, image scene, picture position Classification and Identification etc., obtain identical effect.Should be noted that, be anyly familiar with those skilled in the art in the technical scope that the present invention discloses, be equal to replacement according to technical scheme of the present invention and inventive concept thereof or correlation parameter changes, all should be encompassed within protection scope of the present invention.

Claims (2)

1. based on an Aerial Images system for rapidly identifying for multitask topology learning, it is characterized in that comprising: feature construction module, topological characteristic build module, topological characteristic projection module, multitask topology learning module, feature selection module and Classification and Identification module;

Described feature construction module, according to set N number of Images Classification, obtains several tag along sorts under each Images Classification respectively, and several under obtaining each Images Classification respectively comprise the training Aerial Images of self label;

Described feature construction module obtains a target Aerial Images to be identified; Aerial Images is formed by described target Aerial Images and described training Aerial Images;

Utilize SLIC dividing method to be partitioned into multiple zone bit to each Aerial Images in described Aerial Images, and form zone bit set; Utilize probability SVM classifier method to calculate the information entropy of each zone bit to described zone bit set, and whether the information entropy weighing each zone bit is less than set information entropy threshold value; From described zone bit set, remove the zone bit being less than described information entropy threshold value, thus generate the zone bit set after optimizing; According to the positional information of each zone bit in described Aerial Images in the zone bit set after described optimization, and the fillet that any two the adjacent zone bits to be obtained by positional information generate, build the regional distribution chart of each Aerial Images in described Aerial Images;

Several area adjacency subgraphs are generated by the regional distribution chart of described each Aerial Images; Described area adjacency subgraph topological structure is separately formed by the positional information of each zone bit in described area adjacency subgraph and fillet; The Region adjacency graph of each Aerial Images is made up of all area adjacency subgraphs;

Sub-Image Feature construction method and core principle component analysis method several area adjacency subgraphs to described each Aerial Images are utilized to process, obtain the proper vector of several area adjacency subgraphs of each Aerial Images respectively, the proper vector of each area adjacency subgraph and the topological structure one_to_one corresponding of area adjacency subgraph corresponding separately; The eigenmatrix of respective Aerial Images in described Aerial Images is made up of the proper vector of several area adjacency subgraphs of described each Aerial Images; The eigenmatrix of described Aerial Images comprises the eigenmatrix of target Aerial Images and the eigenmatrix of training Aerial Images;

Described topological characteristic builds module and the proper vector in the eigenmatrix of described training Aerial Images with identical topological structure is carried out fusion treatment, obtains the topological characteristic matrix of described training Aerial Images; Thus the topological characteristic matrix of all training Aerial Images under obtaining described each Images Classification;

Described topological characteristic projection module, according to self label of several tag along sorts described and described each training Aerial Images and topological characteristic matrix, obtains the topological characteristic projection matrix of the training Aerial Images under described each Aerial Images classification;

Described multitask topology learning module utilizes multi-task learning framework and principal component analytical method to carry out iteration to described topological characteristic projection matrix, thus obtains the topological encode matrix of described training Aerial Images; Utilize the row vector of F-normal form to described topological encode matrix to calculate, the result of calculation of acquisition sorts by descending, and the topological structure before selecting corresponding to S row vector, forms the topological dictionary of described Aerial Images;

The described feature selection module topological structure corresponding with each proper vector in the eigenmatrix of described Aerial Images according to each topological structure in described topological dictionary mates, if the match is successful, be then made up of the feature selecting vector of described Aerial Images the proper vector that in the eigenmatrix of described Aerial Images, the match is successful; The feature selecting vector of described Aerial Images comprises the feature selecting vector of the feature selecting vector sum training Aerial Images of target Aerial Images; By the Images Classification that feature selecting is vectorial and described training Aerial Images is corresponding separately of each training Aerial Images, form feature selecting two tuple of each training Aerial Images;

Described Classification and Identification module adopts SVM classifier to build multiple disaggregated model to described topological dictionary; And adopt probability SVM method to obtain the multiple disaggregated models optimized according to feature selecting two tuple of described training Aerial Images; Topological Boosting learning model is built by multiple disaggregated models of described optimization; The feature selecting vector of described target Aerial Images is input in described topological Boosting learning model, thus obtains the classification of described target Aerial Images.

2. utilize a method for quickly identifying for the Aerial Images system for rapidly identifying based on multitask topology learning described in claim 1, it is characterized in that,

Defining N number of Images Classification is C={c ₁, c ₂..., c _n..., c _n; c _nrepresent the n-th Images Classification; Several tag along sorts are comprised in each Images Classification; Remember described n-th classification c _ntag along sort be k _nrepresent tag number, 1≤k _n≤ K _n; represent kth _nindividual tag along sort; K _nrepresent described n-th Images Classification c _nthe sum of middle tag along sort; Then the tag along sort set of described N number of Images Classification C is T={T ₁, T ₂..., T _n..., T _n; Note κ is the sum of all tag along sorts in described tag along sort set T, and has

Remember described n-th Images Classification c _nin comprise self label training Aerial Images be represent the n-th Images Classification c _nin m _nindividual training Aerial Images, represent described n-th Images Classification c _nin m _nindividual training Aerial Images self tag set, 1≤m _n≤ M _n; M _nrepresent described n-th Images Classification c _nin the sum of all training Aerial Images; Then the training Aerial Images set of described N number of Images Classification C is I={I ₁, I ₂..., I _n..., I _n; Note target Aerial Images is I ₀;

Described method for quickly identifying carries out as follows:

Step 1, build described n-th Images Classification c _nin m _nindividual training Aerial Images regional distribution chart thus obtain described n-th Images Classification c _nin the regional distribution chart of all training Aerial Images ${\mathrm{RAG}}_n=\{{\mathrm{RAG}}_1^{\left(n\right)},{\mathrm{RAG}}_2^{\left(n\right)},\&CenterDot;\&CenterDot;\&CenterDot;,{\mathrm{RAG}}_{m_n}^{\left(n\right)},\&CenterDot;\&CenterDot;\&CenterDot;,{\mathrm{RAG}}_{M_n}^{\left(n\right)}\};$ Thus obtain the areal distribution set of graphs RAG={RAG of the training Aerial Images set I of described N number of Images Classification C ₁, RAG ₂..., RAG _n..., RAG _n; And described target Aerial Images I ₀regional distribution chart RAG ₀;

Step 1.1, utilize SLIC dividing method by described n-th Images Classification c _nin m _nindividual training Aerial Images be divided into individual zone bit, obtains described m _nindividual training Aerial Images zone bit set represent described n-th Images Classification c _nin m _nindividual training Aerial Images ? individual zone bit; described m _nindividual training Aerial Images in comprise described individual zone bit positional information;

Step 1.2, utilize probability SVM classifier method calculate described m _nindividual zone bit set in individual zone bit belong to the n-th Images Classification c _nprobability thus obtain described individual zone bit belong to the probability of each classification in described Images Classification C;

Step 1.3, formula (1) is utilized to calculate described individual zone bit information entropy thus obtain described m _nindividual zone bit set in the information entropy of each zone bit:

$\mathrm{\&epsiv;}\left(r_{{h_m}_n}^{(n,m_n)}\right)=-{\mathrm{\&Sigma;}}_{c_n\&Element;C}p(r_{{h_m}_n}^{(n,m_n)}\&RightArrow;c_n)\&times;\log p(r_{{h_m}_n}^{(n,m_n)}\&RightArrow;c_n)---\left(1\right)$

Step 1.4, judge described information entropy whether be less than information entropy threshold value δ, if be less than, from described m _nindividual zone bit set middle removal is less than the zone bit of described information entropy threshold value δ; Thus obtain described m _nindividual training Aerial Images optimization after zone bit set

Step 1.5, with described m _nindividual training Aerial Images optimization after zone bit set in the place-centric point of each zone bit as particle, adopt adjacent side to connect the particle of two adjacent zone bits any in described each zone bit, thus obtain described m _nindividual training Aerial Images regional distribution chart

Step 2, build described n-th Images Classification c _nin m _nindividual training Aerial Images region adjacency graph

From described m _nindividual training Aerial Images regional distribution chart in exhaustive several particles comprising fillet, form described m by several particles described and corresponding fillet _nindividual training Aerial Images area adjacency subgraph represent the individual area adjacency subgraph, is made up of the topological structure of described area adjacency subgraph the positional information of each zone bit in described area adjacency subgraph and fillet; represent described m _nindividual training Aerial Images the sum of area adjacency subgraph; 1≤q _mn≤ Q _mn; By described m _nindividual training Aerial Images area adjacency subgraph form described n-th Images Classification c _nin m _nindividual training Aerial Images region adjacency graph thus obtain described n-th Images Classification c _nin the Region adjacency graph of all training Aerial Images thus obtain the area adjacency set of graphs G={G of the training Aerial Images set I of described N number of Images Classification C ₁, G ₂..., G _n..., G _n; And described target Aerial Images I ₀regional distribution chart G ₀;

Step 3, obtain described n-th Images Classification c _nin m _nindividual training Aerial Images eigenmatrix thus obtain described n-th Images Classification c _nthe eigenmatrix of training Aerial Images thus obtain the eigenmatrix set X={X of the training Aerial Images set I of described N number of Images Classification C ₁, X ₂..., X _n..., X _n; And described target Aerial Images I ₀eigenmatrix X ₀;

Step 3.1, to described m _nindividual Aerial Images region adjacency graph the area adjacency subgraph that middle particle number is identical with sub-Image Feature construction method is adopted to obtain Sub-Image Feature matrix respectively with $1\&le;i_{m_n},j_{m_n}\&le;Q_{m_n};$

Step 3.2, formula (2) is utilized to obtain described Sub-Image Feature matrix with golub-Werman distance

$d_{\mathrm{GW}}(B_{i_{m_n}}^{(n,m_n)},B_{j_{m_n}}^{(n,m_n)})={\left|\right|{\left(B_{i_{m_n}}^{(n,m_n)}\right)}^U-{\left(B_{j_{m_n}}^{(n,m_n)}\right)}^V\left|\right|}_2---\left(2\right)$

In formula (2), with represent described Sub-Image Feature matrix respectively with orthogonal basis; || || ₂represent 2-normal form;

Step 3.3, formula (3) is utilized to calculate described area adjacency subgraph with gaussian kernel

In formula (3), represent gaussian kernel parameter;

Step 3.4, repetition step 3.1 to step 3.3, thus obtain described m _nindividual Region adjacency graph in all gaussian kernel with the area adjacency subgraph of phase homologous pints;

Step 3.5, employing core principle component analysis method obtain described m _nindividual Aerial Images eigenmatrix for described individual area adjacency subgraph proper vector, d is individual proper vector dimension; Described individual area adjacency subgraph proper vector with described individual area adjacency subgraph topological structure one_to_one corresponding;

Step 4, obtain described n-th Images Classification c _nin m _nindividual training Aerial Images topological characteristic vector thus obtain described n-th Images Classification c _nthe topological characteristic matrix of middle training Aerial Images thus obtain the topological characteristic set of matrices X ' of the training Aerial Images set I of described N number of Images Classification C=X ' ₁, X ' ₂..., X ' _n..., X ' _n;

Step 4.1, retain described n-th Images Classification c _nin m _nindividual training Aerial Images eigenmatrix in in topological structure corresponding to all proper vectors topological structure not identical and particle number is no more than the topological structure of upper limit ν, thus obtain described n-th Images Classification c _nin m _nindividual training Aerial Images topological structure set thus obtain described n-th Images Classification c _nthe topological structure set of training Aerial Images utilize union process thus obtain the topological structure set tp of the training Aerial Images set I of described N number of Images Classification C;

Step 4.2, each topological structure in the topological structure set tp of the training Aerial Images set I of described N number of Images Classification C to be numbered, obtain σ topological structure be designated as tp ' ₁, tp ' ₂..., tp ' _ο..., tp ' _σ; Tp ' _οrepresent ο topological structure; 1≤ο≤σ;

Step 4.3, by described m _nindividual Aerial Images eigenmatrix in topological structure corresponding to each proper vector and described ο topological structure tp ' _οcompare, select the proper vector with identical topological structure and carry out suing for peace and being averaged process; Obtain described m _nindividual Aerial Images at ο topological structure tp ' _οon topology merge vector thus obtain described m _nindividual Aerial Images a described σ topological structure tp ' ₁, tp ' ₂..., tp ' _ο..., tp ' _σon topology merge vector

Step 4.3, by described m _nindividual Aerial Images a described σ topological structure tp ' ₁, tp ' ₂..., tp ' _ο..., tp ' _σon topology merge vector carry out end to end fusion treatment, obtain described m _nindividual training Aerial Images topological characteristic vector and have d σ=d × σ;

Step 5, obtain the topological projection matrix set M={M of the training Aerial Images set I of described N number of Images Classification C ₁, M ₂..., M _n..., M _n}:

Step 5.1, formula (4) is utilized to obtain described n-th Images Classification c _nmiddle training Aerial Images I _ntopological transitions matrix

${\stackrel{\&OverBar;}{X}}_n=X_n^{\&prime;}(I_{M_n}-\frac{1}{M_n}{1}_{M_n}{\left(1_{M_n}\right)}^T)---\left(4\right)$

In formula (4), for unit matrix, represent M _nindividual element is all a dimensional vector of 1,

Step 5.2, formula (5) is utilized to obtain described n-th Images Classification c _nin m _nindividual training Aerial Images kth _nindividual tag element thus obtain described n-th Images Classification c _nin m _nindividual training Aerial Images label vector $y_{m_n}^{\left(n\right)}=[y_1^{(n,m_n)},y_2^{(n,m_n)},\&CenterDot;\&CenterDot;\&CenterDot;,y_{k_n}^{(n,m_n)},\&CenterDot;\&CenterDot;\&CenterDot;,y_{K_n}^{(n,m_n)}],$ thus obtain described n-th Images Classification c _ntraining Aerial Images I _nlabel matrix $Y_n=[y_1^{\left(n\right)},y_2^{\left(n\right)},\&CenterDot;\&CenterDot;\&CenterDot;,y_{m_n}^{\left(n\right)},\&CenterDot;\&CenterDot;\&CenterDot;,y_{M_n}^{\left(n\right)}];$ And have

$y_{k_n}^{(n,m_n)}=\left\{\begin{array}{cc}1& \mathrm{if}(t_{k_n}^{\left(n\right)}\&Element;{\mathrm{Tag}}_{m_n}^{\left(n\right)})\\ 0& \mathrm{else}\end{array}\right.---\left(5\right)$

Step 5.3, formula (6) is utilized to obtain described n-th Images Classification c _nmiddle training Aerial Images I _nlabel measure oriental matrix F _n,

F _n＝Y _n[(Y _n) ^TY _n] ^-1/2(6)

Step 5.4, formula (7) is utilized to obtain described n-th Images Classification c _nmiddle training Aerial Images I _ntopological projection matrix M _n, thus obtain the topological projection matrix set M={M of the training Aerial Images set I of described N number of Images Classification C ₁, M ₂..., M _n..., M _n;

$M_n={[{\stackrel{\&OverBar;}{X}}_n(I_{M_n}-{\left(F_n\right)}^T{F}_n){\left({\stackrel{\&OverBar;}{X}}_n\right)}^T-\mathrm{\&gamma;}{\left({\stackrel{\&OverBar;}{X}}_n\right)}^T{\stackrel{\&OverBar;}{X}}_n]}^T---\left(7\right)$

Step 6, multitask topology learning feature coding:

Step 6.1, definition iterations are t; Iterations t=1 described in initialization;

And the topological encode matrix of the t time iteration is described in during initialization the t time iteration and the n-th topological encode submatrix when having the t time iteration in each element obtained by random function;

Step 6.2, the topological encode matrix utilizing formula (8) to obtain described the t time iteration are W ^(t)code weight matrix D ^(t)the capable v column element of v thus code weight matrix D when obtaining the t time iteration ^(t):

$d_{v,v}^{\left(t\right)}=\frac{1}{2{\left|\right|w_v^{\left(t\right)}\left|\right|}_2}---\left(8\right)$

In formula (4), represent that the topological encode matrix of described the t time iteration is W ^(t)v every trade vector, described code weight matrix D ^(t)for diagonal matrix,

Step 6.3, definition tasking learning number of times are l; 1≤l≤N; And tasking learning number of times l=1 described in initialization;

Step 6.4, formula (9) is utilized to obtain the weight selection matrix of the t time iteration l subtask study in the capable v column element of v thus obtain the weight selection matrix of the t time iteration l subtask study

$d_{l,v,v}^{\left(t\right)}=\frac{1}{2{\left|\right|w_{l,v}^{(t-1)}\left|\right|}_2}---\left(9\right)$

In formula (4), l topological encode submatrix when representing the t time iteration v row vector, described weight selection matrix for diagonal matrix,

Step 6.5, from described topological characteristic projection matrix set M={M ₁, M ₂..., M _n..., M _nmiddle extraction l topological characteristic projection matrix M _l; Formula (10) is utilized to obtain the code optimization matrix of the t time iteration l subtask study

$U_l^{\left(t\right)}=M_l+\mathrm{\&lambda;}(D_l^{\left(t\right)}+\mathrm{\&mu;}{D}^{\left(t\right)})---\left(10\right)$

In formula (8), λ presentation code Optimal Parameters; μ presentation code internal coordination parameter;

Step 6.6, note K _lfor l Images Classification c in described N number of Images Classification C _lthe sum of middle tag along sort; Principal component analytical method is adopted to obtain the code optimization matrix of the t time iteration l subtask study multiple eigenwert and the proper vector corresponding with multiple eigenwert, described multiple eigenwert is sorted from small to large, and selects front K _lproper vector corresponding to individual eigenwert is as the n-th topological encode submatrix during described the t time iteration

Step 6.7, judge whether l=N sets up, if set up, then complete the study of N subtask, thus obtain the topological encode matrix of the t+1 time iteration and perform step 6.8; Otherwise the value of l+1 is assigned to l, repeated execution of steps 6.4-step 6.7;

Step 6.8, judge whether t=Γ sets up, Γ is iteration total degree; If set up, perform step 6.9; Otherwise the value of t+1 is assigned to t, returns step 6.2 order and perform;

Step 6.9, utilize F-normal form to described the t+1 time iteration topological encode matrix W ^(t+1)row vector calculate, the result of calculation of acquisition sorts by descending, selects described the t+1 time iteration topological encode matrix W ^(t+1)s row vector before middle sequence, obtains described front S row vector in described the t+1 time iteration topological encode matrix W ^(t+1)in corresponding line number { h ₁, h ₂..., h _s..., h _s, h _sto represent before described sequence that in S row vector, s row vector is in described the t+1 time iteration topological encode matrix W ^(t+1)in line number; 1≤s≤S;

Step 6.10, from described topological structure tp ' ₁, tp ' ₂..., tp ' _ο..., tp ' _σin select the individual topological structure, thus the topological structure before obtaining described sequence in S row vector corresponding to s row vector; Thus the topological structure before obtaining described sequence corresponding to S row vector, and form the topological dictionary Tp={tp of described Aerial Images " ₁, tp " ₂..., tp " _{s '}..., tp " _{s '}; Tp " _{s '}represent the individual topological structure of s ' in described topological dictionary Tp; 1≤s '≤S ';

Step 7, obtain described n-th Images Classification c _nin m _nindividual training Aerial Images feature selecting vector and described target Aerial Images I ₀feature selecting vector x " ₀; Thus obtain described n-th Images Classification c _nin m _nindividual training Aerial Images feature selecting two tuple thus obtain described n-th Images Classification c _nmiddle training Aerial Images I _nfeature selecting two tuple $X_{\mathrm{FS}}^{\left(n\right)}=\{<x_1^{\&prime;\&prime;\left(n\right)},c_n>,<x_2^{\&prime;\&prime;\left(n\right)},c_n>,\&CenterDot;\&CenterDot;\&CenterDot;,<x_{m_n}^{\&prime;\&prime;\left(n\right)},c_n>,\&CenterDot;\&CenterDot;\&CenterDot;,<x_{M_n}^{\&prime;\&prime;\left(n\right)},c_n>\};$ Thus obtain feature selecting two tuple-set of the training Aerial Images set I of described N number of Images Classification C $X_{\mathrm{FS}}=\{X_{\mathrm{FS}}^{\left(1\right)},X_{\mathrm{FS}}^{\left(2\right)},\&CenterDot;\&CenterDot;\&CenterDot;,X_{\mathrm{FS}}^{\left(m_n\right)},\&CenterDot;\&CenterDot;\&CenterDot;,X_{\mathrm{FS}}^{\left(M_n\right)}\};$ Note $X_{\mathrm{FS}}=\{<x_1^{\mathrm{FS}},c_1^{\mathrm{FS}}>,<x_2^{\mathrm{FS}},c_2^{\mathrm{FS}}>,\&CenterDot;\&CenterDot;\&CenterDot;,<x_e^{\mathrm{FS}},c_e^{\mathrm{FS}}>,\&CenterDot;\&CenterDot;\&CenterDot;,<x_E^{\mathrm{FS}},c_E^{\mathrm{FS}}>\};$ represent the feature selecting two tuple-set X of the training Aerial Images set I of described N number of Images Classification C _fSe feature selecting two tuple, 1≤e≤E;

Step 7.1, definition select number of times to be ρ, and initialization ρ=1; 1≤ρ≤S ';

Define described the ρ time described m _nindividual training Aerial Images feature Choice Matrix be and the ρ time described m described in initialization _nindividual training Aerial Images feature Choice Matrix be for sky; Define the ρ time described m _nindividual training Aerial Images feature extraction matrix be and initialization

Step 7.2, from topological dictionary Tp, choose ρ topological structure tp " _ρ;

Step 7.3, judge described ρ topological structure tp " _ρin particle number whether be greater than described m _nindividual training Aerial Images regional distribution chart particle number, if be greater than, then perform step 7.2 after the value of ρ+1 being assigned to ρ; Otherwise, perform step 7.4;

Step 7.4, by described ρ topological structure tp " _ρwith the ρ time described m _nindividual training Aerial Images eigenmatrix in topological structure corresponding to each proper vector mate; If the match is successful, then from the ρ time described m _nindividual training Aerial Images feature extraction matrix in remove and described ρ topological structure tp " _ρproper vector corresponding to the topological structure that the match is successful, and put into the ρ time described m _nindividual training Aerial Images feature Choice Matrix be in;

Step 7.5, the value of ρ+1 is assigned to ρ, judges whether ρ=S ' sets up, if set up, adopt maximum pond strategy to described m _nindividual training Aerial Images feature Choice Matrix process, obtain described m _nindividual training Aerial Images feature selecting vector otherwise, repeated execution of steps 7.2-step 7.5;

Step 8, build topological Boosting learning model:

Step 8.1, the one-against-rest of SVM classifier is utilized to build the individual topological structure tp of s ' in described topological dictionary Tp " _{s '}disaggregated model ψ _{s '}; Thus obtain the individual disaggregated model { ψ of S ' of described topological dictionary Tp ₁, ψ ₂..., ψ _{s '}..., ψ _{s '};

The iterations of step 8.2, definition Boosting learning model is τ, and iterations τ=1 described in initialization;

Definition represent the feature selecting two tuple-set X of the training Aerial Images set I of described N number of Images Classification C during the τ time iteration _fSe feature selecting two tuple class weight; The then feature selecting two tuple-set X of the training Aerial Images set I of described N number of Images Classification C during the τ time iteration _fSclass weight be and initialization

Step 8.3, formula (11) is utilized to obtain the disaggregated model ψ of the optimization of the τ time iteration ^(τ);

${\mathrm{\&psi;}}^{\left(\mathrm{\&tau;}\right)}=\arg {\min }_{{\mathrm{\&psi;}}^{\left(\mathrm{\&tau;}\right)}\&Element;\{{\mathrm{\&psi;}}_1,{\mathrm{\&psi;}}_2,...,{\mathrm{\&psi;}}_{s^{\&prime;}},...,{\mathrm{\&psi;}}_{S^{\&prime;}}\}}{\mathrm{\&Sigma;}}_{e=1}^E{\mathrm{\&omega;}}_e^{\left(\mathrm{\&tau;}\right)}\&times;[1-p_{\mathrm{\&tau;}}(x_e^{\mathrm{FS}}\&RightArrow;c_e^{\mathrm{FS}})]---\left(11\right)$

In formula (11), represent and adopt probability SVM method to described disaggregated model ψ ^(τ)carry out processing described e feature selecting two tuple obtained feature selecting vector belong to Images Classification probability;

Step 8.4, identification error err when utilizing formula (12) to calculate the τ time iteration _τ:

${\mathrm{err}}_{\mathrm{\&tau;}}={\mathrm{\&Sigma;}}_{e=1}^E{\mathrm{\&omega;}}_e^{\left(\mathrm{\&tau;}\right)}\&times;[1-p_{\mathrm{\&tau;}}(x_e^{\mathrm{FS}}\&RightArrow;c_e^{\mathrm{FS}})]---\left(12\right)$

Step 8.5, identification weight α when utilizing formula (13) to calculate the τ time iteration _τ:

${\mathrm{\&alpha;}}_{\mathrm{\&tau;}}=\log \frac{(1-{\mathrm{err}}_{\mathrm{\&tau;}})}{{\mathrm{err}}_{\mathrm{\&tau;}}}+\log (N-1)---\left(13\right)$

Step 8.6, e class weight described in when utilizing formula (14) to calculate the τ time iteration corresponding classification transition weight thus obtain E classification transition weight

${\mathrm{\&omega;}}_e^{\&prime;\left(\mathrm{\&tau;}\right)}={\mathrm{\&omega;}}_e^{\left(\mathrm{\&tau;}\right)}\&times;\exp \{{\mathrm{\&alpha;}}_{\mathrm{\&tau;}}\&times;[1-p_{\mathrm{\&tau;}}(x_e^{\mathrm{FS}}\&RightArrow;c_e^{\mathrm{FS}})\left]\right\}---\left(14\right)$

Step 8.7, to described E classification transition weight be normalized, obtain class weight during τ+1 iteration

Step 8.8, the value of τ+1 is assigned to τ, repeated execution of steps 8.3-step 8.8; Till τ equals iterations A, thus obtain identification weight set { α ₁, α ₂..., α _a..., α _aand the disaggregated model { ψ of optimization of described A iteration ⁽¹⁾, ψ ⁽²⁾..., ψ ^(a)..., ψ ^(A); α _arepresent that in the set of described identification weight, a identifies weight; ψ ^(a)represent the disaggregated model identifying the optimization that weight is corresponding with described a;

Step 8.9, the topological Boosting learning model shown in formula (15) is utilized to obtain described target Aerial Images I ₀images Classification Δ; Thus the quick identification of realize target Aerial Images:

$\mathrm{\&Delta;}=\arg {\min }_{\mathrm{\&Delta;}\&Element;C=\{c_1,c_2,\&CenterDot;\&CenterDot;\&CenterDot;,c_n,\&CenterDot;\&CenterDot;\&CenterDot;,c_N\}}{\mathrm{\&Sigma;}}_{a=1}^A{\mathrm{\&alpha;}}_a\&times;p_a(x_0^{\&prime;\&prime;}\&RightArrow;\mathrm{\&Delta;})---\left(15\right)$

In formula (15), p _a(x " ₀→ Δ) represent the disaggregated model ψ adopting probability SVM method to described a optimization ^(a)carry out processing obtained described target Aerial Images I ₀feature selecting vector x ₀" belong to the probability of Images Classification Δ.