CN105550712A - Optimized convolution automatic encoding network-based auroral image sorting method - Google Patents

Abstract

The invention discloses an optimized convolution automatic encoding network-based auroral image sorting method, and mainly aims at solving the problem that the existing technology is relatively low in auroral image sorting accuracy. The method comprises the following realization steps: 1, solving a saliency map of auroral images and extracting training samples on the basis of the saliency map; 2, carrying out pre-whitening on the training samples; 3, training an automatic encoding network AE; 4, solving the convolution self-encoding characteristics of the auroral images by utilizing the trained automatic encoding network; 5, carrying out mean pooling on the convolution self-encoding characteristics of the auroral images; and 6, inputting the pooled convolution self-encoding characteristics into a softmax sorter so as to sort the auroral images.

Description

Based on the aurora image classification method optimizing convolution automatic coding network

Technical field

The invention belongs to technical field of image processing, further relate to the sorting technique of aurora image, can be used for scene classification and the target identification of image.

Background technology

Aurora are various Magnetic storm process ionosphere traces the most intuitively, the all-sky imaging system All-skyCamera of Chinese Arctic Yellow River Station is simultaneously to three typical spectral coverage 427.8nm of aurora, 557.7nm and 630.0nm carries out Continuous Observation, produce ten hundreds of aurora images, data volume is huge.Rationally effective aurora Images Classification to study of various aurora phenomenon and and Magnetic storm process between relation particularly important.

Early stage aurora sort research is based on visual inspection, and manual realization marks and classification work, but aurora image is millions of every year, and the mode of manually carrying out key words sorting no longer meets requirement large-scale data being carried out to objective classification.Until 2004 document " m.T., andDonovanE.F., Diurnalauroraloccurrencestatisticsobtainedviamachinevisi on.AnnalesGeophysicae, 22 (4): 1103-1113,2004. " just image processing techniques is incorporated into aurora classification of images in.The people such as Wang in 2007 at article " WangQian, LiangJimin, HuZeJun, HuHaiHong, ZhaoHeng, HuHongQiao, GaoXinbo, YangHuigen.Spatialtexturebasedautomaticclassificationofd aysideaurorainall-skyimages.JournalofAtmosphericandSolar-TerrestrialPhysics, 2010, 72 (5): 498 – 508. " in use the gray feature of principal component analysis (PCA) PCA to aurora image to extract, propose a kind of aurora sorting technique based on presentation, certain progress is achieved in Coronal aurorae sort research direction, 2008, the people such as Gao publish an article " L.Gao, X.B.Gao, andJ.M.Liang.Daysidecoronaautoradetectionbasedonsamplese lectionandadaBoostalgorithm.J.ImageGraph, 2010,15 (1): 116-121. ", aurora image classification method based on Gabor transformation is proposed, have employed local Gabor filter and extract characteristics of image, reducing feature redundant information when guaranteeing computational accuracy, achieving good classifying quality, 2009, morphology constituent analysis (MCA) combines with aurora image procossing by the people such as Fu in article " FuRu, JieLiandX.B.Gao..Automaticauroraimagesclassificationalgo rithmbasedonseparatedtexture.Proc.Int.Conf.RoboticsandBi omimetics, 2009:1331-1335. ", feature is extracted from the aurora texture subgraph obtained after MCA is separated, for the classification of arc crown two class aurora image, improve the accuracy of arc crown aurora classification.Follow-up correlative study also has: the people such as Han propose again the aurora classification based on the classification of BIFs characteristic sum C average in article " BingHan; XiaojingZhao; DachengTao, etal.DaysideauroraclassificationviaBIFs-basedsparserepre sentationusingmanifoldlearning.InternationalJournalofCom puterMathematics.Publishedonline:12Nov2013. "; The people such as Yang propose multi-level Wavelet Transform conversion and represent aurora characteristics of image in article " YangXi; LiJie; HanBing; GaoXinbo.Wavelethierarchicalmodelforauroraimagesclassifi cation.JournalofXidianUniversity; 2013; 40 (2): 18-24. ", achieve higher classification accuracy; 2013, the people such as Han introduce implicit Dirichlet distribute model LDA in article " HanB; YangC; GaoXB.AuroraimageclassificationbasedonLDAcombiningwithsa liencyinformation.RuanJianXueBao/JournalofSoftware; 2013; 24 (11): 2758-2766. ", and combining image conspicuousness information, further increase again the classification accuracy of aurora image.

But existing aurora image processing algorithm is all based on shallow-layer feature, and characteristic present ability and classification accuracy are all greatly limited.Article " A.Krizhevsky; I.Sutskever; andG.Hinton.ImageNetclassificationwithdeepconvolutionaln euralnetworks.InNIPS; 2012. " proposes convolutional neural networks, its outstanding image characteristics extraction ability is persistently overheating in academia, and its great potential be applied in aurora image characteristics extraction is worth further investigation.

But, still there is following problem in feature extraction degree of depth convolutional network being directly used in aurora image: be first divide owing to there is many complete full blackboards without any information in aurora image, existing degree of depth learning algorithm does not have disposal route for this part redundancy; Secondly due to number of training restriction, existing degree of depth convolutional network technology is not high to the classification accuracy of aurora image; 3rd, degree of depth convolutional network time consumption for training is serious.

Summary of the invention

The object of the invention is to the deficiency existed for above-mentioned prior art, proposing a kind of aurora image classification method based on optimizing convolution automatic coding network, to complete network training fast, improve classification accuracy rate.

The technical scheme realizing above-mentioned purpose of the present invention is: carry out significance analysis to aurora image, significantly scheme to extract the training sample for training automatic coding network A E based on aurora, then the convolution own coding feature of aurora image is extracted with the automatic coding network characterization trained, and utilize average pond about to subtract convolution own coding feature, realize the classification to aurora image finally by softmax sorter.Implementation step comprises as follows:

(1) input aurora image, extract totally 100000 training block of pixels according to aurora image saliency map, composition training block of pixels collection P _{8 × 8 × 100000};

(2) to training block of pixels collection P _{8 × 8 × 100000}carry out whitening pretreatment, obtain the training sample set x after albefaction _pCAwhite;

(3) the training sample set x after albefaction is utilized _pCAwhite, training automatic coding network A E:

3a) training sample set is expressed as x _pCAwhite={ xp ⁽¹⁾, xp ⁽²⁾, xp ⁽³⁾..., xp ⁽ⁱ⁾..., xp ^(m), wherein xp ⁽ⁱ⁾represent i-th training sample, xp ⁽ⁱ⁾∈ R ⁶⁴, i=1,2 ..., m, m represent number of training; According to training sample xp ⁽ⁱ⁾ask an automatic coding network A E hidden layer jth neuronic average active degree:

$\widehat{{\mathrm{\ρ}}_j}=\frac{1}{m}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\[a_{W,b}\left({\mathrm{xp}}^{\left(i\right)}\right)\],$

Wherein, j=1,2 ..., n, n represent node in hidden layer, a _w,b(xp ⁽ⁱ⁾) represent when being input as xp ⁽ⁱ⁾time an automatic coding network A E hidden layer jth neuronic activity, (W, b)=(W ⁽¹⁾, b ⁽¹⁾, W ⁽²⁾, b ⁽²⁾) represent the parameter of automatic coding network A E, wherein, W ⁽¹⁾represent the weight of input layer to hidden layer node, W ⁽²⁾represent the weight of hidden layer node to output layer node, b ⁽¹⁾represent input layer being biased to hidden layer node, b ⁽²⁾represent hidden layer node being biased to output layer node;

3b) according to backpropagation BP coaching method principle, with parameter (W, b) and the hidden layer average active degree of automatic coding network A E construct a sparse cost function J _sparse(W, b):

$\begin{array}{c}{J}_{sparse}\left(W,b\right)=\[\frac{1}{m}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\left(\frac{1}{2}\left|\right|h_{W,b}\left({\mathrm{xp}}^{\left(i\right)}\right)-{\mathrm{xp}}^{\left(i\right)}|{|}^2\right)\]\\ +\frac{{\mathrm{\λ}}_1}{2}\mathrm{\Σ}{\left(W^{\left(1\right)}\right)}^2+\frac{{\mathrm{\λ}}_2}{2}\mathrm{\Σ}{\left(W^{\left(2\right)}\right)}^2+\mathrm{\β}\underset{l=1}{\overset{n}{\mathrm{\Σ}}}KL\left(\mathrm{\ρ}|\widehat{{\mathrm{\ρ}}_j}\right)\end{array}$

In formula, h _w,b() represents nonlinear automatic coding network A E function, represent with ρ be average and with for average two Bernoulli random variables between relative entropy, λ ₁and λ ₂represent the weight attenuation parameter of hidden layer and output layer respectively, ρ represents degree of rarefication coefficient, and its value is a constant being less than 0.1;

3c) by minimizing cost function J _sparse(W, b) be optimized after the parameter (W of automatic coding network A E _opt, b _opt):

$(W_{opt},b_{opt})=\underset{W,b}{min}{J}_{sparse}(W,b)=({W_{opt}}_{\left(jq\right)}^{\left(1\right)},{b_{opt}}_{\left(j\right)}^{\left(1\right)},{W_{opt}}_{\left(kj\right)}^{\left(2\right)},{b_{opt}}_{\left(k\right)}^{\left(2\right)})$

Wherein, represent the weight of q node to a jth node of hidden layer of input layer after optimizing, represent the weight of a hidden layer jth node to hidden layer to an output layer kth node after optimizing, represent input layer being biased to a hidden layer jth node after optimizing, represent hidden layer node being biased to a kth node after optimizing, q=k=1,2 ..., 64,64 represent input layer numbers, and output layer nodes equals input layer number, j=1,2 ..., n, n represent node in hidden layer;

(4) by the weight of q the node to a jth node of hidden layer of optimizing rear input layer ask the convolution own coding feature Fr of aurora image I;

(5) pondization that is averaged by the convolution own coding feature Fr of aurora image I operates, the block of 11 × 11 sizes is on average divided into by convolution own coding feature Fr, every block is all merged into a mean value, then these mean values are reconfigured and obtain pond feature F ∈ R ^{11 × 11 × n};

(6) the pond feature F of aurora image is input to softmax sorter, obtains a class label, be the classification of this aurora image.

The present invention compared with prior art tool has the following advantages:

The first, the present invention utilizes image saliency map to carry out the preferred of training sample, effectively eliminates invalid training sample, improves network training efficiency, improves the classification accuracy of model to aurora image simultaneously;

The second, automatic coding network A E pre-training convolution filter is selected in invention, builds convolutional network, effectively overcomes the lower problem of classification accuracy that aurora image lack of training samples causes.

Accompanying drawing explanation

Fig. 1 is realization flow figure of the present invention;

Fig. 2 is that the present invention is to aurora image saliency map, significantly figure binaryzation and extraction training fragment result figure;

Fig. 3 is the part convolution filter that the present invention is obtained by training automatic coding network A E;

Fig. 4 is corresponding classification accuracy and classification time comparison diagram when automatic coding network A E node in hidden layer is different in the present invention;

Fig. 5 is that automatic coding network A E hidden layer degree of rarefication of the present invention is to the effect diagram of classification accuracy.

Embodiment

Below in conjunction with accompanying drawing, performing step of the present invention and technique effect are described in further detail.

With reference to Fig. 1, performing step of the present invention is as follows:

Step 1, input aurora image, extracts training block of pixels collection P _{8 × 8 × 100000}.

1.1) the aurora image of a width as shown in Fig. 2 (a) is inputted, obtain each pixel I (x in this image, y), brightness L (x, y), Gradient Features H (x, and edge binaryzation feature B (x y), y), and by these three kinds of Fusion Features, aurora image slices vegetarian refreshments I (x is obtained, y) conspicuousness value of information S (x, y):

S(x,y)＝L(x,y)+H(x,y)+B(x,y)；

By aurora image the aurora image saliency map S of conspicuousness value of information S (x, y) composition as shown in Fig. 2 (b) a little;

1.2) carry out binaryzation operation to image saliency map S, the binaryzation obtained as Fig. 2 (c) significantly schemes S ₁;

1.3) significantly S is schemed in binaryzation at random ₁the training block of pixels ps of upper extraction 8 × 8 size _{8 × 8}, judge the value of this block of pixels: if block of pixels ps _{8 × 8}in 1 value proportion be greater than 0.8, then extract the block of pixels p of original image I in this position _{8 × 8}; If block of pixels ps _{8 × 8}in 1 value proportion be less than or equal to 0.8, then do not deal with;

1.4) according to 1.3) method, extract aurora images totally 100000 training block of pixels, composition training block of pixels collection P _{8 × 8 × 100000}.

Step 2, to training block of pixels collection P _{8 × 8 × 100000}carry out whitening pretreatment, ask for the training sample set x after albefaction _pCAwhite.

Whitening pretreatment technology comprises: PCA albefaction, ZCA albefaction and spectral whitening etc.What this example adopted is ZCA whitening approach, and its concrete operation step is described below:

2.1) block of pixels collection P will be trained _{8 × 8 × 100000}carry out matrix deformation, obtain deformation matrix: x ∈ R ^{64 × 100000};

2.2) covariance matrix of x is asked:

$\mathrm{\φ}=\frac{1}{m}{\mathrm{\Σ}}_{i=1}^m\left(x^{\left(i\right)}\right){\left(x^{\left(i\right)}\right)}^T,$

Wherein, m=100000 represents number of training, x ⁽ⁱ⁾i-th row of representing matrix x;

2.3) SVD decomposition is carried out to deformation matrix x: x=U φ V, obtain left basis matrix U and right basis matrix V, and x is expressed as in left basis matrix U direction: x _rot=U ^tx;

2.4) according to 2.2) and 2.3) obtain training sample set x _pw:

$x_{Pw}=U\frac{x_{rot}}{\sqrt{\mathrm{\φ}+\mathrm{\ϵ}}}$

Wherein, ε represent one be not 0 minimum number, its value is 10 ^-5.

Step 3, utilizes training sample set x _pw, training autocoder.

3.1) training sample set is expressed as x _pw={ xp ⁽¹⁾, xp ⁽²⁾, xp ⁽³⁾..., xp ⁽ⁱ⁾..., xp ^(m), wherein xp ⁽ⁱ⁾represent i-th training sample, xp ⁽ⁱ⁾∈ R ⁶⁴, m represents number of training; According to training sample xp ⁽ⁱ⁾ask an automatic coding network A E hidden layer jth neuronic average active degree

$\widehat{{\mathrm{\ρ}}_j}=\frac{1}{m}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\[a_{W,b}\left({\mathrm{xp}}^{\left(i\right)}\right)\],$

Wherein, j=1,2 ..., n, n represent node in hidden layer, a _w,b(xp ⁽ⁱ⁾) represent when being input as xp ⁽ⁱ⁾time an automatic coding network A E hidden layer jth neuronic activity, (W, b)=(W ⁽¹⁾, b ⁽¹⁾, W ⁽²⁾, b ⁽²⁾) represent the parameter of automatic coding network A E, wherein, W ⁽¹⁾represent the weight of input layer to hidden layer node, W ⁽²⁾represent the weight of hidden layer node to output layer node, b ⁽¹⁾represent input layer being biased to hidden layer node, b ⁽²⁾represent hidden layer node being biased to output layer node;

3.2) according to backpropagation BP coaching method principle, with parameter (W, b) and the hidden layer average active degree of automatic coding network A E construct a sparse cost function J _sparse(W, b):

$\begin{array}{c}{J}_{sparse}\left(W,b\right)=\[\frac{1}{m}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\left(\frac{1}{2}\left|\right|h_{W,b}\left({\mathrm{xp}}^{\left(i\right)}\right)-{\mathrm{xp}}^{\left(i\right)}|{|}^2\right)\]\\ +\frac{{\mathrm{\λ}}_1}{2}\mathrm{\Σ}{\left(W^{\left(1\right)}\right)}^2+\frac{{\mathrm{\λ}}_2}{2}\mathrm{\Σ}{\left(W^{\left(2\right)}\right)}^2+\mathrm{\β}\underset{l=1}{\overset{n}{\mathrm{\Σ}}}KL\left(\mathrm{\ρ}|\widehat{{\mathrm{\ρ}}_j}\right)\end{array}$

In formula, h _w,b() represents nonlinear automatic coding network A E function, represent with ρ be average and with for average two Bernoulli random variables between relative entropy, λ ₁and λ ₂represent the weight attenuation parameter of hidden layer and output layer respectively, ρ represents degree of rarefication coefficient, and its value is a constant being less than 0.1;

3.3) by minimizing cost function J _sparse(W, b) be optimized after the network parameter (W of automatic coding network A E _opt, b _opt):

$(W_{opt},b_{opt})=\underset{W,b}{min}{J}_{sparse}(W,b)=({W_{opt}}_{\left(jq\right)}^{\left(1\right)},{b_{opt}}_{\left(j\right)}^{\left(1\right)},{W_{opt}}_{\left(kj\right)}^{\left(2\right)},{b_{opt}}_{\left(k\right)}^{\left(2\right)}),$

Wherein, represent the weight of q node to a jth node of hidden layer of input layer after optimizing, represent the weight of a hidden layer jth node to hidden layer to an output layer kth node after optimizing, represent input layer being biased to a hidden layer jth node after optimizing, represent hidden layer node being biased to a kth node after optimizing, q=k=1,2 ..., 64,64 represent input layer numbers, and output layer nodes equals input layer number, j=1,2 ..., n, n represent node in hidden layer.

Step 4, asks the convolution own coding feature Fr of aurora image I.

4.1) by the weight of q the node to a jth node of hidden layer of optimizing rear input layer ask the training block of pixels feature pf that a hidden layer jth node is corresponding _j:

${\mathrm{pf}}_j=\frac{{W_{opt}}_{\left(jq\right)}^{\left(1\right)}}{\sqrt{{\mathrm{\Σ}}_{q=1}^{64}{\left({W_{opt}}_{\left(jq\right)}^{\left(1\right)}\right)}^2}};$

4.2) block of pixels feature pf will be trained _jcarry out matrix deformation, obtain training block of pixels eigenmatrix: f _j∈ R ^{8 × 8};

4.3) by aurora image I and training block of pixels eigenmatrix f _jconvolution, obtains a jth convolution own coding feature Fr of aurora image I _j:

Fr _j＝I*f _j，

Wherein, I ∈ R ^{128 × 128}, Fr _j∈ R ^{121 × 121};

4.4) by n convolution own coding feature Fr _jseries connection, obtains the convolution own coding feature of aurora image I:

Fr∈R ^121×121×n。

Step 5, the pondization that is averaged by the convolution own coding feature Fr of aurora image I operates, on average be divided into the block of 11 × 11 sizes by convolution own coding feature Fr, every block all merged into a mean value, then these mean values are reconfigured and obtain pond feature F ∈ R ^{11 × 11 × n}.

Step 6, is input to softmax sorter by the pond feature F of aurora image I, obtains a class label, is the classification of this aurora image I.

Effect of the present invention further describes by following emulation experiment.

Experiment 1:SCAE model parameter arranges experiment

Experiment condition: the present invention comes from Chinese Arctic Yellow River Station for the 3200 width aurora data of testing, this database comprises multi sphere shape, radiation crown shape, focus crown shape and each 800 width of valance Coronal aurorae image.

Experiment content 1: different automatic coding network A E node in hidden layer n is set, use the present invention to carry out aurora Images Classification, result is as Fig. 3, and wherein Fig. 3 (a) is classification accuracy, and Fig. 3 (b) is time consumption for training;

Experiment content 2: arrange different automatic coding network A E hidden layer degree of rarefication ρ, use the present invention to carry out aurora Images Classification, classification accuracy result is as Fig. 4.

As seen from Figure 3, as node in hidden layer n=400, the classification accuracy of aurora image is the highest; Node in hidden layer n is larger, and time consumption for training is longer.

As seen from Figure 4, when the hidden layer degree of rarefication of automatic coding network A E is 0.03, the classification accuracy of aurora image is the highest.

Experiment 2: contrast with the classification accuracy of different model to aurora image.

Experiment condition: 3200 width aurora images in experiment 1 have been used in this experiment.

Experiment content: use existing Le-net5 method, CAE method respectively, and S-CAE method proposed by the invention is classified to aurora image, classification accuracy result is as Fig. 5.

As seen from Figure 5, S-CAE method proposed by the invention is adopted effectively to improve the classification accuracy of aurora image.

Claims (4)

1., based on an aurora image classification method for the convolution autoencoder network optimized, comprise the steps:

(1) input aurora image, ask aurora image saliency map and significantly scheme to extract training block of pixels ps based on it _{8 × 8}, composition training block of pixels collection P _{8 × 8 × 100000};

(2) to training block of pixels collection P _{8 × 8 × 100000}carry out whitening pretreatment, obtain the training sample set x after albefaction _pw;

(3) the training sample set x after albefaction is utilized _pw, training automatic coding network A E:

3a) training sample set is expressed as v={xp ⁽¹⁾, xp ⁽²⁾, xp ⁽³⁾..., xp ⁽ⁱ⁾..., xp ^(m), wherein xp ⁽ⁱ⁾represent i-th training sample, xp ⁽ⁱ⁾∈ R ⁶⁴, i=1,2 ..., m, m represent number of training; According to training sample xp ⁽ⁱ⁾ask an automatic coding network A E hidden layer jth neuronic average active degree:

$\widehat{{\mathrm{\ρ}}_j}=\frac{1}{m}\underset{i=1}{\overset{m}{\Σ}}\[a_{W,b}\left({\mathrm{xp}}^{\left(i\right)}\right)\],$

Wherein, j=1,2 ..., n, n represent node in hidden layer, a _w,b(xp ⁽ⁱ⁾) represent when being input as xp ⁽ⁱ⁾time an automatic coding network A E hidden layer jth neuronic activity, (W, b)=(W ⁽¹⁾, b ⁽¹⁾, W ⁽²⁾, b ⁽²⁾) represent the parameter of automatic coding network A E, wherein, W ⁽¹⁾represent the weight of input layer to hidden layer node, W ⁽²⁾represent the weight of hidden layer node to output layer node, b ⁽¹⁾represent input layer being biased to hidden layer node, b ⁽²⁾represent hidden layer node being biased to output layer node;

3b) according to backpropagation BP coaching method principle, with parameter (W, b) and the hidden layer of automatic coding network A E

Average active degree construct a sparse cost function J _sparse(W, b):

$\begin{array}{c}{J}_{sparse}\left(W,b\right)=\[\frac{1}{m}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\left(\frac{1}{2}\left|\right|h_{W,b}\left({\mathrm{xp}}^{\left(i\right)}\right)-{\mathrm{xp}}^{\left(i\right)}|{|}^2\right)\]\\ +\frac{{\mathrm{\λ}}_1}{2}\mathrm{\Σ}{\left(W^{\left(1\right)}\right)}^2+\frac{{\mathrm{\λ}}_2}{2}\mathrm{\Σ}{\left(W^{\left(2\right)}\right)}^2+\mathrm{\β}\underset{l=1}{\overset{n}{\mathrm{\Σ}}}KL\left(\mathrm{\ρ}|\widehat{{\mathrm{\ρ}}_j}\right)\end{array}$

In formula, h _w,b() represents nonlinear automatic coding network A E function, represent with ρ be average and with for average two Bernoulli random variables between relative entropy, λ ₁and λ ₂represent the weight attenuation parameter of hidden layer and output layer respectively, ρ represents degree of rarefication coefficient, and its value is a constant being less than 0.1;

3c) by minimizing cost function J _sparse(W, b) be optimized after the parameter (W of automatic coding network A E _opt, b _opt):

$(W_{opt},b_{opt})=\underset{W,b}{min}{J}_{sparse}(W,b)=({W_{opt}}_{\left(jq\right)}^{\left(1\right)},{b_{opt}}_{\left(j\right)}^{\left(1\right)},{W_{opt}}_{\left(kj\right)}^{\left(2\right)},{b_{opt}}_{\left(k\right)}^{\left(2\right)})$

Wherein, represent the weight of q node to a jth node of hidden layer of input layer after optimizing, represent the weight of a hidden layer jth node to hidden layer to an output layer kth node after optimizing, represent input layer being biased to a hidden layer jth node after optimizing, represent hidden layer node being biased to a kth node after optimizing, q=k=1,2 ..., 64,64 represent input layer numbers, and output layer nodes equals input layer number, j=1,2 ..., n, n represent node in hidden layer;

(4) by the weight of q the node to a jth node of hidden layer of optimizing rear input layer ask the convolution own coding feature Fr of aurora image I;

(5) pondization that is averaged by the convolution own coding feature Fr of aurora image I operates, the block of 11 × 11 sizes is on average divided into by convolution own coding feature Fr, every block is all merged into a mean value, then these mean values are reconfigured and obtain pond feature F ∈ R ^{11 × 11 × n};

(6) the pond feature F of aurora image is input to softmax sorter, obtains a class label, be the classification of this aurora image.

2. method according to claim 1, wherein asks aurora image saliency map in step (1) and significantly schemes to extract training block of pixels ps based on it _{8 × 8}, carry out as follows:

1a) for each pixel I (x of any width input aurora image I, y), obtain its brightness L (x, y), Gradient Features H (x, and edge binaryzation feature B (x y), y), and by these three kinds of Fusion Features, obtain the conspicuousness value of information of aurora image slices vegetarian refreshments I (x, y): S (x, y)=L (x, y)+H (x, y)+B (x, y), again by aurora image conspicuousness value of information S (x, y) a little form aurora image saliency map S;

1b) binaryzation operation is carried out to image saliency map S, obtain binaryzation and significantly scheme S ₁;

1c) significantly scheme S in binaryzation at random ₁the training block of pixels ps of upper extraction 8 × 8 size _{8 × 8}, judge the value of this block of pixels: if block of pixels ps _{8 × 8}in 1 value proportion be greater than 0.8, then extract the block of pixels p of original image I in this position _{8 × 8}; If block of pixels ps _{8 × 8}in 1 value proportion be less than or equal to 0.8, then do not deal with.

3. method according to claim 1, wherein step (2) is to training block of pixels collection P _{8 × 8 × 100000}carry out whitening pretreatment, carry out as follows:

Block of pixels collection P 2a) will be trained _{8 × 8 × 100000}carry out matrix deformation, obtain deformation matrix: x ∈ R ^{64 × 100000};

2b) ask the covariance matrix of x:

$\mathrm{\φ}=\frac{1}{m}{\mathrm{\Σ}}_{i=1}^m\left(x^{\left(i\right)}\right){\left(x^{\left(i\right)}\right)}^T,$

Wherein, i=1,2 ..., m, m=100000 represent number of training, x ⁽ⁱ⁾i-th row of representing matrix x;

2c) SVD decomposition is carried out to deformation matrix x: x=U φ V, obtain left basis matrix U and right basis matrix V, and x is expressed as in left basis matrix U direction: x _rot=U ^tx;

2d) according to 2b) and 2c) obtain training sample set x _pCAwhite:

$x_{PCAwhite}=U\frac{x_{rot}}{\sqrt{\mathrm{\φ}+\mathrm{\ϵ}}}$

Wherein, ε represent one be not 0 minimum number, its value is 10 ^-5.

4. method according to claim 1, wherein step (4) asks the convolution own coding feature Fr of aurora image I, carries out as follows:

4a) by the weight of q the node to a jth node of hidden layer of optimizing rear input layer ask the training block of pixels feature pf that a hidden layer jth node is corresponding _j:

${\mathrm{pf}}_j=\frac{{W_{opt}}_{\left(jq\right)}^{\left(1\right)}}{\sqrt{{\mathrm{\Σ}}_{q=1}^{64}{\left({W_{opt}}_{\left(jq\right)}^{\left(1\right)}\right)}^2}};$

4b) block of pixels feature pf will be trained _jcarry out matrix deformation, obtain training block of pixels eigenmatrix: f _j∈ R ^{8 × 8};

4c) by aurora image I and training block of pixels eigenmatrix f _jconvolution, obtains a jth convolution own coding feature Fr of aurora image I _j:

Fr _j＝I*f _j，

Wherein, I ∈ R ^{128 × 128}, Fr _j∈ R ^{121 × 121};

4d) by n convolution own coding feature Fr _jseries connection, obtains the convolution own coding feature of aurora image I:

Fr∈R ^121×121×n。