CN106250856A - A kind of high-definition picture scene classification method based on non-supervisory feature learning - Google Patents

Abstract

The invention belongs to field of remote sensing image processing, particularly relate to a kind of high-resolution remote sensing image scene classification method based on non-supervisory feature learning.The present invention is first with the significance information of significance Detection and Extraction image, and according to the significance information of image diverse location, from image, the size stochastical sampling according to significance chooses image block；The openness object function of feature and minimal reconstruction residual error object function study image block characteristics is utilized to express；Finally, utilize the feature operator learning out, the view data of remote sensing image to be identified is carried out convolution operation, extracts characteristics of image, the feature extracted input support vector machine is classified.The present invention utilizes the significance information of significance Detection and Extraction image, utilize sparse own coding from image learning feature representation, make study to feature in scene Recognition, have more robustness.

Description

A kind of high-definition picture scene classification method based on non-supervisory feature learning

Technical field

The invention belongs to field of remote sensing image processing, particularly relate to a kind of high-resolution based on non-supervisory feature learning distant Sense image scene classification method.

Background technology

High-resolution remote sensing image (referring generally to more than the remote sensing images of 1 meter of spatial resolution) has the highest sky due to it Between resolution, abundant spatial information can be that Objects recognition provides fine information, has been widely used in every field.So And, owing to image spatial resolution is higher, the scene being identified generally comprises multiple different classes of atural object and is mixed to get Pixel, these atural objects often have different structural informations, but owing to spectral resolution is relatively low, it tends to be difficult to distinguish.With The maturation of high-resolution imaging technology and the reduction of cost, high-definition picture is more and more used, but high-resolution The scene Recognition of rate image yet suffers from some restrictive conditions:

1) high-resolution existence makes scene be often made up of Various Complex atural object, causes foreign body with composing phenomenon, makes essence Really interpretation becomes a difficult point, because cannot determine which kind of material this pixel belongs on earth.

2) carry out scene Recognition, need to utilize the space of image, structure, texture and the semantic information obtained, from semanteme Interpret image in aspect, but adaptive feature learning is a difficult point.

3) some algorithms only make use of the shallow-layer information in high-definition picture, such as artificial designs such as structure, texture informations Feature, often have ignored the construction features of data itself.These methods can not efficiently extract the feature representation of image.

Accordingly, it would be desirable to one can utilize the high-level feature of data itself, the method having again the semantic information of excellence to express is come Carry out adaptive feature learning.

Summary of the invention

The present invention mainly provides a kind of non-supervisory feature learning method and solves the problem existing for existing method, it is provided that One can utilize the high-level feature of data itself, has again the method that the semantic information of excellence is expressed.

The technical scheme that the present invention provides is, a kind of high-definition picture scene classification side based on non-supervisory feature learning Method, comprises the steps of

Step 1, significance detect, utilize diverse location in image image block local and the overall situation similarity to calculate image Significance, it is achieved mode is as follows,

For different image blocks, first one image block similar function (1) of definition,

$d(x^i,x^j)=\frac{d_{color}(x^i,x^j)}{1+c\·d_{position}(x^i,x^j)}---\left(1\right)$

Wherein d_colorFor image block at the Euclidean distance of CIELab color space, d_positionFor image block in picture position The Euclidean distance in space, xⁱ,x^jBeing the image block of optional position, c is constant, chooses and correspondence image block xⁱK most like figure As block x^k, wherein k=1,2 ... K, utilize formula (2) finally to calculate the significance of image block position,

$S^i=1-\exp \{-\frac{1}{K}\underset{k=1}{\overset{K}{\Σ}}d(x^i,x^k)\}---\left(2\right)$

Wherein, SⁱRefer to image block xⁱSignificance；

Step 2, specific image block samples, and detects according to the significance described in step 1 and obtains the notable of image block in image Degree information, the random image block choosing default fixed size from image；

Step 3, utilizes the openness object function of feature and minimal reconstruction residual error object function study image block characteristics to express, According to the openness and minimal reconstruction residual error of objective optimization formula (5) optimization aim feature simultaneously,

$J(X,Z)=\frac{1}{2}{\mathrm{\Σ}}_{i=1}^m\left|\right|x^i-z^i|{|}^2+\frac{\mathrm{\λ}}{2}|\left|W\right|{|}^2---\left(3\right)$

$KL\left(\mathrm{\ρ}\right|\left|\widehat{\mathrm{\ρ}}\right)=\mathrm{\ρ}log\frac{\mathrm{\ρ}}{\widehat{\mathrm{\ρ}}}+(1-\mathrm{\ρ})log\frac{1-\mathrm{\ρ}}{1-\widehat{\mathrm{\ρ}}}---\left(4\right)$

$Y=J(X,Z)+\mathrm{\β}\underset{j=1}{\overset{K}{\Σ}}KL\left(\mathrm{\ρ}\right|\left|\widehat{\mathrm{\ρ}}\right)---\left(5\right)$

Wherein J (X, Z) is original image and the residual sum weights size through decoded image, and wherein X represents original graph As block, Z represents the image block of reconstruct, and i represents that image block is numbered, and m represents the number of image block, zⁱFor image block xⁱReconstruct knot Really, W represents the weights of object function, is intended to the feature extraction operator learnt, and λ is bound term weights；It it is neuron Degree of rarefication, ρ is target characteristic degree of rarefication,Being current signature degree of rarefication, β is the weights of degree of rarefication；Y is global error letter Number；

Step 4, is strengthened by data and random drop further enhances the image block characteristics expression that step 3 study is arrived；

Step 5, solves and updates feature operator, according to the objective optimization formula (5) described in step 3, utilizes stochastic gradient Descent method solves and updates feature operator；

Step 6, it is judged that whether iteration reaches maximum frequency of training, if reaching maximum frequency of training, goes to step 7；Otherwise Go to step 5；

Step 7, utilizes feature operator that image carries out convolutional neural networks operation and extracts characteristics of image；

Step 8, utilizes support vector machine to carry out scene Recognition.

And, the data described in step 4 strengthen by image carries out Random-Rotation translation realization, and described random drop leads to Crossing in network training process, random masked segment neuron realizes.

Compared with prior art, the invention has the beneficial effects as follows:

The present invention uses sparse own coding that feature carries out adaptive learning, and introduces data enhancing and random drop to spy The optimal solution levying operator is updated, it is to avoid data " over-fitting " problem.At the design aspect of object function, ingenious combination is sparse Object function and minimal reconstruction residual error object function carry out feature learning, can realize feature under rational time cost Study, learn to feature in scene Recognition, have more robustness.

Accompanying drawing explanation

Fig. 1 is the non-supervisory feature learning schematic diagram in the embodiment of the present invention.

Fig. 2 is that the convolutional neural networks in the embodiment of the present invention extracts characteristics of image schematic diagram.

Detailed description of the invention

Below in conjunction with the drawings and specific embodiments, the present invention is described further.

The present invention provides a kind of method for high-resolution remote sensing image scene classification, by the characterology in scene classification Practise and regard a non-supervisory problem concerning study as, utilize the significance information of significance Detection and Extraction image, utilize sparse own coding From image learning feature representation, make feature learning in conjunction with the openness object function of feature and minimal reconstruction residual error object function Result has robustness.

Introducing significance detection algorithm, significance detection algorithm is a kind of calculation learnt by human eye vision mechanism introduced feature Method.The image block with significance feature generally comprises the feature that image most represents, and therefore can ensure that study is to optimum Feature representation.Embodiment utilizes the image block of the image space positions comprising notable information to carry out feature coding study.So learn The reason practising coding is, the image block comprising significance information often has the image spatial feature most represented with sufficiently Semantic information.By the statistical information of image, it is known that texture and structural information in image are static, different local letters Breath is the most similar, utilizes the representational image block of significance information searching, can be with higher probability learning to complete The feature representation that office is optimum.

Introduce sparse own coding and feature is carried out coding study.The feature representation of different images usually contains different stricture of vaginas Reason, structurally and semantically information.Utilize sparse own coding method, the immanent structure of adaptive study image and feature representation, this Sample can make study to feature there is best scene separability.Utilize sparse constraint simultaneously, to study to feature carry out Optimize so that feature more robust.Embodiment, in order to avoid occurring data " over-fitting " problem during feature learning, introduces Data enhancing and random drop carry out the renewal of optimal solution position.Adapt to feature representation the most one-dimensional abandons operation, So can make learning algorithm more robust, improve the effectiveness of feature simultaneously.

Fig. 1 is non-supervisory feature learning schematic diagram, and non-supervisory feature learning is divided into two stages to carry out, and the first stage is base Image block in significance is sampled, first with the significance information of significance Detection and Extraction image, according to image diverse location Significance information, from image, the size stochastical sampling according to significance chooses image block (such as 10 × 10 sizes).Second-order Section is to utilize sparse own coding to carry out feature learning, based on the image block choosing out, utilizes adaptive of sparse own coding Practise feature extraction operator.Sparse own coding utilizes stochastic gradient descent method to solve.Stochastic gradient descent method is a kind of tradition Optimum algorithm of multi-layer neural network.Finally, utilize the feature operator learning out, the view data of remote sensing image to be identified is entered Row convolution operation, extracts characteristics of image, then the feature input support vector machine extracted is classified, such as, be divided into residence People district, runway, meadow, airport, industrial occupancy.

The flow process that embodiment provides specifically includes following steps:

Step 1, significance detects: utilize the similarity of the local of the image block of diverse location and the overall situation in image to calculate The significance of image.For different image blocks (embodiment is 10 × 10 sizes), first one image block similar function of definition (1),

$d(x^i,x^j)=\frac{d_{color}(x^i,x^j)}{1+c\·d_{position}(x^i,x^j)}---\left(1\right)$

Wherein d_colorFor image block at the Euclidean distance of CIELab color space, d_positionFor image block in picture position The Euclidean distance in space, xⁱ,x^jBeing the image block of optional position, c is constant, and the present embodiment takes c=3.According to similar function, its It is worth two image blocks of the biggest expression the most similar, chooses and correspondence image block xⁱK most like image block x^k(k=1,2 ... K, tool When body is implemented, K can be preset value voluntarily by those skilled in the art, it is proposed that value 9-15), utilize formula (2) finally to calculate The significance of this image block position,

$S^i=1-\exp \{-\frac{1}{K}\underset{k=1}{\overset{K}{\Σ}}d(x^i,x^k)\}---\left(2\right)$

Wherein, SⁱRefer to image block xⁱSignificance.

Step 2, specific image block sampling: according to the significance information of image, random choosing from image is preset fixing The image block (such as 10 × 10) of size.It is the image comprising notable information that the present embodiment sets significance more than remarkable threshold 0.75 Block, less than non-significant threshold value 0.25 for non-significant degree block；200,000 image blocks are randomly selected from remotely-sensed data to be identified, Wherein 75% is significance block, and 25% is non-significant degree block.When being embodied as, remarkable threshold, non-significant threshold value, image block numbers Can be preset voluntarily by those skilled in the art.

Step 3, utilizes the openness object function of feature and minimal reconstruction residual error object function study image block characteristics to express.

Utilize sparse own coding adaptive learning characteristic operator from image block.Sparse own coding is that one comprises one The neutral net of hidden layer, mainly encodes according to input picture and decodes, and obtains the forced coding method of image, simultaneously Make to solve code error minimum.Openness and the minimal reconstruction residual error of optimization aim feature while of coming according to objective optimization formula (5):

$J(X,Z)=\frac{1}{2}{\mathrm{\Σ}}_{i=1}^m\left|\right|x^i-z^i|{|}^2+\frac{\mathrm{\λ}}{2}|\left|W\right|{|}^2---\left(3\right)$

$L\left(\mathrm{\ρ}\right|\left|\widehat{\mathrm{\ρ}}\right)=\mathrm{\ρ}log\frac{\mathrm{\ρ}}{\widehat{\mathrm{\ρ}}}+(1-\mathrm{\ρ})log\frac{1-\mathrm{\ρ}}{1-\widehat{\mathrm{\ρ}}}---\left(4\right)$

$Y=J(X,Z)+\mathrm{\β}\underset{j=1}{\overset{K}{\Σ}}KL\left(\mathrm{\ρ}\right|\left|\widehat{\mathrm{\ρ}}\right)---\left(5\right)$

Wherein J (X, Z) is original image and the residual sum weights size through decoded image, and wherein X represents original graph As block, Z represents the image block of reconstruct, and i represents that image block is numbered, and m represents the number (being 200,000 in embodiment) of image block, zⁱFor Image block xⁱReconstruction result, W represents the weights of object function, the feature extraction operator namely learnt, and λ is bound term Weights, are used for retraining the numerical values recited of W, general value 0.005.Being the degree of rarefication of neuron, KL represents Kullback-Leibler divergence^[1], ρ is target characteristic degree of rarefication, is used for keeping feature openness, when being embodied as Can be pre-set by those skilled in the art, typically take 0-1,It it is current signature degree of rarefication.β is the weights of degree of rarefication, typically Value 0-1.Y is global error function.

In embodiment, the ingenious openness object function of combination feature and minimal reconstruction residual error object function carry out optimum special That levies chooses.The openness object function of feature requires that feature has diversity, can the sparse expression of learning characteristic effectively；? Little reconstructed residual object function calculates the reconstructed error of feature, and the optimum combination of learning characteristic is expressed.So at feature learning During, use the openness object function of feature and minimal reconstruction residual error is object function simultaneously, the optimum combination of learning characteristic Expression model, guarantees the openness of feature simultaneously.

[1]S.Kullback and R.A.Leibler,“On information and sufficiency,” Ann.Math.Stat.,vol.22,pp.79–86,1951.

Step 4, is strengthened by data and random drop further enhances the image block characteristics expression that study is arrived.Due to god Need substantial amounts of data through network training, and be difficult at present obtain.Embodiment, by image is carried out Random-Rotation translation, reaches The effect strengthened to data, and simulate the change between different pieces of information.Simultaneously special in order to improve the data of step 3 study further Levy expression, by adding random drop mechanism, i.e. in network training process, random masked segment neuron, add random Property, make feature more sane^[2]。

[2]G.E.Hinton,N.Srivastava,A.Krizhevsky,I.Sutskever,and R.R.Salakhutdinov,“Improving neural networks by preventing co-adaptation of feature detectors,”Arxiv preprint arXiv:1207.0580,2012.

Step 5, solves and updates feature operator.Objective optimization formula (5) according to step 3, utilizes stochastic gradient descent Method^[3]Carry out solving and updating feature operator.

[3]D.E.Rumelhart,G.E.Hinton,and R.J.Williams,“Learning representations by back-propagating errors,”Nature,vol.323,pp.533–536,1986.

Step 6, it is judged that whether iteration reaches maximum frequency of training, if reaching maximum frequency of training, goes to step 7；Otherwise Go to step 5.

Step 7, utilizes feature operator that image carries out convolutional neural networks operation and extracts characteristics of image.

Fig. 2 is the schematic diagram that convolutional neural networks extracts characteristics of image, inputs the data of a panel height image in different resolution, has k =3 wave bands.First with study to feature operator image carried out convolution operation, wherein n be input picture size (as N × n), w is the size (such as w × w) of feature operator, and s is the step-length of convolution operation, represents that being spaced several pixels once rolls up Long-pending operation, finally obtains the size characteristic pattern for (n-w)/s+1.What k represented is the number of feature operator, calculates through k feature The convolution operation of son, can obtain k width characteristic image, i.e. k wave band, obtain characteristic of correspondence figure, and characteristic pattern is by each Feature operator obtains.Utilize characteristic pattern, proceed pondization operation, reduce the dimension of characteristic pattern, obtain more effective feature. Pond is a kind of characteristic statistics mode, and characteristic pattern divides according to net lattice control, then asks for the average or of each grid Big value.As in figure 2 it is shown, characteristic pattern is divided into the grid of 4 × 4, each grid asks for maximum therein, obtains the spy of Chi Huahou Levy figure.Finally characteristic pattern is carried out vectorization, the feature of corresponding 16 dimensions of each characteristic pattern, obtain k × 16 after full connection and (figure is remembered For k*16) the final characteristic of division of dimension, input support vector machine with this and classify.

Step 8, utilizes support vector machine^[4]Carry out scene Recognition.

[4]Cortes C,Vapnik V.Support-vector networks[J].Machine learning, 1995,20(3):273-297.

Based on significance the non-supervisory feature learning method being above the present embodiments relate to realize step.Pass through The study of significance detection algorithm, sparse own coding, data strengthen and the introducing of random drop, can be at more rational time cost Under, extract the non-supervisory feature learning operator of data.

Following points for attention are also had implementing when:

Maximum iteration time in the optimizing stage arranges aspect, owing to initial value is randomly provided, it is possible that need More iteration searches out the optimal solution of feature operator；Random drop introduces again more randomness, algorithm the convergence speed Relatively slow, those skilled in the art can preset maximum iteration time the most based on experience value.

When being embodied as, method provided by the present invention can realize automatic operational process based on software engineering.

It is emphasized that embodiment of the present invention is illustrative rather than determinate.Bag the most of the present invention Include the embodiment being not limited to described in detailed description of the invention, every by those skilled in the art according to technical scheme Other embodiments drawn, also belong to the scope of protection of the invention.

Claims (2)

1. a high-definition picture scene classification method based on non-supervisory feature learning, it is characterised in that comprise following step Rapid:

Step 1, significance detects, and utilizes the image block local of diverse location in image and overall situation similarity to calculate the aobvious of image Work degree, it is achieved mode is as follows,

For different image blocks, first one image block similar function (1) of definition,

$d(x^i,x^j)=\frac{d_{color}(x^i,x^j)}{1+c\·d_{position}(x^i,x^j)}---\left(1\right)$

Wherein d_colorFor image block at the Euclidean distance of CIELab color space, d_positionFor image block in space, picture position Euclidean distance, xⁱ,x^jBeing the image block of optional position, c is constant, chooses and correspondence image block xⁱK most like image block x^k, wherein k=1,2 ... K, utilize formula (2) finally to calculate the significance of image block position,

$S^i=1-\exp \{-\frac{1}{K}\underset{k=1}{\overset{K}{\Σ}}d(x^i,x^k)\}---\left(2\right)$

Wherein, SⁱRefer to image block xⁱSignificance；

Step 2, specific image block is sampled, and detects the significance letter obtaining image block in image according to the significance described in step 1 Breath, the random image block choosing default fixed size from image；

Step 3, utilizes the openness object function of feature and minimal reconstruction residual error object function study image block characteristics to express, according to Openness and the minimal reconstruction residual error of objective optimization formula (5) optimization aim feature simultaneously,

$J(X,Z)=\frac{1}{2}{\mathrm{\Σ}}_{i=1}^m\left|\right|x^i-z^i|{|}^2+\frac{\mathrm{\λ}}{2}|\left|W\right|{|}^2---\left(3\right)$

$KL\left(\mathrm{\ρ}\right|\left|\widehat{\mathrm{\ρ}}\right)=\mathrm{\ρ}log\frac{\mathrm{\ρ}}{\widehat{\mathrm{\ρ}}}+(1-\mathrm{\ρ})log\frac{1-\mathrm{\ρ}}{1-\widehat{\mathrm{\ρ}}}---\left(4\right)$

$Y=J(X,Z)+\mathrm{\β}\underset{j=1}{\overset{K}{\Σ}}KL\left(\mathrm{\ρ}\right|\left|\widehat{\mathrm{\ρ}}\right)---\left(5\right)$

Wherein J (X, Z) is original image and the residual sum weights size through decoded image, and wherein X represents original picture block, Z represents the image block of reconstruct, and i represents that image block is numbered, and m represents the number of image block, zⁱFor image block xⁱReconstruction result, W table Showing the weights of object function, be intended to the feature extraction operator learnt, λ is bound term weights；It is the dilute of neuron Dredging degree, ρ is target characteristic degree of rarefication,Being current signature degree of rarefication, β is the weights of degree of rarefication；Y is global error function；

Step 4, is strengthened by data and random drop further enhances the image block characteristics expression that step 3 study is arrived；

Step 5, solves and updates feature operator, according to the objective optimization formula (5) described in step 3, utilizes stochastic gradient descent Method solves and updates feature operator；

Step 6, it is judged that whether iteration reaches maximum frequency of training, if reaching maximum frequency of training, goes to step 7；Otherwise go to Step 5；

Step 7, utilizes feature operator that image carries out convolutional neural networks operation and extracts characteristics of image；

Step 8, utilizes support vector machine to carry out scene Recognition.

A kind of high-definition picture scene classification method based on non-supervisory feature learning, it is special Levy and be: the data described in step 4 strengthen by image carries out Random-Rotation translation realization, and described random drop is by net During network training, random masked segment neuron realizes.