CN105184298A - Image classification method through fast and locality-constrained low-rank coding process - Google Patents

Abstract

The invention discloses an image classification method through the fast and locality-constrained low-rank coding process. The method comprises a dictionary learning part, a training part and a testing part. During the dictionary learning part, a corresponding dictionary is obtained based on SIFT features extracted out of all data sets according to the k-means clustering method or other dictionary learning methods. During the training part, one part of images in each data set are selected as training images, and then the corresponding low-rank representation corresponding to each data set is learned based on the tags of the images. Through the SPM quantization process and the SVM classifier-based training process, category labels and a kernel function corresponding to a feature representation matrix are obtained. During the testing part, all remaining images in each data set are adopted as to-be-classified test images. For any unclassified image, the SIFT features of the image are extracted firstly, and then a low-rank coding matrix thereof in the dictionary is found out. Through the SPM quantization process, a final feature representation matrix is obtained. The final feature representation matrix is input into an SVM classifier obtained during the training part, so that the category of the image can be known.

Description

A kind of image classification method of quick local restriction low-rank coding

Technical field

The present invention relates to a kind of method utilizing quick local restriction low-rank to encode and carry out Images Classification, also precision is higher fast for the method, is a kind of sorting technique of data-driven of image content-based.

Background technology

Images Classification belongs to an important research direction of computer vision, and it comprises the pre-service of image, the expression of feature and the design of sorter, and wherein feature representation comprises the coding of the extraction of characteristics of image, the dimensionality reduction of feature and feature.Feature extraction mainly by attributes such as the space distributions of various content presentation color out in image, brightness, texture, shape and pixel to iamge description.Feature expression refers to carry out adding up (vector quantization), coding or additive method to form the last feature of piece image in the most basic feature base, can have better performance under normal circumstances than original essential characteristic; Good feature representation can promote the performance of Images Classification and identification.

The Measures compare represented for feature interpretation is in recent years many, there is visual word bag model (bagofwords, BOW), based on the spatial pyramid model (SparseCodingSpatialPyramidMatching of sparse coding, ScSPM), local restriction uniform enconding (Locality-constrainedLinearCoding, LLC), image classification method (the StructuredLow-rankRepresentation that structure low-rank represents, SLRR) and low-rank sparse decompose image classification method (Low-rankSparseCoding, LRSC), these methods overcome some difficult problems to a certain extent, but still all Shortcomings.

Visual word bag model (bagofwords, BOW), although be easy to build, have ignored similar image and has certain space structure similarity.Spatial pyramid model (SparseCodingSpatialPyramidMatching, ScSPM) based on sparse coding have ignored the correlativity between local description, and to the change of feature and noise-sensitive, and calculated amount is very large, consumption internal memory.Local restriction uniform enconding (Locality-constrainedLinearCoding, LLC), this locality result in openness to a certain extent, relative to sparse coding, the operand of algorithm is reduced greatly, speed improves, but LLC have ignored the spatial information of the one-piece construction feature between feature.Image classification method (the StructuredLow-rankRepresentation that structure low-rank represents, SLRR), there is good effect to recognition of face, be particularly useful for having the Images Classification of critical noisy (as block, lighting angle change etc.).Image classification method (Low-rankSparseCoding) (LRSC) nicety of grading that low-rank sparse is decomposed obtains certain raising, but have ignored local spatial information; Wherein sparse part is still very sensitive to coding.Based on the image classification method (FastLow-rankRepresentationbasedSpatialPyramidMatching that the quick low-rank of spatial pyramid represents, LrrSPM), the classification speed of the method is 5 ~ 16 times of ScSPM, but nicety of grading is lower than ScSPM and LLC method.

The coded system that current image classification method mostly adopts sparse coding and low-rank to represent, but in view of sparse coding to the change of image and noise more responsive, very consume internal memory; The coded system adopting low-rank to represent obtains a balance between precision and computation complexity.

Summary of the invention

In view of the defect of current sorting technique in classifying quality and classification speed, the present invention proposes a kind of image classification method of quick local restriction low-rank coding, and the method comprises dictionary learning part, training part and part of detecting three parts.Dictionary learning means suitable, to the SIFT feature of image zooming-out images all in image library, obtains dictionary through K-means cluster; Training part is first to the image zooming-out SIFT feature in training set, then try to achieve the feature coding of SIFT feature under dictionary to represent, adopt the low-rank coding method of local restriction, consider global structure consistance and the local spatial simlanty of image simultaneously, obtain feature coding matrix, behind spatial pyramid coupling core, pond, obtain final character representation matrix, adopt SVM method to train character representation matrix and class mark, set up disaggregated model.First part of detecting carries out feature extraction to any image, then encoder matrix under its dictionary is tried to achieve, after spatial pyramid coupling core, pond obtain character representation matrix, the SVM classifier obtained is trained in input, obtains the classification results belonging to this image.

The technical scheme adopted for realizing the object of the invention is such, and a kind of image classification method of quick local restriction low-rank coding, comprises dictionary learning part, training part and part of detecting.

Described dictionary learning part steps process is as follows:

1) utilize SIFT feature extracting method to carry out local shape factor to all images of image library, obtain input feature value matrix X=[x ₁, x ₂..., x _n] ∈ R ^{m × n}, wherein x _i∈ R ⁿbe i-th component vector of eigenvectors matrix, wherein m is the dimension of vector, and n is vectorial number.

2) utilize K-means clustering or other dictionary learning methods, clustering processing is carried out to the eigenvectors matrix X of all image zooming-out, complete dictionary D=[d must be ₁, d ₂..., d _k] ∈ R ^{m × k}, wherein m is the dimension of dictionary, and k is the number of dictionary base, and meet dictionary and cross completeness m < < k, k is generally 256,1024,4096 etc.

Described training department step by step process is as follows:

I) from image library, selected part image, as training image, utilizes SIFT feature extracting method to carry out local shape factor to all training images, obtains input feature value matrix X'=[x ₁', x ₂' ..., x _n'] ∈ R ^{m × n}.Wherein x _i' ∈ R ⁿbe i-th component vector of eigenvectors matrix, wherein m is the dimension of vector, and n is vectorial number.

Ii) each proper vector x is calculated _iand the Euclidean distance p between dictionary vector _i, computing formula is as follows:

$p_i=\exp \left(\frac{dist({x_i}^{\&prime;},D)}{\mathrm{\&sigma;}}\right)$

Dist (x _i', D)=[dist (x _i', d ₁) ..., dist (x _i', d _k)] ^t, dist (x _i', d _i) be proper vector x _i' d vectorial with each dictionary _ibetween Euclidean distance, σ is the weight of adjustment local restriction speed.

Iii) calculate the character representation vector Z of local restriction low-rank coding, computing formula is:

$Z={({\mathrm{\&lambda;}}_{1}I+{\mathrm{DD}}^T+{\mathrm{\&lambda;}}_1{\mathrm{\&lambda;}}_2\underset{i=1}{\overset{k}{\&Sigma;}}{p_i}^T{p}_i)}^{-1}{D}^T{X}^{\&prime;}$

Wherein λ ₁, λ ₂> 0 is weight; p _ifor the Euclidean distance p that above-mentioned steps ii calculates _i.

Iv) adopt spatial pyramid matching process to process to the character representation vector Z obtained, Z is divided into 2 ^l× 2 ^lblock, wherein l=1,2,4, represent the number of different three layers of corresponding spatial pyramid piecemeal respectively.Then for the z of each block in every layer _icarry out maximum pond, i.e. z _i=max{|z _1i|, | z _2i| ..., | z _ji|, obtain the i-th vectorial z of final low-rank coding _i, wherein z _jii-th element in jth block, j=2 ^l× 2 ^l, obtain last character representation matrix Z' by this process.

V) character representation matrix Z' is inputted SVM classifier, according to every class image category label and its characteristic of correspondence representing matrix Z', study obtains the kernel function of SVM classifier.

Described part of detecting step process is as follows:

I) to all image zooming-out SIFT feature remaining in image library, input feature value matrix X*=[x is obtained ₁*, x ₁* ..., x _n*] ∈ R ^{m × n}, wherein x _i* ∈ R ⁿi-th component vector of eigenvectors matrix.

II) each proper vector x is calculated _i* the Euclidean distance p and between dictionary vector _i, computing formula is as follows:

$p_i*=\exp \left(\frac{dist(x_i*,D)}{\mathrm{\&sigma;}}\right)$

Dist (x _i*, D)=[dist (x _i*, d ₁) ..., dist (x _i*, d _k)] ^t, dist (x _i*, d _i) be proper vector x _i* with each dictionary vector d _ibetween Euclidean distance, σ is the weight of adjustment local restriction speed.

III) calculate the character representation vector Z * of local restriction low-rank coding, computing formula is:

$Z*={({\mathrm{\&lambda;}}_{1}I+{\mathrm{DD}}^T+{\mathrm{\&lambda;}}_1{\mathrm{\&lambda;}}_2\underset{i=1}{\overset{k}{\&Sigma;}}{p}_i{*}^T{p}_i*)}^{-1}{D}^{T}X*$

Wherein λ ₁, λ ₂> 0 is weight; p _i* be Euclidean distance that above-mentioned steps II calculates.

IV) adopt spatial pyramid matching process, character representation vector Z * is divided into 2 ^l× 2 ^lblock, wherein l=1,2,4, represent the number of different three layers of corresponding spatial pyramid piecemeal respectively.Then for the z of each block in every layer _i* carry out maximum pond, keep character representation vector to have certain robustness, i.e. z _i*=max{|z _1i* |, | z _2i* | ..., | z _ji* | }, obtain the i-th vectorial z of final low-rank coding _i*, wherein z _ji* be i-th element in jth block, j=2 ^l× 2 ^l, obtain last character representation matrix Z*' like this.

V) character representation matrix Z*' is inputted training department and divide the SVM classifier obtained, obtain the class label belonging to image to be classified.

VI) the statistical test classified image number percent of correctly classifying, obtains the nicety of grading that every class testing image is final.

Further, in the image classification method of described a kind of quick local restriction low-rank coding, the low-rank that in training part, in step I ii and part of detecting, Step II I calculates feature coding represents vector.

The coding method that existing low-rank represents is that the low-rank finding proper vector in the error range allowed represents, namely

$\underset{Z,E}{min}\left|\right|Z|{|}_{*}+{\mathrm{\&lambda;}}_1|\left|E\right|{|}_p$

s.t.X＝DZ+E

Wherein μ ₁, μ ₂> 0 is error penalty term, λ ₁> 0 is weights, and Z is the expression of feature vector, X under dictionary D.Wherein ‖ Z ‖ _*nuclear norm, ‖ E ‖ _pthe l of constraint sparse coding item ₁norm or retrain a certain class interference l _2-1norm.

Existing low-rank optimized algorithm commonly adopts Lagrange multiplier to optimize AugmentedLagrangeMultiplealgorithm (ALM), and optimizing expression is as follows:

$\begin{array}{c}\underset{Z,D}{\min }\left|\right|Z|{|}_{*}+\mathrm{\&lambda;}||D-D_0|{|}_F+tr\&lsqb;Y^T(D-D_0-E_1)\&rsqb;+\\ tr\&lsqb;Y^T(X-DZ-E_2)\&rsqb;+\frac{{\mathrm{\&mu;}}_1}{2}\left|\right|D-D_0-E_1|{|}_F^2+\frac{{\mathrm{\&mu;}}_2}{2}||X-DZ-E_2|{|}_F^2\end{array}---\left(1\right)$

Wherein ‖ ‖ _ff norm (Frobeniusnorm), λ ₁> 0 is weights, Y ₁, Y ₂lagrange multiplier, μ ₁, μ ₂> 0 is error penalty term.

Above formula 1 can adopt typical exactALM or inexactALM method, adopt alternative manner step by step respectively regeneration characteristics represent vector Z and initial dictionary D ₀, the computation complexity of this method is O (mn ²), n is the number of proper vector.This calculating comes that is very large, and is a kind of method of iteration optimization of off-line, cannot obtain classification results (can not asking character representation if do not belonged to such other image) to the input picture of increment type.Meanwhile, existing low-rank constraint only considered the global structure consistance of characteristics of image, have ignored the local space similarity of feature, comprehensively can not represent the related information between characteristics of image.

The low-rank coding method step calculating local restriction in this method is as follows:

A) when dictionary error is zero, E ₁=D-D ₀=0, F norm ‖ ‖ _fnuclear norm ‖ ‖ can be replaced _*.Above-mentioned optimizing expression 1 becomes

$\underset{Z}{min}\left|\right|Z|{|}_F+{\mathrm{\&lambda;}}_1/2||X-DZ|{|}_F^2---\left(2\right)$

B) combined coding mode is taked in low-rank constraint, on the conforming basis of characteristics of image global structure, makes up the local space similarity of feature, adds local restriction item above-mentioned expression formula 2 becomes:

Wherein λ ₁, λ ₂> 0 is unequal weight, and symbol ⊙ represents that the corresponding element of two vectors is multiplied, p _ifor proper vector x _iwith the Euclidean distance p of dictionary vector D _i

$p_i=\exp \left(\frac{dist(x_i,D)}{\mathrm{\&sigma;}}\right)$

Dist (x _i, D) and=[dist (x _i, d ₁) ..., dist (x _i, d _k)] ^t, dist (x _i, d _i) be proper vector x _iwith each dictionary vector d _ibetween Euclidean distance, σ is the weight of adjustment local restriction speed.

C) differentiate is carried out to above-mentioned expression formula 3

Finally calculate character representation vector: $Z={({\mathrm{\&lambda;}}_{1}I+{\mathrm{DD}}^T+{\mathrm{\&lambda;}}_1{\mathrm{\&lambda;}}_2\underset{i=1}{\overset{k}{\&Sigma;}}{p_i}^T{p}_i)}^{-1}{D}^{T}X,$ Wherein I is the vector of unit length of k × k.

In differentiate process, take threshold value constraint item by character representation vector z _iconstrain within 0.98.

Technique effect of the present invention is mathematical, the present invention adopts a kind of low-rank coding specification method of local restriction fast, consider global structure consistance and the local spatial simlanty of characteristics of image simultaneously, better can not only represent the related information between characteristics of image, obtain effective nicety of grading, and calculated amount reduces, classification speed is a bit larger tham LLC, is only 19.4 ~ 25% of ScSPM.

Accompanying drawing explanation

Fig. 1 is Images Classification main-process stream;

Fig. 2 is local restriction low-rank coding (FLCLR) process flow diagram;

Fig. 3 is the example images of classification in the middle part of table 1;

Fig. 4 is the example images of classification in the middle part of table 2;

Fig. 5 is the example images of classification in the middle part of table 4;

In figure: 1-feature extraction, 2-feature coding, 3-pond process.

Embodiment

Below in conjunction with embodiment, the invention will be further described, but should not be construed the above-mentioned subject area of the present invention and be only limitted to following embodiment.Without departing from the idea case in the present invention described above, according to ordinary skill knowledge and customary means, make various replacement and change, all should be included in protection scope of the present invention.

Below implement to choose Images Classification conventional authoritative image library Scene13 and Caltech101 as class test data set.Scene13 is the most frequently used scene classification data set, has 13 class images, comprises suburb (CALsuburb), seashore (MITcoast), and totally 4485 width such as forest (MITforest), every class comprises 200 to 400 pictures.Caltech101 image library comprises 101 classes (as animal, flower, face etc.) and a background classes image, totally 9144 width.Each class comprises 31 to 800 width images, and has difference in larger class, is the most frequently used database of Images Classification.

As shown in Figure 1, whole implementation step is divided into three parts, and Part I is dictionary learning part, extracts all SIFT feature according to data set, uses the dictionary learning methods such as k-menas clustering to obtain corresponding dictionary.Part II is training part, chooses the parts of images in the every class of each data set, and wherein Scene13 gets 100 at random as training image, and Caltech101 gets 30 at random as training image.Then the low-rank learning each classification corresponding for the label of these training images represents, quantizes to obtain obtaining class label and kernel function corresponding to character representation through svm classifier training through SPM.Part III is part of detecting, in each database, remaining all images are as test pattern to be sorted, the image of classification is not known for any one, first SIFT feature is extracted, then its low-rank coding vector under dictionary is tried to achieve, quantize to obtain final character representation matrix through SPM, inputted the SVM classifier that training part gained arrives, and then obtain the classification belonging to this image.

Pre-service illustrates: image size is adjusted to maximum 300 × 300 pixels, and the grid of the intensive Bian sample of feature extraction is 16 × 16 pixels, and its step-length is 4 pixels.

Dictionary learning part:

1) to image zooming-out SIFT local features all in image library, input feature value matrix X=[x is obtained ₁, x ₂..., x _n] ∈ R ^{m × n}, wherein x _i∈ R ⁿbe i-th local feature vectors of image, wherein m is the dimension of vector, and n is vectorial number.

2) utilize K-means clustering or other dictionary learning methods, clustering processing is carried out to the proper vector of all image zooming-out, complete dictionary D=[d must be ₁, d ₂..., d _k] ∈ R ^{m × k}, wherein k is the dimension of dictionary, and m is the number of dictionary base, meets dictionary and crosses completeness m < < k, and wherein k is that the dimension of dictionary gets 1024.

Training part:

I) SIFT feature is extracted to the training image chosen, obtain input feature value matrix X'=[x' ₁, x' ₂..., x' _n] ∈ R ^{m × n}, wherein x _i' ∈ R _ni-th component vector of eigenvectors matrix.

Ii) each proper vector x in calculation training image _i' and dictionary vector between Euclidean distance p _i, computing formula is as follows:

$p_i=\exp \left(\frac{dist({x_i}^{\&prime;},D)}{\mathrm{\&sigma;}}\right)$

Dist (x _i', D)=[dist (x _i', d ₁) ..., dist (x _i', d _k)] ^t, dist (x _i', d _i) be proper vector x _i' d vectorial with each dictionary _ibetween Euclidean distance; σ is the weight of adjustment local restriction speed, gets 100.

Iii) calculate the character representation vector Z of local restriction low-rank coding, computing formula is:

$Z={({\mathrm{\&lambda;}}_{1}I+{\mathrm{DD}}^T+{\mathrm{\&lambda;}}_1{\mathrm{\&lambda;}}_2\underset{i=1}{\overset{k}{\&Sigma;}}{p}_i^T{p}_i)}^{-1}{D}^T{X}^{\&prime;}$

Wherein I is unit matrix, and dimension is consistent with Z, λ ₁, λ ₂> 0 is weight, λ ₁, λ ₂0.7 is respectively, 0.0001 according to the empirical value value of document Wang and Peng.P _ifor the Euclidean distance calculated in above-mentioned steps 2.

Step I i and step I ii seeks to obtain character representation vector z, the i.e. X=DZ of eigenvectors matrix X under dictionary D _,wherein Z=[z ₁, z ₁..., z _n] ∈ R ^{k × n}.Algorithm flow chart as shown in Figure 2.

Iv) adopt spatial pyramid matching process, Z is divided into 2 ^l× 2 ^lblock, wherein l=1,2,4, represent the number of different three layers of corresponding spatial pyramid piecemeal respectively.Then for the z of each block in every layer _icarry out maximum pond, keep character representation vector to have certain robustness, i.e. z _i=max{|z _1i|, | z _2i| ..., | z _ji|, obtain the i-th vectorial z of final low-rank coding _i, wherein z _jii-th element in jth block, j=2 ^l× 2 ^l, obtain last character representation matrix Z' like this.

V) character representation matrix Z' is inputted SVM classifier, according to every class image category label and its characteristic of correspondence representing matrix Z', study obtains the kernel function of SVM classifier.

Part of detecting:

I) to all image zooming-out SIFT feature remaining in image library, input feature value matrix X*=[x is obtained ₁*, x ₁* ..., x _n*] ∈ R ^{m × n}, wherein x _i* ∈ R ⁿi-th component vector of eigenvectors matrix.

II) each proper vector x is calculated _i* the Euclidean distance p and between dictionary vector _i, computing formula is as follows:

$p_i*=\exp \left(\frac{dist(x_i*,D)}{\mathrm{\&sigma;}}\right)$

Dist (x _i*, D)=[dist (x _i*, d ₁) ..., dist (x _i*, d _k)] ^t, dist (x _i*, d _i) be proper vector x _i* with each dictionary vector d _ibetween Euclidean distance, σ is the weight of adjustment local restriction speed.

III) calculate the character representation vector Z * of local restriction low-rank coding, computing formula is:

$Z*={({\mathrm{\&lambda;}}_{1}I+{\mathrm{DD}}^T+{\mathrm{\&lambda;}}_1{\mathrm{\&lambda;}}_2\underset{i=1}{\overset{k}{\&Sigma;}}{p}_i{*}^T{p}_i*)}^{-1}{D}^{T}X*$

Wherein λ ₁, λ ₂> 0 is weight; p _i* be Euclidean distance that above-mentioned steps II calculates.

IV) adopt spatial pyramid matching process, Z* is divided into 2 ^l× 2 ^lblock, wherein l=1,2,4, represent the number of different three layers of corresponding spatial pyramid piecemeal respectively.Then for the z of each block in every layer _i* carry out maximum pond, keep character representation vector to have certain robustness, i.e. z _i*=max{|z _1i* |, | z _2i* | ..., | z _ji* | }, obtain the i-th vectorial z of final low-rank coding _i*, wherein z _ji* be i-th element in jth block, j=2 ^l× 2 ^l, obtain last character representation matrix Z*' like this.

V) character representation matrix Z*' is inputted training department and divide the SVM classifier obtained, obtain the class label belonging to image to be classified.

VI) the statistical test classified image number percent of correctly classifying, obtains the nicety of grading that every class testing image is final.

In order to contrast the validity of the inventive method, have chosen the corresponding programs of the most frequently used several associated picture sorting techniques in test, arranging of the experiment parameter of contrast test is identical, all obtains experiment comparison of classification result, as shown in list.

Table 1 is the part classifying precision (first group) of Scene13;

Table 2 is the part classifying precision (second group) of Scene13;

Table 3 is overall nicety of grading and the time of Scene13;

Table 4 is the part classifying precision (first group) of Caltech101;

Table 5 is overall nicety of grading and the time of Caltech101.

Wherein ScSPM is the spatial pyramid model image sorting technique based on sparse coding, LLC is local restriction uniform enconding image classification method, LrrSPM is the image classification method represented based on the quick low-rank of spatial pyramid, and FLCLR is the image classification method of the quick local restriction geocoding that this experiment proposes.

Table 1

Classes	ScSPM	LLC	LrrSPM	FLCLR
					MITtallbuilding	0.7182	0.7727	0.8148	0.85714
bedroom	0.8787	0.8129	0.9074	0.9858
					MITforest	0.8397	0.8431	0.8538	0.9414
MIThighway	0.6387	0.6774	0.7027	0.7818
					MITcoast	0.83654	0.8846	0.8125	0.89375
kitchen	0.9118	0.9161	0.8992	0.9423
					MITmountain	0.8980	0.9102	0.9025	0.93226
Mean	0.8174	0.8310	0.8420	0.9049

Table 2

Classes	ScSPM	LLC	LrrSPM	FLCLR
					Livingroom	0.8550	0.8438	0.83125	0.8437
MITstreet	0.6638	0.6840	0.6293	0.6207
					MITopencountry	0.9478	0.9391	0.8783	0.8922
MITinsidecity	0.8802	0.8654	0.8594	0.8608
					CALsuburb	0.8323	0.8308	0.8154	0.8123
PARoffice	0.6725	0.6032	0.6032	0.6009
					Mean	0.8086	0.7944	0.7695	0.7718

Table 3

Table 4

Classes	ScSPM	LLC	LrrSPM	FLCLR
					car_side	0.7000	0.7318	0.7536	0.8364
chair	0.4615	0.6223	0.6485	0.72230
					Faces	0.6470	0.6647	0.7059	0.7765
mayfly	0.7192	0.7192	0.7587	0.8462
					pyramid	0.6667	0.7578	0.7783	0.8519
water_lilly	0.4286	0.6429	0.6943	0.7714
					widsor_char	0.84615	0.9315	0.9534	0.9931

Table 5

Claims (1)

1. an image classification method for quick local restriction low-rank coding, is characterized in that: comprise dictionary learning part, training part and part of detecting;

Described dictionary learning part comprises the following steps process;

1) utilize SIFT feature extracting method to carry out local shape factor to all images of image library, obtain input feature value matrix X=[x ₁, x ₂..., x _n] ∈ R ^{m × n}; Wherein x _i∈ R ⁿbe i-th component vector of eigenvectors matrix, m is the dimension of vector, and n is vectorial number;

2) utilize K-means clustering or other dictionary learning methods, clustering processing is carried out to the eigenvectors matrix X of all image zooming-out, complete dictionary D=[d must be ₁, d ₂..., d _k] ∈ R ^{m × k}; Wherein m is the dimension of dictionary, and k is the quantity of dictionary base;

Described training department divides the process of comprising the following steps;

I) from image library, selected part image, as training image, utilizes SIFT feature extracting method to carry out local shape factor to all training images, obtains input feature value matrix X'=[x ₁', x ₂' ..., x _n'] ∈ R ^{m × n}; Wherein x _i' ∈ R ⁿbe i-th component vector of eigenvectors matrix, wherein m is the dimension of vector, and n is vectorial number;

Ii) each proper vector x is calculated _i' and dictionary vector D between Euclidean distance p _i, computing formula is as follows:

$p_i=\exp \left(\frac{dist({x_i}^{\&prime;},D)}{\mathrm{\&sigma;}}\right)$

Dist (x _i', D)=[dist (x _i', d ₁) ..., dist (x _i', d _k)] ^t, dist (x _i', d _i) be proper vector x _i' d vectorial with each dictionary _ibetween Euclidean distance; σ is the weight of adjustment local restriction speed;

Iii) calculate the character representation vector Z of local restriction low-rank coding, computing formula is:

$Z={({\mathrm{\&lambda;}}_{1}I+{\mathrm{DD}}^T+{\mathrm{\&lambda;}}_1{\mathrm{\&lambda;}}_2\underset{i=1}{\overset{k}{\&Sigma;}}{p_i}^T{p}_i)}^{-1}{D}^T{X}^{\&prime;}$

Wherein λ ₁, λ ₂> 0 is weight, and I is unit vector, Z=[z ₁, z ₁..., z _n] ∈ R ^{k × n}; p _ifor the Euclidean distance that above-mentioned steps ii calculates;

Iv) adopt spatial pyramid matching process to process to the character representation vector Z obtained, Z is divided into 2 ^l× 2 ^lblock, wherein l=1,2,4, represent the number of different three layers of corresponding spatial pyramid piecemeal respectively; Then for the z of each block in every layer _icarry out maximum pond, i.e. z _i=max{|z _1i|, | z _2i| ..., | z _ji|, obtain the i-th vectorial z of final low-rank coding _i, wherein z _jii-th element in jth block, j=2 ^l× 2 ^l, obtain last character representation matrix Z' by this process;

V) character representation matrix Z' is inputted SVM classifier; Each image is to there being a label, and according to every class image category label and its characteristic of correspondence representing matrix Z', study obtains the kernel function of SVM classifier;

Described part of detecting comprises the following steps process;

I) to remaining image after choosing training image in image library, extract SIFT feature, obtain input feature value matrix X*=[x ₁ ^*, x ₁ ^*..., x _n ^*] ∈ R ^{m × n}, wherein x _i ^*∈ R ⁿi-th component vector of eigenvectors matrix;

II) each proper vector x is calculated _i ^*and the Euclidean distance p between dictionary vector _i, computing formula is as follows:

$p_i*=\exp \left(\frac{dist(x_i*,D)}{\mathrm{\&sigma;}}\right)$

Dist (x _i ^*, D) and=[dist (x _i ^*, d ₁) ..., dist (x _i ^*, d _k)] ^t, dist (x _i ^*, d _i) be proper vector x _i ^*with each dictionary vector d _ibetween Euclidean distance; σ is the weight of adjustment local restriction speed;

III) calculate the character representation vector Z * of local restriction low-rank coding, computing formula is:

$Z*={({\mathrm{\&lambda;}}_{1}I+{\mathrm{DD}}^T+{\mathrm{\&lambda;}}_1{\mathrm{\&lambda;}}_2\underset{i=1}{\overset{k}{\&Sigma;}}{p}_i{*}^T{p}_i*)}^{-1}{D}^{T}X*$

Wherein λ ₁, λ ₂> 0 is weight, and wherein I is unit vector; p _i ^*for the Euclidean distance that above-mentioned steps II calculates;

IV) adopt spatial pyramid matching process, character representation vector Z * is divided into 2 ^l× 2 ^lblock, wherein l=1,2,4, represent the number of different three layers of corresponding spatial pyramid piecemeal respectively; Then for the z of each block in every layer _i ^*carry out maximum pond, keep character representation vector to have certain robustness, i.e. z _i ^*=max{|z _1i ^*|, | z _2i ^*| ..., | z _ji ^*|, obtain the i-th vectorial z of final low-rank coding _i ^*, wherein z _ji ^*i-th element in jth block, j=2 ^l× 2 ^l, obtain last character representation matrix Z*' like this;

V) character representation matrix Z*' is inputted training department and divide the SVM classifier obtained, obtain the class label belonging to image to be classified;

VI) the statistical test classified image number percent of correctly classifying, obtains the nicety of grading that every class testing image is final;

Step II I in step I ii and part of detecting in described training part, calculate the character representation vector of local restriction low-rank coding, the low-rank coding method calculating local restriction comprises the following steps;

A) when dictionary error is zero, E ₁=D-D ₀=0, F norm || || _fcan nuclear norm be replaced || || _*; By existing expression formula $\begin{array}{c}\underset{Z,D}{\min }\left|\right|Z|{|}_{*}+\mathrm{\&lambda;}||D-D_0|{|}_F+tr\&lsqb;Y^T(D-D_0-E_1)\&rsqb;+\\ tr\&lsqb;Y^T(X-DZ-E_2)\&rsqb;+\frac{{\mathrm{\&mu;}}_1}{2}\left|\right|D-D_0-E_1|{|}_F^2+\frac{{\mathrm{\&mu;}}_2}{2}||X-DZ-E_2|{|}_F^2\end{array}$ Optimization is turned into

$\underset{Z}{\min }\left|\right|Z|{|}_F+{\mathrm{\&lambda;}}_1/2||X-DZ|{|}_F^2$

B) combined coding mode is taked in low-rank constraint, on the conforming basis of characteristics of image global structure, makes up the local space similarity of feature, adds local restriction item expression formula after optimizing in above-mentioned steps a becomes:

Wherein λ ₁, λ ₂> 0 is unequal weight; Symbol ⊙ represents that the corresponding element of two vectors is multiplied; p _ifor proper vector x _iwith the Euclidean distance p of dictionary vector D _i:

$p_i=\exp \left(\frac{dist(x_i,D)}{\mathrm{\&sigma;}}\right)$

Dist (x _i, D) and=[dist (x _i, d ₁) ..., dist (x _i, d _k)] ^t, dist (x _i, d _i) be proper vector x _iwith each dictionary vector d _ibetween Euclidean distance; σ is the weight of adjustment local restriction speed;

C) differentiate is carried out to the optimizing expression added in above-mentioned steps b after local restriction item

Finally calculate character representation vector: $Z={({\mathrm{\&lambda;}}_{1}I+{\mathrm{DD}}^T+{\mathrm{\&lambda;}}_1{\mathrm{\&lambda;}}_2\underset{i=1}{\overset{k}{\&Sigma;}}{p_i}^T{p}_i)}^{-1}{D}^{T}X,$ Wherein I is the vector of unit length of k × k.