CN103177264B - The image classification method that view-based access control model dictionary Global Topological is expressed - Google Patents

Abstract

The invention discloses the image classification method that a kind of view-based access control model dictionary Global Topological is expressed, including training and two processes of identification, specifically include step: the target image having marked classification is carried out feature extraction, the feature extracted is carried out Global Topological coding on visual dictionary, coding result is trained and models；The image of unknown classification is carried out feature extraction, the feature extracted is carried out Global Topological coding on visual dictionary, be input to train the model obtained by coding result, it is thus achieved that the classification of target image.Expressing the manifold for image due to Global Topological to express and have invariance, therefore the present invention uses the Global Topological of view-based access control model dictionary to express the precision improving image recognition, and this technology has great importance for the understanding of dynamic image.The present invention is expressed by the Global Topological of study visual dictionary, can identify the classification of image accurately, and this technology can be widely applied to safety verification, the field such as web search and digital entertainment.

Description

The image classification method that view-based access control model dictionary Global Topological is expressed

Technical field

The present invention relates to area of pattern recognition and computer vision field, open up particularly to a kind of view-based access control model dictionary overall situation Flutter the image classification method of expression.

Background technology

In computer vision, image classification is basic studying a question.In the past few decades, researchers taste Examination goes the physiology according to human cognitive according to studying classification problem.And as one of them important foundation, researchers Think that the fundamental that image is classified is the potential manifold expression of image.This manifold is expressed and is come from neuron coding, And researchers propose utilize interneuronal topological relation to obtain certain topological invariance, and this invariance for figure The manifold of picture is expressed and is served important function.Therefore, vision topology is to connect neuron coding and the bridge of manifold expression, and The feature coding that this relationship mechanism has been widely applied in image classification is expressed with manifold.Numerous studies in the last few years Show, describe vision topology and can preferably reconstruct local feature so that the manifold of image is expressed more smooth, to reach to promote The purpose of image classification accuracy.Therefore, the vision topology in the classification of research image is necessary.

In recent years, increasing researcher considered topological structure in image is classified, and especially compiled at local feature Code part.Topological structure is used for feature coding and comes from topology expression neutral net, and researchers in a network the earliest Utilize the topological relation between neuron to come more accurately neuron is encoded.Expressed by this topology and inspired, many Local feature is encoded by the topological relation that researcher starts to be devoted to utilize the vision word in visual dictionary model.Mesh Front topological structure mainly considers melting of the topological relation of two vision word, such as two closest vision word Close, the distance of two vision word and angular relationship etc., and this relation of topological structure two-by-two achieves certain effect.

But, the theory expressed according to manifold, a limitation of this topological structure two-by-two is that it can not good table Levy the manifold invariance of feature space.Accordingly, it is considered to one can allow manifold express more smooth higher order topology structure and just show Must be highly desirable to, such as, consider the global Topological Structure between multiple vision word.Another one limitation is due to vision Single contamination forms exponential increase along with the increase of vision word number, it is considered to the topological structure of this high-order is very challenging 's.Therefore, in the present invention, it will thus provide the image classification method that a kind of view-based access control model dictionary Global Topological is expressed overcomes above Two limitation, to reach better image recognition effect.

Summary of the invention

In order to solve the problem that prior art exists, the present invention provides the figure that a kind of view-based access control model dictionary Global Topological is expressed As sorting technique.

The image classification method that a kind of view-based access control model dictionary Global Topological that the present invention proposes is expressed, comprises the following steps:

Step 1, gathers several training images, and several training images carry out local sampling respectively, and in the local obtained Extract Scale invariant features transform (SIFT) feature on sampling block, thus obtain the SIFT feature set of training image；

Step 2, carries out cluster and generates multiple cluster centres, and with described cluster centre be the SIFT feature set obtained Vision word composition visual dictionaryWherein, C represents visual dictionary, its vision tieed up by M D Word c_iComposition,Represent the subspace that M point in D dimension space forms；

Step 3, carries out Global Topological coding to each SIFT feature of every width training image；

Step 4, all SIFT feature to every width training imageGlobal Topological coding V₁, V₂..., V_NCarry out maximum aggregation operator, generate the image expression F of this training image；

Step 5, sends into the image expression of all training images in grader and is trained, generate training pattern；

Step 6, similar to described step 1, every image to be identified is carried out local sampling, and at the local sampling obtained Extract Scale invariant features transform SIFT feature on block, obtain the SIFT feature set of every image to be identified；

Step 7, based on the visual dictionary being made up of vision word obtained in described step 2, utilizes described step 3 to often Each SIFT feature of image to be identified carries out Global Topological coding；

Step 8, similar to described step 4, the Global Topological of all SIFT feature of every image to be identified is coded into The maximum aggregation operator of row, generates the image expression of image to be identified；

Step 9, the image expression of the image to be identified described step 8 obtained sends into the training mould that described step 5 generates Type is tested, thus obtains the other recognition result of target class in image to be identified.

The method according to the invention, in complex situations, the topological structure of image still can be by robust for target image Ground obtains, thus carries out the robust control policy of image.In intelligent visual surveillance system, this technology can be used for identification monitoring system field The classification of target in scape so that monitoring system can well identify the behavior that target object is current.

Accompanying drawing explanation

Fig. 1 is the image classification method flow chart that view-based access control model dictionary Global Topological of the present invention is expressed.

Detailed description of the invention

For making the object, technical solutions and advantages of the present invention clearer, below in conjunction with specific embodiment, and reference Accompanying drawing, the present invention is described in more detail.

Fig. 1 is the image classification method flow chart that view-based access control model dictionary Global Topological of the present invention is expressed, as it is shown in figure 1, this The image classification method that the view-based access control model dictionary Global Topological that invention is proposed is expressed comprises the following steps:

Step 1, gathers several training images, and several training images carry out local sampling, local sampling block pixel respectively Size can be such as 16*16, and on the local sampling block obtained extract Scale invariant features transform SIFT feature, thus Obtain the SIFT feature set of training image；

Step 2, carries out cluster and generates multiple cluster centres, and with described cluster centre be the SIFT feature set obtained Vision word composition visual dictionaryWherein, C represents visual dictionary, its vision tieed up by M D Word c_iComposition,Represent the subspace that M point in D dimension space forms；

In this step, described cluster can use clustering algorithm commonly used in the prior art, such as k-means clustering algorithm.

Step 3, carries out Global Topological coding to each SIFT feature of every width training image；

Described step 3 farther includes three sub-steps:

Step 3.1, calculates the dependency between vision word in described visual dictionary；

In an embodiment of the present invention, utilize a kind of independent increment formula algorithm based on distance and angle obtain vision word it Between dependency, before described independent increment formula algorithm based on distance and angle is introduced, first define several parameter:

At visual dictionaryIn, for vision word c_iIf, c_jWith c_iIt is relevant, Then define them in feature space, form a cone:

$c_i\&RightArrow;c_j=\left\{y\right|{\left|\right|y-c_i\left|\right|}_2^2\&le;{\mathrm{\&mu;}}_{i,}\mathrm{\&Delta;}(c_{i,}y,c_j)<\mathrm{\&theta;}\}---\left(1\right)$

Wherein, c_i→c_jRepresent c_iAnd c_jCone in the feature space formed between two vision word, y is appointing in cone Anticipate a characteristic point, y} represents this cone,Represent y and c_iBetween Euclidean distance, μ_iIt is c_iAfter clustering with k-means Belong to c_iCharacteristic point between maximum Euclidean distance, θ be control cone angle；Δ(c_i, y, c_j) it is vector y-c_iWith vector c_j- c_iBetween angle, be defined as follows:

$\mathrm{\&Delta;}(c_{i,}y,c_j)=\arccos \frac{\&lang;y-c_{i,}{c}_j-c_i\&rang;}{{\left|\right|y-c_i\left|\right|}_2\&CenterDot;{\left|\right|c_j-c_i\left|\right|}_2}---\left(2\right)$

Wherein, < y-c_i, c_j-c_i> represent vector y-c_iWith vector c_j-c_iBetween inner product, " " represents between two vectors Dot product, | | | |₂Represent two norms of vector.

Next the angle between the vector that the cone defined according to formula (1) and formula (2) define, it is possible to based on distance With the independent increment algorithm of angle obtains the dependency between vision word, the base of described independent increment algorithm based on distance and angle This thinking is that each relevant vision word occupies an independent cone region.Specifically, described based on away from walk-off angle The independent increment algorithm of degree includes:

Firstly, for a certain vision word c_i, all of vision word arrives c according to it_iEuclidean distance carry out from the near to the remote Sequence, wherein, using nearest vision word as initial relevant vision word c_i1；

Secondly, except vision word c_iOther outer vision word check according to distance and angle criterion one by one, if A certain vision word c_jBe satisfied by described distance and angle criterion through inspection, then this vision word is marked as multi view list Word c_ij, and add set S to_i, finally give vision word c_iThe set S of multi view word_i；

Wherein, described distance criterion is defined as:

||c_i-c_ij||₂＜ τ | | c_i-c_i1||₂ (3)

Wherein, τ is used for controlling c_iEuclidean distance to other vision word, say, that distance c_iThe most remote vision word Do not account for；

Described angle rule definition is:

Δ(c_i, c_ij, c_k) ＞ θ, $\&ForAll;\left\{c_k\right\}\&Element;S_i---\left(4\right)$

Wherein, S_iIt it is vision word c_iThe set of multi view word, and c_kRepresentative has been added to gather S_iIn phase Close vision word.

Finally, to all of vision word traversal processing, i.e. can get the related words set S of visual dictionary C:

Wherein, D represents the dimension of each vision word, S_iIt it is vision word c_iThe set of related words, Q_iIt is S_iIn Vision word number.

Step 3.2, utilizes dependency between the vision word obtained in described step 3.1 to each SIFT feature K the vision word of neighbour sets up Global Topological model；

In this step, it is assumed that the SIFT feature collection obtained is extracted for every width training image and is combined intoWherein, D is intrinsic dimensionality, and N is Characteristic Number.Then for each feature x_j, its overall situation is opened up Flutterring model is:

Δ (C, x_j, S) and=[Δ (c₁, x_j, S₁) ..., Δ (c_M, x_j, S_M)] (7)

$\&ForAll;\mathrm{\&Delta;}(c_i,x_j,S_i)=[\mathrm{\&Delta;}(c_i,x_j,c_{i1}),...,\mathrm{\&Delta;}(c_i,x_j,c_{{\mathrm{iQ}}_i})]\&Element;R^{1\&times;Q_i}---\left(8\right)$

Wherein, Δ (c_i, x_j, S_i) it is vector x_j-c_iWith set S_iIn each multi view word c_ijWith c_iConstitute to Amount c_ij-c_iBetween angle:

$\mathrm{\&Delta;}(c_i,x_j,S_i)=[\mathrm{\&Delta;}(c_i,x_j,c_{i1}),...,\mathrm{\&Delta;}(c_i,x_j,c_{{\mathrm{iQ}}_i})],$

$\mathrm{\&Delta;}(c_i,x_j,c_{\mathrm{ij}})=\arccos \frac{\&lang;x_j-x_i,c_{\mathrm{ij}}-c_i\&rang;}{{\left|\right|x_j-c_i\left|\right|}_2\&CenterDot;{\left|\right|c_{\mathrm{ij}}-c_i\left|\right|}_2},$ <x_j-c_i, c_ij-c_i> represent vector x_j-c_iAnd vector c_ij-c_iBetween inner product, " " represents the dot product between two vectors, | | | |₂Represent two norms of vector.

Step 3.3, the Global Topological model utilizing described step 3.2 to obtain carries out Global Topological volume to each SIFT feature Code.

In this step, utilize following formula to each SIFT feature x_iCarry out Global Topological coding:

$\underset{V_i}{\arg \min }{\left|\right|x_i-{\mathrm{SV}}_i^T\left|\right|}_2^2+[\mathrm{\&lambda;T}\left(V_i\right)+\mathrm{\&alpha;}{\left|\right|V_i\&CenterDot;\mathrm{\&Delta;}(C,x_i,S)\left|\right|}_2^2]---\left(9\right)$

Wherein, λ is penalty term coefficient, T (V_i) it is to V_iAny form of penalty term, α is the punishment of Global Topological model , " " represents the dot product between two vectors, V_iIt is x_iCoding on S, is defined as follows:

Wherein, (v_k1)_iIt is this local feature x_iIn vision word c_kFirst multi view word c_k1On response, depend on Secondary analogize.

Then, formula (9) is optimized and solves, i.e. can get SIFT feature x_iGlobal Topological coding V_i。

Step 4, all SIFT feature to every width training imageGlobal Topological coding V₁, V₂..., V_NCarry out maximum aggregation operator, generate the image expression F of this training image；

In this step, carry out maximum aggregation operator according to the following formula, the image expression F of generation training image:

F=max_column[V₁ ^T, V₂ ^T..., V_N ^T]^T (12)

Wherein, max_columnRepresent the every string for matrix and only retain its maximum.

Step 5, sends into the image expression of all training images in grader and is trained, generate training pattern；

In this step, described grader can use grader commonly used in the prior art, such as support vector machine.

Step 6, similar to described step 1, every image to be identified is carried out local sampling, and at the local sampling obtained Extract Scale invariant features transform SIFT feature on block, obtain the SIFT feature set of every image to be identified；

Step 7, based on the visual dictionary being made up of vision word obtained in described step 2, utilizes described step 3 (ratio Such as formula (7)-formula (11)), each SIFT feature of every image to be identified is carried out Global Topological coding；

Step 8, similar to described step 4, utilize the image expression that formula (12) defines, the institute to every image to be identified The Global Topological coding having SIFT feature carries out maximum aggregation operator, generates the image expression of image to be identified；

Step 9, the image expression of the image to be identified described step 8 obtained sends into the training mould that described step 5 generates Type is tested, thus obtains the other recognition result of target class in image to be identified.

From the foregoing, it will be observed that the present invention propose view-based access control model dictionary Global Topological express image classification method comprise training and Identifying two processes, next as a example by the vehicle detecting system in some monitoring scene, enforcement to the method is said Bright.Described vehicle detecting system may determine that in monitoring scene whether contain vehicle.

First having to collect substantial amounts of vehicle image (1000) and non-vehicle image (1000), these images are used for training Vehicle identification model.Training step S1 is:

Step S11: 1000 vehicle images (positive sample) and 1000 non-vehicle images (negative sample) are carried out SIFT spy Levy extraction, generate 2000 groups of SIFT feature, calculate containing 2000 SIFT feature with average often group, altogether extract 4000000 (2000 × 2000) individual SIFT feature；

Step S12: 4000000 SIFT feature carry out k-menas cluster operation, generates 1 and comprises 1000 visions The visual dictionary of word；

Step S13: in actual application scenarios, 3 vision word of the arest neighbors taking each SIFT local feature are entered Row Global Topological encodes, and code length is 3000 dimensions；

Step S14: the Global Topological coding of 2000 SIFT local features of every training image is carried out maximum gathering Operation, obtains the image expression of image, a length of 3000 dimensions；

Step S15: the image expression of all training images is sent in grader and is trained, obtain training pattern；

At cognitive phase, camera signals is accessed computer by capture card and carrys out collecting test picture, specifically identify step Rapid S2 is:

Step S21: input a test image, it is carried out SIFT feature and extracts operation, generate 1 group of SIFT feature, contain There are 2000 SIFT feature.

Step S22: in actual application scenarios, 3 vision word of the arest neighbors taking each SIFT local feature are entered Row Global Topological encodes, and code length is 3000 dimensions；

Step S23: the Global Topological coding of 2000 SIFT local features of every test image is carried out maximum gathering Operation, obtains the image expression of image, a length of 3000 dimensions；

Step S24: the image expression of the image to be identified step S23 obtained sends into the training pattern that step S15 generates Test, thus obtain the other recognition result of target class in image to be identified.

To sum up, the present invention proposes the Image Classfication Technology that a kind of effective view-based access control model dictionary Global Topological is expressed.This Invention is easily achieved, stable performance, it is possible to increase the intelligent monitor system understandability to monitoring scene, is intelligence of future generation simultaneously Key technology in energy video monitoring system.

Particular embodiments described above, has been carried out the purpose of the present invention, technical scheme and beneficial effect the most in detail Describe in detail bright, be it should be understood that the specific embodiment that the foregoing is only the present invention, be not limited to the present invention, all Within the spirit and principles in the present invention, any modification, equivalent substitution and improvement etc. done, should be included in the guarantor of the present invention Within the scope of protecting.

Claims (8)

1. the image classification method that a view-based access control model dictionary Global Topological is expressed, it is characterised in that the method includes following step Rapid:

Step 1, gathers several training images, and several training images carry out local sampling respectively, and at the local sampling obtained Extract Scale invariant features transform (SIFT) feature on block, thus obtain the SIFT feature set of training image；

Step 2, carries out cluster and generates multiple cluster centres the SIFT feature set obtained, and with described cluster centre as vision Word composition visual dictionaryWherein, C represents visual dictionary, its vision word tieed up by M D c_iComposition,Represent the subspace that M point in D dimension space forms；

Step 3, carries out Global Topological coding to each SIFT feature of every width training image；

Step 4, all SIFT feature to every width training imageGlobal Topological coding V₁, V₂..., V_NCarry out maximum aggregation operator, generate the image expression F of this training image；

Step 5, sends into the image expression of all training images in grader and is trained, generate training pattern；

Step 6, similar to described step 1, every image to be identified is carried out local sampling, and on the local sampling block obtained Extract Scale invariant features transform SIFT feature, obtain the SIFT feature set of every image to be identified；

Step 7, based on the visual dictionary being made up of vision word obtained in described step 2, utilizes described step 3 to treat every width Identify that each SIFT feature of image carries out Global Topological coding；

Step 8, similar to described step 4, the Global Topological coding of all SIFT feature of every image to be identified is carried out Big aggregation operator, generates the image expression of image to be identified；

Step 9, the training pattern that the image expression described step 5 of feeding of the image to be identified described step 8 obtained generates is entered Row test, thus obtain the other recognition result of target class in image to be identified；

Described step 3 farther includes three sub-steps:

Step 3.1, calculates the dependency between vision word in described visual dictionary；

Step 3.2, utilizes the arest neighbors to each SIFT feature of the dependency between the vision word obtained in described step 3.1 K vision word set up Global Topological model；

Step 3.3, the Global Topological model utilizing described step 3.2 to obtain carries out Global Topological coding to each SIFT feature；

In described step 3.2, it is assumed that the SIFT feature collection obtained is extracted for every width training image and is combined intoWherein, D is intrinsic dimensionality, and N is Characteristic Number, then for each feature x_j, its Global Topological Model representation is:

Δ (C, x_j, S) and=[Δ (c₁, x_j, S₁) ..., Δ (c_M, x_j, S_M)],

$\&ForAll;\mathrm{\&Delta;}(c_i,x_j,S_i)=\&lsqb;\mathrm{\&Delta;}(c_i,x_j,c_{i1}),...,\mathrm{\&Delta;}(c_i,x_j,c_{{\mathrm{iQ}}_i})\&rsqb;\&Element;R^{1\&times;Q_i},$

Wherein, Δ (c_i, x_j, S_i) it is vector x_j-c_iWith set S_iIn each related words c_ijWith c_iThe vectorial c constituted_ij-c_iIt Between angle:

$\mathrm{\&Delta;}(c_i,x_j,S_i)=[\mathrm{\&Delta;}(c_i,x_j,c_{i1}),...,\mathrm{\&Delta;}(c_i,x_j,c_{{\mathrm{iQ}}_i})],$

<x_j-c_i, c_ij-c_i> represent vector x_j-c_iWith vector c_ij-c_i Between inner product, " " represents the dot product between two vectors, | | | |₂Represent two norms of vector.

Method the most according to claim 1, it is characterised in that in described step 2, uses k-means clustering algorithm to carry out Cluster.

Method the most according to claim 1, it is characterised in that utilize independent increment formula algorithm based on distance and angle to obtain Dependency between vision word.

Method the most according to claim 3, it is characterised in that described independent increment algorithm based on distance and angle includes:

Firstly, for a certain vision word c_i, all of vision word arrives c according to it_iEuclidean distance arrange from the near to the remote Sequence, wherein, using nearest vision word as initial relevant vision word c_i1；

Secondly, except vision word c_iOther outer vision word check according to distance and angle criterion one by one, regard if a certain Feel word c_jBe satisfied by described distance and angle criterion through inspection, then this vision word is marked as multi view word c_ij, And add set S to_i, finally give vision word c_iThe set S of multi view word_i；

Finally, to all of vision word traversal processing, i.e. can get the related words set S of visual dictionary C.

Method the most according to claim 4, it is characterised in that wherein, described distance criterion is defined as:

||c_i-c_ij||₂＜ τ | | c_i-c_i1| | 2,

Wherein, τ is used for controlling c_iEuclidean distance to other vision word；

Described angle rule definition is:

Δ(c_i, c_ij, c_k) ＞ θ,

Wherein, S_iIt it is vision word c_iThe set of multi view word, and c_kRepresentative has been added to gather S_iIn vision list Word.

Method the most according to claim 4, it is characterised in that the related words set S of described visual dictionary C can represent For:

Wherein, D represents the dimension of each vision word, S_iIt it is vision word c_iThe set of multi view word, Q_iIt is S_iIn Vision word number.

Method the most according to claim 1, it is characterised in that in described step 3.3, utilizes following formula to each SIFT feature x_iCarry out Global Topological coding:

$\underset{V_i}{\mathrm{argmin}}\left|\right|x_i-{{\mathrm{SV}}_i}^T|{|}_2^2+\&lsqb;\mathrm{\&lambda;}T\left(V_i\right)+\mathrm{\&alpha;}||V_i\&CenterDot;\mathrm{\&Delta;}(C,x_i,S)|{|}_2^2\&rsqb;,$

Wherein, λ is penalty term coefficient, T (V_i) it is to V_iAny form of penalty term, α is the penalty term of Global Topological model, " " represents the dot product between two vectors, V_iIt is x_iGlobal Topological coding on S:

Wherein, (v_k1)_iIt it is each SIFT feature x_iIn vision word c_kFirst multi view word c_k1On response, successively Analogize.

Method the most according to claim 1, it is characterised in that in described step 4, carries out grasping maximum gathering according to the following formula Make, the image expression F of generation training image:

F=max_column[V₁ ^T, V₂ ^T..., V_N ^T]^T

Wherein, max_columnRepresent the every string for matrix and only retain its maximum.