CN103426006A - Self-adaption multi-feature fusion image feature learning method - Google Patents

Abstract

The invention discloses a self-adaption multi-feature fusion image feature learning method. The method comprises the steps of first respectively extracting image visual features of various types according to the characteristic that an L1 norm has automatic data sample selection, then respectively adopting data sparse expression constrained by the L1 norm for every visual feature, establishing a directed L1 map, minimizing reestablishment errors of various feature data on a low-dimensional space, introducing feather weight vectors to perform global coordinate alignment operation on all the features, obtaining an objective function of the self-adaption multi-feature fusion image feature learning method, finally solving the objective function, and obtaining the optimal image low-dimensional feature expression by learning on the basis of original multiple image features. The image low-dimensional feature expression obtained by the method serves as input of a nearest neighbor classifier algorithm; compared with other algorithms on a public image data set Core15K and NUS-WIDE-OBJECT, the method improve the classification accuracy rates respectively by 5% and 2%, and accordingly the image features obtained by the method have high expression capability.

Description

The characteristics of image learning method of the many Fusion Features of a kind of self-adaptation

Technical field

The present invention relates to sparse expression, figure study, feature learning, relate in particular to the characteristics of image study of many Fusion Features.

Background technology

In computer vision, pattern-recognition and information retrieval field, owing to inevitably being faced with the semantic gap problem, low-level image feature can't be expressed high-layer semantic information fully exactly; Therefore, in order to express as far as possible all-sidedly and accurately the object that will describe, often can, to the polytype visual signature of this object extraction, from different visual angles, it be expressed.For example: in Images Classification, to coloured image, often can extract its color histogram feature, profile moment characteristics, textural characteristics, sift point feature etc.From traditional to express object based on single type feature different, these characteristics, be distributed in different feature spaces, has its inherent statistical property and real physical significance, and its ability to express (or resolving ability) is the property of there are differences also.In order to take full advantage of the characteristic of different characteristic, accelerating algorithm is processed, and usually needs to carry out many feature learnings according to these many features, obtains optimum low-dimensional feature representation.

Researchers wish by utilizing many feature learnings fully to excavate out the complementary information between different characteristic, thereby obtain a low-dimensional nested result better than the single feature learning of direct utilization.In existing algorithm, the spectrum nested algorithm has very important status.Then, traditional spectrum nested algorithm is all that tentation data is to be distributed in same vector space, thereby can't be applied directly in many feature learnings.The solution existed at present has two: one is that different features first is stitched together, and is then composed nested study.Obvious this way lacks clear and definite physical significance support.Another way is to use the nested study of distributed spectrum (DSE) algorithm, and different features is learnt respectively.Yet many characteristics original in this method are sightless for final learning process, can't fully effectively utilize the complementary information between different characteristic.Under the impact of the many feature learnings universal model proposed people such as Long, the people such as Xia have proposed characteristic spectrum nested algorithm more than, in different characteristic study, obtain simultaneously a low-dimensional, enough level and smooth nested, thereby fully effectively utilize the complementary information between different characteristic.

Yet, because all being based on spectral graph theory, overwhelming majority various visual angles learning algorithm develops, and a core of its algorithm is the adjacency matrix (similar matrix) that utilizes space structure relation between various criteria construction expression datas.The Usefulness Pair of the construction algorithm of adjacency matrix plays vital effect in the performance of whole algorithm.Now widely used two kinds of adjacency matrix construction methods: based on k nearest neighbor and the patterning process based on thermonuclear (heat kernel), be not the patterning process of data adaptive.The former need to set in advance neighboring node number k, and the latter need to arrange nuclear parameter.In actual applications, the setting of these two parameters often by virtue of experience or utilize the mode of cross validation to select best parameter, lacks simple directly method and obtains optimal parameter.Meanwhile, the algorithm of this class based on figure, be again highstrung for the selection of these two parameters sometimes.The quality that parameter is selected directly has influence on the performance of algorithm.Therefore, several existing work noted earlier all is subject to lacking the puzzlement of data adaptive problem.How building the figure with data adaptive, improve the adaptivity of algorithm, is a problem in the urgent need to address.

In the last few years, the L1 norm was subject to extensive concern in sparse expression and compressed sensing research field.Be different from traditional L2 norm constraint, the L1 norm constraint can, so that the solution of problem becomes sparse, have stronger anti-noise ability.According to the people's such as Cheng research, show, L1 figure has three key properties: (1) noise resisting ability is strong; (2) sparse property; (3) neighborhood relationships of data adaptive.

From the above, in order to overcome traditional poor problem of many feature learnings algorithm adaptivity based on figure, the characteristics of image learning method of the many Fusion Features of self-adaptation that this paper proposes is to have according to the L1 norm characteristic that the automaticdata sample is selected, at first extract respectively polytype Image Visual Feature from image, then to each visual signature, adopt respectively the Sparse of L1 norm constraint to express, be built with to L1 and scheme, the good characteristic that L1 figure is had is dissolved in many feature learnings algorithm, and descend the reconstruction error of data on lower dimensional space under the various features of change most, secondly the introduced feature weight vectors is done the world coordinates alignment operation to all features, obtain the objective function of the image feature selection method of the many Fusion Features of self-adaptation, last this objective function of Optimization Solution, can obtain adaptively the low-dimensional feature representation of image optimum from original multiple characteristics of image basis learning.

Summary of the invention

The objective of the invention is to overcome the deficiencies in the prior art, propose the characteristics of image learning method of the many Fusion Features of a kind of self-adaptation.

The characteristics of image learning method of the many Fusion Features of self-adaptation is to comprise the steps:

1) extract respectively the Image Visual Feature of K type from N width image

Wherein

It is the visual signature of v type extracted on i width image;

2) there is the characteristic of automaticdata feature selecting according to the L1 norm, to each visual signature, adopt respectively the Sparse of L1 norm constraint to express:

$\min {\left|\right|x_i^v-X^v{w}_i^v\left|\right|}_F^2+\mathrm{\λ}{\left|\right|w_i^v\left|\right|}_1$ 1

$s.t.w_{\mathrm{ii}}^v=0$

According to what calculate

Be built with the g to L1 figure ^v={ X ^v, W ^v, the type characteristics of image X wherein ^vThe weight on the ，Tu De limit, summit of pie graph

Descend most the reconstruction error of data on lower dimensional space under the various features of change:

$\underset{Y^v}{\min }{\mathrm{\Σ}}_{i=1}^N{\left|\right|Y_i^v-Y^v{w}_i^v\left|\right|}_2^2---2$

3) introduced feature weight vectors α=[α ₁..., α _K], all features are done to the world coordinates alignment operation, the characteristics of image learning method objective function that obtains the many Fusion Features of self-adaptation is as follows:

$\underset{\mathrm{\α},Y}{\min }{\mathrm{\Σ}}_{v=1}^K{\left({\mathrm{\α}}_v\right)}^r{{\mathrm{\Σ}}_{i=1}^N\left|\right|Y-Y{w}_i^v\left|\right|}_2^2$ 3

$s.t.{\mathrm{YY}}^T=I,{\mathrm{\Σ}}_{v=1}^K{\mathrm{\α}}_v=1,{\mathrm{\α}}_v\≥0,\∀v$

Wherein, r is evolution number of times parameter;

4) top objective function 3 is optimized and solves, adopt Lagrangian multiplication, obtain computing formula as follows:

${\mathrm{\α}}_v=\frac{(1/\mathrm{tr}{\left({\mathrm{YG}}^v{Y}^T\right)}^{1/(r-1)})}{{\mathrm{\Σ}}_{v=1}^K(1/\mathrm{tr}{\left({\mathrm{YG}}^v{Y}^T\right)}^{1/(r-1)})}---4$

Y＝select_svd(G,d) 5

Wherein: tr () is the computing of Matrix Calculating mark, G ^v=(I-W ^v) (I-W ^v) ^T,

Select_svd (G, d) means to G is done Eigenvalues Decomposition and gets front d the corresponding proper vector of eigenwert of eigenwert minimum, formation matrix, the low-dimensional feature representation of the image optimum that Y is from original multiple characteristics of image basis to learn to obtain.

The beneficial effect that the present invention compared with prior art has is, algorithm has data adaptive, fast operation, and the characteristics of image of learning to obtain has stronger ability to express, can effectively improve successive image sorting algorithm accuracy rate.

The accompanying drawing explanation

Fig. 1 is the part sample image on the Core5K image data set;

Fig. 2 is the part sample image on the NUS-WIDE-OBJECT data set;

Fig. 3 be on the Corel5K image data set each algorithm pattern as the classification accuracy comparing result;

Fig. 4 be on the NUS-WIDE-OBJECT image data set each algorithm pattern as the classification accuracy comparing result;

Fig. 5 is the neighbour's statistics of the constructed L1 figure of this method on the Corel5K image data set;

Fig. 6 is the neighbour's statistics of the constructed L1 figure of this method on the NUS-WIDE-OBJECT image data set.

Embodiment

The characteristics of image learning method of the many Fusion Features of self-adaptation comprises the steps:

1) extract respectively the Image Visual Feature of K type from N width image

Wherein

It is the visual signature of v type extracted on i width image;

2) there is the characteristic of automaticdata feature selecting according to the L1 norm, to each visual signature, adopt respectively the Sparse of L1 norm constraint to express:

$\min {\left|\right|x_i^v-X^v{w}_i^v\left|\right|}_F^2+\mathrm{\λ}{\left|\right|w_i^v\left|\right|}_1$ 1

$s.t.w_{\mathrm{ii}}^v=0$

According to what calculate Be built with the g to L1 figure ^v={ X ^v, W ^v, the type characteristics of image X wherein ^vThe weight on the ，Tu De limit, summit of pie graph

Descend most the reconstruction error of data on lower dimensional space under the various features of change:

$\underset{Y^v}{\min }{\mathrm{\Σ}}_{i=1}^N{\left|\right|Y_i^v-Y^v{w}_i^v\left|\right|}_2^2---2$

3) introduced feature weight vectors α=[α ₁..., α _K], all features are done to the world coordinates alignment operation, the characteristics of image learning method objective function that obtains the many Fusion Features of self-adaptation is as follows:

$\underset{\mathrm{\α},Y}{\min }{\mathrm{\Σ}}_{v=1}^K{\left({\mathrm{\α}}_v\right)}^r{{\mathrm{\Σ}}_{i=1}^N\left|\right|Y-Y{w}_i^v\left|\right|}_2^2$ 3

$s.t.{\mathrm{YY}}^T=I,{\mathrm{\Σ}}_{v=1}^K{\mathrm{\α}}_v=1,{\mathrm{\α}}_v\≥0,\∀v$

Wherein, r is evolution number of times parameter;

4) top objective function 3 is optimized and solves, adopt Lagrangian multiplication, can obtain computing formula as follows:

${\mathrm{\α}}_v=\frac{(1/\mathrm{tr}{\left({\mathrm{YG}}^v{Y}^T\right)}^{1/(r-1)})}{{\mathrm{\Σ}}_{v=1}^K(1/\mathrm{tr}{\left({\mathrm{YG}}^v{Y}^T\right)}^{1/(r-1)})}---4$

Y＝select_svd(G,d) 5

Wherein: tr () is the computing of Matrix Calculating mark, G ^v=(I-W ^v) (I-W ^v) ^T,

Select_svd (G, d) means to G is done Eigenvalues Decomposition and gets front d the corresponding proper vector of eigenwert of eigenwert minimum, formation matrix, the low-dimensional feature representation of the image optimum that Y is from original multiple characteristics of image basis to learn to obtain.

Will be on two common image test data set Corel5K and NUS-WIDE-OBJECT data set, the impact of the characteristics of image learning algorithm (being designated as AMVL) of the many Fusion Features of contrast test self-adaptation on the Images Classification accuracy rate.The characteristics of image of announcing according to the NUS-WIDE-OBJECT data set, respectively these two concentrated images of data are extracted to following 5 kinds of features: LAB color histogram (64 dimension), hsv color autocorrelogram (144 dimension), edge orientation histogram (73 dimension), wavelet texture (128 dimension) and piecemeal color moment feature (225 dimension) obtain 5 different visual angles features.Test the specifying information of data set used in Table 1.

Table 1. embodiment data set information

Use the PCA(principal component analysis (PCA)), the classification of SRC(sparse expression), the mono-feature adaptive learning of ASVL() compose from various visual angles nested study with MSE() method of testing as a comparison.Wherein, PCA and SRC are two kinds of unsupervised learning algorithms at present commonly used, and ASVL is single feature learning special case of the AMVL algorithm that proposes of this paper, and MSE is a various visual angles learning algorithm.Adopt nearest neighbor classifier as Images Classification benchmark algorithm, contrast each algorithm performance.

In force, the needed parameter lambda of this method=0.01 is set, r=2, and all adopt parameter traversal mode to obtain optimum parameter setting for other algorithms, then contrast in the situation that acquire the characteristics of image of different dimensions, the performance change situation of each algorithm.

The characteristics of image learning method that Fig. 3 and Fig. 4 have showed the many Fusion Features of self-adaptation that propose and contrast algorithm be the classification accuracy comparing result above image data set at two.Can find out that from the result of Fig. 3 and Fig. 4 the classification accuracy of the resulting characteristics of image of this method on most characteristic dimension all can be much better than other contrast algorithms.In characteristic dimension, be 60 o'clock, Corel5K is upper, and the method AMVL proposed improves 5% than other algorithms, at NUS-WIDE-OBJECT, improves 2%.Fig. 5 and Fig. 6 have showed neighbour's statistics of the L1 figure that on Corel5K and NUS-WIDE-OBJECT image data set this method is constructed.This method can be selected neighbour's number adaptively as can be known from Fig. 5 and Fig. 6.

Claims (1)

1. the characteristics of image learning method of the many Fusion Features of self-adaptation, is characterized in that comprising the steps:

1) extract respectively the Image Visual Feature of K type from N width image

Wherein

It is the visual signature of v type extracted on i width image;

2) there is the characteristic of automaticdata feature selecting according to the L1 norm, to each visual signature, adopt respectively the Sparse of L1 norm constraint to express:

$\min {\left|\right|x_i^v-X^v{w}_i^v\left|\right|}_F^2+\mathrm{\λ}{\left|\right|w_i^v\left|\right|}_1$ 1

$s.t.w_{\mathrm{ii}}^v=0$

According to what calculate

Be built with the g to L1 figure ^v={ X ^v, W ^v, the type characteristics of image X wherein ^vThe weight on the ，Tu De limit, summit of pie graph Descend most the reconstruction error of data on lower dimensional space under the various features of change:

$\underset{Y^v}{\min }{\mathrm{\Σ}}_{i=1}^N{\left|\right|Y_i^v-Y^v{w}_i^v\left|\right|}_2^2---2$

3) introduced feature weight vectors α=[α ₁..., α _K], all features are done to the world coordinates alignment operation, the characteristics of image learning method objective function that obtains the many Fusion Features of self-adaptation is as follows:

$\underset{\mathrm{\α},Y}{\min }{\mathrm{\Σ}}_{v=1}^K{\left({\mathrm{\α}}_v\right)}^r{{\mathrm{\Σ}}_{i=1}^N\left|\right|Y-Y{w}_i^v\left|\right|}_2^2$ 3

$s.t.{\mathrm{YY}}^T=I,{\mathrm{\Σ}}_{v=1}^K{\mathrm{\α}}_v=1,{\mathrm{\α}}_v\≥0,\∀v$

Wherein, r is evolution number of times parameter;

4) top objective function 3 is optimized and solves, adopt Lagrangian multiplication, obtain computing formula as follows:

${\mathrm{\α}}_v=\frac{(1/\mathrm{tr}{\left({\mathrm{YG}}^v{Y}^T\right)}^{1/(r-1)})}{{\mathrm{\Σ}}_{v=1}^K(1/\mathrm{tr}{\left({\mathrm{YG}}^v{Y}^T\right)}^{1/(r-1)})}---4$

Y＝select_svd(G,d) 5

Wherein: tr () is the computing of Matrix Calculating mark, G ^v=(I-W ^v) (I-W ^v) ^T,

Select_svd (G, d) means to G is done Eigenvalues Decomposition and gets front d the corresponding proper vector of eigenwert of eigenwert minimum, formation matrix, the low-dimensional feature representation of the image optimum that Y is from original multiple characteristics of image basis to learn to obtain.

Images