CN102722732B - Image set matching method based on data second order static modeling - Google Patents

Abstract

The invention provides an image set matching method based on data second order static modeling, which comprises the following steps of firstly providing two image sets S1 and S2 to be matched, analyzing principle components on the training set data to obtain a feature extraction projection matrix, projecting an image sample to be matched in a target feature sub-space, then building a second order static model for two image sets to be matched, then removing the noises in the second order static model, and finally calculating the similarity of the two image sets to be matched and completing the matching of the two image sets according to the similarity. The image set matching method provided by the invention has no prior assumption for the set data distribution form and the set sample scale, the noise data probably existing in the sets can be well tolerated by the method, the algorithm model is intuitive and efficient, and the calculation is simple.

Description

A kind of image collection matching process based on the modeling of data second-order statistic

Technical field

The present invention relates to technical field of computer vision, particularly a kind of image collection matching process based on the modeling of data second-order statistic.

Background technology

Computer vision is one and studies the science how making machine " see ", a vital task of computer vision is the object classification in recognition image or video, in conventional methods where, for object type in static single image, other identifies existing more ripe research, in recent years, the increasing interest of researcher is then caused for the object identification in video.Along with the universal development of the hardware technology such as video camera and mass-memory unit, at much new application scenarios as in the task such as video monitoring, video frequency searching, the great amount of images data of object can be got, thus providing a large amount of samples for the training and testing stage of rear end classification problem, these samples usually carry out modeling with the pattern of image collection and represent.In this kind of identification problem, each set usually comprises and belongs to other multiple image pattern of same object type, and these images cover object apparent changing pattern widely, the change, non-rigid deformation, illumination variation etc. at such as attitude visual angle.

Difficult point based on the classification problem of image collection is, how effectively to portray the distribution with multiple image in modeling set, and according to the information that built model comprehensive utilization multisample provides.In recent years, typical way mainly contains two classes, respectively from parameter type and nonparametric formula two angles to image collection modeling, the former utilizes probability distribution function to represent the sample distribution of image collection usually, and then adopt such as K-L divergence (Kullback-Leibler Divergence, and so on KLD) tolerance calculates the similarity between two probability distribution functions, latter is modeled as linear subspaces or more general non-linearity manifold according to the essential change pattern of sample in image collection, based on this modeling pattern, the problem of sets match classification just can be converted into the problem of subspace or stream shape coupling, thus adopt the various possible similarity measurements flow function on subspace or stream shape and matching strategy to classify.In general, this two class set of current employing builds modeling method jointly all has hypothesis to a certain degree to the form of sample distribution in image collection, and the source of gathering sample in practical problems is normally diversified, when the sample distribution form difference supposed with model is larger, classifying quality just has very large deviation.

Summary of the invention

The present invention is intended at least solve the technical matters existed in prior art, especially innovatively proposes a kind of image collection matching process based on the modeling of data second-order statistic.

In order to realize above-mentioned purpose of the present invention, the invention provides a kind of image collection matching process based on the modeling of data second-order statistic, it comprises the steps:

S1: given two image collection S to be matched ₁and S ₂, training set data is carried out principal component analysis (PCA) and obtains feature extraction projection matrix, image pattern to be matched is projected in target signature subspace;

S2: in described target signature subspace, sets up the second-order statistic model of two image collections to be matched;

S3: remove the noise in described second-order statistic model;

S4: the similarity calculating described two image collections to be matched, completes the coupling of image collection according to described similarity size.

Image collection matching process of the present invention is to collective data distribution form and gather sample size without any a priori assumption, has good tolerance to the noise data that may exist in set.The present invention can realize the image collection/video sequence classification of high-precision and high-stability in true application scenarios, has the potentiality be applied to as in the digital multimedia systems such as Video security monitoring, video frequency searching, the management of large scale network digital album.

In one preferred embodiment of the invention, the covariance matrix of sample set or the related function matrix second-order statistic model as two image collections is got.And the filtering second-order statistic model employing diagonal line perturbation scheme removal singular problem corresponding to be matched two image collections, then adopts the similarity of the determinant logarithm divergence compute matrix in Riemann manifold.

Image collection matching process of the present invention adopts the sample covariance matrix or related function matrix gathered as descriptor, naturally the distribution pattern of data is portrayed, measure the similarity of two set further by the matrix divergence calculated in Riemann manifold, complete the coupling classification of set.This method is to the scale of the distribution form and set sample of gathering sample all without any a priori assumption, and algorithm model is intuitively efficient, calculates easy.

Additional aspect of the present invention and advantage will part provide in the following description, and part will become obvious from the following description, or be recognized by practice of the present invention.

Accompanying drawing explanation

Above-mentioned and/or additional aspect of the present invention and advantage will become obvious and easy understand from accompanying drawing below combining to the description of embodiment, wherein:

Fig. 1 is the process flow diagram of the image collection matching process that the present invention is based on the modeling of data second-order statistic;

Fig. 2 is the image collection second-order statistic modeling method schematic diagram adopted in a kind of preferred implementation of the present invention.

Embodiment

Be described below in detail embodiments of the invention, the example of described embodiment is shown in the drawings, and wherein same or similar label represents same or similar element or has element that is identical or similar functions from start to finish.Being exemplary below by the embodiment be described with reference to the drawings, only for explaining the present invention, and can not limitation of the present invention being interpreted as.

Fig. 1 is the process flow diagram of the image collection matching process that the present invention is based on the modeling of data second-order statistic, and as seen from the figure, this image collection matching process comprises the steps:

First, carry out statistical modeling, specifically:

The first step: pre-service is carried out to sample characteristics, given two image collection S to be matched ₁and S ₂, training set data is carried out principal component analysis (PCA) and obtains feature extraction projection matrix, image pattern to be matched is projected in target signature subspace, in the present embodiment, when carrying out principal component analysis (PCA), adopt the gray-scale value of image pattern as primitive character;

Second step: in target signature subspace, sets up the second-order statistic model of two image collections to be matched, in the present embodiment, gets the covariance matrix of sample set or the related function matrix second-order statistic model as two image collections;

Then, carry out image collection coupling, specifically:

3rd step: remove the noise in second-order statistic model, for the second-order statistic model of original sample obtained in the previous step, carried out Eigenvalues Decomposition, the eigenwert of gained and characteristic of correspondence vector are arranged in order according to size order, setting threshold value also carries out filtering to eigenwert, utilize and calculate new filtering second-order statistic model higher than the eigenwert of threshold value and corresponding proper vector, thus the noise effect of removing in primary statistics model, in the present embodiment, select to arrange and filtering according to order from big to small by the eigenwert of second-order statistic model and with its characteristic of correspondence vector,

4th step: the similarity calculating two image collections to be matched, the coupling of image collection is completed according to similarity size, the filtering second-order statistic model corresponding to be matched two image collections, first diagonal line perturbation scheme is adopted to remove singular problem, then adopt the similarity of the determinant logarithm divergence compute matrix in Riemann manifold, complete sets match classification task according to similarity size value.

In the present embodiment, any secondary statistic can be selected to carry out modeling, be specifically as follows but be not limited to covariance matrix or the related function matrix of sample set, during owing to selecting different second-order statistic model, image collection matching process is identical, therefore, in the preferred embodiment of the present invention, be only described for covariance matrix as second-order statistic model, this can not be interpreted as limitation of the present invention to better the present invention is described.

Fig. 2 is the image collection second-order statistic modeling method schematic diagram adopted in a kind of preferred implementation of the present invention, illustrates that concrete image collection matching process is in conjunction with formula:

First, sample characteristics pre-service, given two image collection S to be matched ₁={ x _i: i=1,2 ..., m} and S ₂={ y _j: j=1,2 ..., n}, adopt the gray-scale value of image pattern as primitive character, training set data is carried out principal component analysis (PCA) and obtains feature extraction projection matrix, projected to by image pattern to be matched in target signature subspace, feature extraction formula is as follows:

x _i′＝Wx _i

y _j′＝Wy _j （1）

Wherein, W is that training set data is gone to school the principal component projection matrix that acquistion arrives, x _i', y _j' represent for the sample after feature extraction;

Then, set up second-order statistic model, in the proper subspace of previous step gained, the sample mean vector of computed image set, is normalized sample, calculates the covariance matrix of sample set afterwards, as the second-order statistic model of set, particularly, for image collection S ₁and S ₂, calculate its mean vector respectively as follows:

${\stackrel{\‾}{x}}^{\′}=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{x}_i^{\′}$

（2）

${\stackrel{\‾}{y}}^{\′}=\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{y}_j^{\′}$

Corresponding covariance matrix second-order statistic is as follows:

$C_1=\frac{1}{m-1}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}(x_i^{\′}-{\stackrel{\‾}{x}}^{\′}){(x_i^{\′}-{\stackrel{\‾}{x}}^{\′})}^T---\left(3\right)$

$C_2=\frac{1}{n-1}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}(y_{\mathrm{ij}}^{\′}-{\stackrel{\‾}{y}}^{\′}){(y_{\mathrm{ij}}^{\′}-{\stackrel{\‾}{y}}^{\′})}^T---\left(4\right)$

Again, for original sample covariance matrix C obtained in the previous step ₁, C ₂, carried out Eigenvalues Decomposition, the eigenwert of gained and characteristic of correspondence vector arranged, with Matrix C according to order from big to small ₁for example, feature decomposition is:

C ₁=U ₁∑ ₁U ₁ ^T （5）

Wherein, ∑ ₁for the eigenvalue matrix after the descending sequence of value, U ₁for characteristic of correspondence vector matrix.Select certain threshold value δ to carry out filtering to eigenvalue matrix, the eigenwert picked out higher than threshold value δ forms diagonal matrix sigma ₁', be expressed as U after its corresponding eigenvectors matrix extracts ₁', calculating new filtering covariance matrix is:

C ₁′=U ₁′∑ ₁′U ₁′ ^T （6）

Thus the noise effect of removing in primary statistics model, for original covariance matrix C ₂also doing same Denoising disposal, to obtain filtering covariance matrix be C ' ₂.

Finally, for the filtering covariance matrix C ' that to be matched two image collections are corresponding ₁, C ' ₂, first adopt diagonal line perturbation scheme to remove singular problem:

$C_1^{*}=C_1^{\′}+{\mathrm{\λ}}_{1}I---\left(7\right)$

$C_2^{*}=C_2^{\′}+{\mathrm{\λ}}_{2}I---\left(8\right)$

Wherein, I is unit matrix, λ ₁=10 ^-3× trace (C ' ₁), λ ₂=10 ^-3× trace (C ' ₂).Adopt the similarity of the determinant logarithm divergence compute matrix in Riemann manifold afterwards, formula specific as follows:

$d_{\mathrm{ld}}(C_1^{*},C_2^{*})=\mathrm{trace}\left(C_1^{*}{\left(C_2^{*}\right)}^{-1}\right)-\log \left|C_1^{*}{\left(C_2^{*}\right)}^{-1}\right|-p---\left(9\right)$

Wherein, p is covariance matrix or covariance matrix exponent number, namely the line number of matrix or columns, complete sets match classification task according to the above-mentioned similarity size value calculated.

Present embodiment adopts the sample covariance matrix of set as descriptor, naturally portrays the distribution pattern of data, measures the similarity of two set further by the matrix divergence calculated in Riemann manifold, completes the coupling classification of set.This algorithm model is intuitively efficient, calculates easy, and to the distribution form of set sample and the scale of set sample all without any a priori assumption, has good tolerance to the noise data that may exist in set.

In the description of this instructions, specific features, structure, material or feature that the description of reference term " embodiment ", " some embodiments ", " example ", " concrete example " or " some examples " etc. means to describe in conjunction with this embodiment or example are contained at least one embodiment of the present invention or example.In this manual, identical embodiment or example are not necessarily referred to the schematic representation of above-mentioned term.And the specific features of description, structure, material or feature can combine in an appropriate manner in any one or more embodiment or example.

Although illustrate and describe embodiments of the invention, those having ordinary skill in the art will appreciate that: can carry out multiple change, amendment, replacement and modification to these embodiments when not departing from principle of the present invention and aim, scope of the present invention is by claim and equivalents thereof.

Claims (3)

1., based on an image collection matching process for data second-order statistic modeling, it is characterized in that, comprise the steps:

S1: given two image collection S to be matched ₁and S ₂, training set data is carried out principal component analysis (PCA) and obtains feature extraction projection matrix, projected in target signature subspace by image pattern to be matched, wherein step S1 is specially:

Pre-service is carried out to sample characteristics, given two image collection S to be matched ₁={ x _i: i=1,2 ..., m} and S ₂={ y _j: j=1,2 ..., n}, training set data is carried out principal component analysis (PCA) and obtains feature extraction projection matrix, projected to by image pattern to be matched in target signature subspace, feature extraction formula is as follows:

x′ _i＝Wx _i

y′ _j＝Wy _j，

Wherein, W is that training set data is gone to school the principal component projection matrix that acquistion arrives, x ' _i, y ' _jfor the sample after feature extraction represents;

S2: set up second-order statistic model, in the proper subspace of above-mentioned gained, the sample mean vector of computed image set, sample is normalized, calculates the covariance matrix of sample set afterwards, as the second-order statistic model of set, it specifically comprises, for image collection S ₁and S ₂, calculate its mean vector respectively as follows:

${\stackrel{\‾}{x}}^{\′}=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{x}_i^{\′}$

${\stackrel{\‾}{y}}^{\′}=\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{y}_j^{\′},$

Corresponding covariance matrix second-order statistic is as follows:

$C_1=\frac{1}{m-1}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}(x_i^{\′}-{\stackrel{\‾}{x}}^{\′}){(x_i^{\′}-{\stackrel{\‾}{x}}^{\′})}^T,$

$C_2=\frac{1}{n-1}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}(y_{\mathrm{ij}}^{\′}-{\stackrel{\‾}{y}}^{\′}){(y_{\mathrm{ij}}^{\′}-{\stackrel{\‾}{y}}^{\′})}^T;$

S3: remove the noise in described second-order statistic model, specifically comprise:

For the original sample covariance matrix C obtained ₁, C ₂, carried out Eigenvalues Decomposition, the eigenwert of gained and characteristic of correspondence vector arranged, with Matrix C according to order from big to small ₁for example, feature decomposition is:

C ₁＝U ₁Σ ₁U ₁ ^T，

Wherein, Σ ₁for the eigenvalue matrix after the descending sequence of value, U ₁for characteristic of correspondence vector matrix, and select certain threshold value δ to carry out filtering to eigenvalue matrix, the eigenwert picked out higher than threshold value δ forms diagonal matrix Σ ₁', be expressed as U after its corresponding eigenvectors matrix extracts ₁', calculating new filtering covariance matrix is:

C ₁′＝U ₁′Σ ₁′U ₁ ^′T，

Thus the noise effect of removing in primary statistics model, for original covariance matrix C ₂also doing same Denoising disposal, to obtain filtering covariance matrix be C ' ₂;

S4: the similarity calculating described two image collections to be matched, the size according to described similarity completes the coupling of image collection, and the computing method of the similarity of described two image collections to be matched comprise:

For the filtering covariance matrix C ' that to be matched two image collections are corresponding ₁, C ' ₂, first adopt diagonal line perturbation scheme to remove singular problem:

$C_1^{*}=C_1^{\′}+{\mathrm{\λ}}_{1}I,$

$C_2^{*}=C_2^{\′}+{\mathrm{\λ}}_{2}I,$

Wherein, I is unit matrix, λ ₁=10 ^-3× trace (C ' ₁) λ ₂=10 ^-3× trace (C ' ₂),

And adopt the similarity of the determinant logarithm divergence compute matrix in Riemann manifold, formula specific as follows:

$d_{\mathrm{ld}}(C_1^{*},C_2^{*})=\mathrm{trace}\left(C_1^{*}{\left(C_2^{*}\right)}^{-1}\right)-\log \left|C_1^{*}{\left(C_2^{*}\right)}^{-1}\right|-p,$

Wherein, p is covariance matrix or covariance matrix exponent number, i.e. the line number of matrix or columns.

2. as claimed in claim 1 based on the image collection matching process of data second-order statistic modeling, it is characterized in that, when carrying out principal component analysis (PCA), adopt the gray-scale value of image pattern as primitive character.

3. as claimed in claim 1 based on the image collection matching process of data second-order statistic modeling, it is characterized in that, arrange according to order from big to small by the eigenwert of second-order statistic model and with its characteristic of correspondence vector.