CN101393608A - Visual object recognition method and apparatus based on manifold distance analysis - Google Patents

Visual object recognition method and apparatus based on manifold distance analysis Download PDF

Info

Publication number
CN101393608A
CN101393608A CNA200810225525XA CN200810225525A CN101393608A CN 101393608 A CN101393608 A CN 101393608A CN A200810225525X A CNA200810225525X A CN A200810225525XA CN 200810225525 A CN200810225525 A CN 200810225525A CN 101393608 A CN101393608 A CN 101393608A
Authority
CN
China
Prior art keywords
subspace
distance
shape
identified
stream shape
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CNA200810225525XA
Other languages
Chinese (zh)
Inventor
谢旭东
高跃
戴琼海
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tsinghua University
Original Assignee
Tsinghua University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tsinghua University filed Critical Tsinghua University
Priority to CNA200810225525XA priority Critical patent/CN101393608A/en
Publication of CN101393608A publication Critical patent/CN101393608A/en
Pending legal-status Critical Current

Links

Images

Landscapes

  • Image Analysis (AREA)

Abstract

The invention provides a method for identifying a visual object based on a manifold distance analysis, which comprises the following steps: constructing a manifold of an object to be identified according to data of the visual object to be identified, and constructing the manifold of a known object according to the data of a known visual object; reducing the dimensions of the manifold of the object to be identified and the manifold of the known object to subspaces; extracting the characteristic data in each subspace of the manifold of the object to be identified and the manifold of the known object; calculating the distance between the subspaces so as to obtain the distance between the manifold of the object to be identified and the manifold of the known object; and when the distance between the manifold of the object to be identified and the manifold of the known object meets the identification condition, confirming an object corresponding to the manifold of the known object as the visual object to be identified. The method compares and analyzes the distances between all the subspaces of the manifold of the object to be identified and the manifold of the known object, considers the whole position relations between the manifolds comprehensively, thus the method is not sensitive to noise data and has better identification effect of the visual object.

Description

A kind of visual object recognition methods and device based on manifold distance analysis
Technical field
The present invention relates to visual object identification field, particularly relate to a kind of visual object recognition methods and device based on manifold distance analysis.
Background technology
Pattern-recognition is a kind of from bulk information and data, on the basis of expertise and existing understanding, utilizes the method for computing machine and mathematical reasoning shape, pattern, curve, numeral, character format and figure to be finished automatically the process of identification.Visual object is discerned as one of main application fields of pattern-recognition, is just bringing into play more and more important effect such as aspects such as recognition of face, identifications.
In visual object identification field, the object of paying close attention to is all trained from a spot of sample usually.Along with the fast development of video frequency pick-up head and high capacity multi-medium data memory technology, be used for object pattern recognition data amount also in continuous growth, very big thereby the amount of images of the object that is used to train also becomes.In this case, visual object identification has developed into based on the training of an object picture set and the situation of seeking in an image library.The complicacy of the relation between the huge and different images data of media data has increased the difficulty of visual object identification, thereby the visual object recognition technology has been proposed new challenge.A lot of in recent years methods are paid close attention to and are used stream shape to visual object set carrying out modeling, by carrying out visual object identification apart from comparative analysis between the convection current shape.
Stream shape is meant a hypersurface in the abstract curved space, and it has the euclidean geometry structure in the part.A point in the stream shape can be represented with one group of particular value of variable element, and all formation stream shapes of these all points itself, variable element is called the coordinate that flows shape.Usually, the figure in the higher dimensional space that stream shape is made up of high dimensional data reduces the low dimension data that dimension obtains with the stream shape of higher-dimension and is called the subspace of flowing shape, flows point in the shape and be exactly the point in the higher dimensional space.If the single image sample is regarded as a point that flows in the shape, the subspace of flowing shape so is exactly the part of being made up of the part sample of stream shape with local linear feature.In visual object identification, base of recognition is the similarity of visual object, and the distance in a kind of simple similarity Free Region feature space defines.Specific to stream shape, define similarity with the distance between the stream shape exactly, the distance between the stream shape is more little, and similarity is just high more; Distance is big more, and similarity is low more.Regard the object picture set that is retrieved as a stream shape, the object picture that will be used to retrieve set is considered as other stream shape, and then visual object identification can be considered to find out the process of the destination object that similarity is the highest between a plurality of stream shapes.
The paper of by name " manifold distance based on the application in the recognition of face of image set " (Manifold-Manifold Distance With Application to Face Recognition based onImage Set) in U.S. electric and electronics engineers computer vision and the pattern-recognition international conference 2008 nd Annual Meeting collection (In Proceedings of IEEE International Conference on Computer Vision andPattern Recognition 2008) has been announced a kind of up-to-date manifold distance analysis method, this method is by the distance of definition stream shape mid point and point, distance between point and the subspace, distance between distance between subspace and the subspace and point and the stream shape, thereby obtain flowing the distance between shape and the stream shape, at last based on manifold distance analysis recognition visible sensation object.This method has higher susceptibility to media data, can effectively handle a large amount of, complicated visual object media data.But the manifold distance analysis method that this method provides is only paid close attention between two stream shapes the distance of close part, and with it as general manifold distance.Simplified manifold distance calculating though do like this, but because, the distance of close part is often very responsive to noise data between the stream shape, when in two stream shapes because the influence of noise data when making that it has nearer subspace, then the manifold distance computational analysis is easy to generate wrong result, thereby causes the mistake identification of visual object.
In a word, need the urgent technical matters that solves of those skilled in the art to be exactly: how can effectively reduce in the manifold distance analysis process the sensitivity of noise data, improve the visual object discrimination.
Summary of the invention
Technical matters to be solved by this invention provides a kind of visual object recognition methods and device based on manifold distance analysis, can effectively reduce in the manifold distance analysis process sensitivity to noise data, improves the visual object discrimination.
In order to address the above problem, the embodiment of the invention provides a kind of visual object recognition methods based on manifold distance analysis, comprising:
According to nonlinear object to be identified stream shape of visual object data configuration to be identified, and, according to the one or more nonlinear known object stream shapes of known visual object data configuration;
The subspace with described object to be identified stream shape and known object stream shape from the higher-dimension dimensionality reduction to low-dimensional;
Extract the characteristic of each subspace of described object to be identified stream shape, and, the characteristic of each subspace of described known object stream shape;
Calculate the distance of described object to be identified stream each subspace of shape and each subspace of known object stream shape, and then obtain the distance of described object to be identified stream shape and known object stream shape;
When the distance of described object to be identified stream shape and known object stream shape satisfied condition for identification, then the object with described known object stream shape correspondence was defined as waiting to look other visual object.
Preferably, the described subspace of shape from the higher-dimension dimensionality reduction to low-dimensional of will flowing is for to carry out dimensionality reduction by using local linear embedding grammar.
Preferably, the characteristic of described each subspace is the orthonormal basis of each subspace.
Preferably, described object to be identified stream each subspace of shape and known object flow the distance between the orthonormal basis that the distance of each subspace of shape is described subspace.
Preferably, the distance of described object to be identified stream shape and known object stream shape is the Hausdorff distance based on the distance of subspace.
Accordingly, the embodiment of the invention also provides a kind of visual object recognition device based on manifold distance analysis, comprising:
Stream shape constructing module is used for according to nonlinear object to be identified stream shape of visual object data configuration to be identified, and, according to the one or more nonlinear known object stream shapes of known visual object data configuration;
The subspace constructing module is used for the subspace from the higher-dimension dimensionality reduction to low-dimensional with described object to be identified stream shape and known object stream shape;
The characteristic extraction module is used to extract the characteristic of each subspace of described object to be identified stream shape, and, the characteristic of each subspace of described known object stream shape;
Distance calculation module is used to calculate the distance of described object to be identified stream each subspace of shape and each subspace of known object stream shape, and then obtains the distance of described object to be identified stream shape and known object stream shape;
The condition for identification judge module is used to judge whether the distance of described object to be identified stream shape and known object stream shape satisfies condition for identification;
The object identification module, be used for when the result of condition for identification judge module when being, then the object with described known object stream shape correspondence is defined as waiting to look other visual object.
Preferably, described distance calculation module further comprises:
Subspace distance calculation submodule is used to calculate the distance of described object to be identified stream each subspace of shape and each subspace of known object stream shape;
The manifold distance calculating sub module is used for the subspace distance according to the acquisition of subspace distance calculation submodule, calculates the distance of object to be identified stream shape and known object stream shape.
Preferably, the described subspace of shape from the higher-dimension dimensionality reduction to low-dimensional of will flowing is for to carry out dimensionality reduction by using local linear embedding grammar.
Preferably, the characteristic of described each subspace is the orthonormal basis of each subspace.
Preferably, described object to be identified stream each subspace of shape and known object flow the distance between the orthonormal basis that the distance of each subspace of shape is described subspace.
Preferably, the distance of described object to be identified stream shape and known object stream shape is the Hausdorff distance based on the distance of subspace.
Compared with prior art, the present invention has the following advantages:
At first, the present invention compares analysis to the distance of all subspaces of object to be identified stream shape and known object stream shape, taken all factors into consideration the integral position relation between the stream shape, therefore insensitive to noise data, still can obtain effect preferably existing under the situation of certain noise, and the comparative analysis of distance is more effective between the convection current shape.
Secondly, the present invention extracts the characteristic of this subspace by the orthonormal basis that calculates each subspace, and then the distance between the calculating subspace, compare with other subspace distance calculating method, only be applicable to distance calculation under the two sub spaces dimension same cases as major component angle (Principal angels) method, even these method two sub spaces dimensions are unequal, still can effectively calculate the distance between the subspace, and use the characteristic computing method of orthonormal basis extraction subspace easy, computation complexity is low, simplicity of design is easy to realize.
Description of drawings
Fig. 1 is the flow chart of steps of a kind of visual object recognition methods embodiment 1 based on manifold distance analysis of the present invention;
Fig. 2 is the flow chart of steps of a kind of visual object recognition methods embodiment 2 based on manifold distance analysis of the present invention;
Fig. 3 is the structured flowchart of a kind of visual object recognition device embodiment based on manifold distance analysis of the present invention;
Fig. 4 is that the present invention uses the flow chart of steps that device embodiment shown in Figure 2 carries out visual object identification.
Embodiment
For above-mentioned purpose of the present invention, feature and advantage can be become apparent more, the present invention is further detailed explanation below in conjunction with the drawings and specific embodiments.
One of core idea of the embodiment of the invention is, compare analysis by distance to all subspaces of object to be identified stream shape and known object stream shape, take all factors into consideration the integral position relation between the stream shape, make the visual object identifying insensitive, obtain recognition effect preferably noise data.
With reference to figure 1, show the flow chart of steps of a kind of visual object recognition methods embodiment 1 based on manifold distance analysis of the present invention, specifically can may further comprise the steps:
Step 101: according to nonlinear object to be identified stream shape of visual object data configuration to be identified, according to the one or more nonlinear known object stream shapes of known visual object data configuration;
Stream shape is the notion of modern mathematics, and in actual applications, stream shape can be considered as closely looking like the object in Euclidean space or other simple relatively spaces.For example: people thought once that the earth was smooth, because we are very little with respect to the earth, this is a understandable illusion.So a desirable mathematical ball also resembles the plane of a linearity in enough little zone, this makes it become a local linear stream shape.
In visual object identification, one group of associated picture of a common object is regarded as a stream shape.This is because same object can form multiple different image under different distance, different directions or different attitude and intensity of illumination.The set of an object all images can regard with position, yardstick, attitude, illumination etc. to be higher dimensional space stream shape of parameter as, and this group image because all be about same object have a stronger correlativity.If each pixel is all corresponding to the one dimension in the space, piece image just can be regarded a point in the higher-dimension image abstraction space as so, and object set of all images in different directions is exactly a continuous stream shape in the image space.
Therefore, in the present embodiment, can be with the associated picture structure object to be identified stream shape of one group of same visual object to be identified; And constructing one or more nonlinear known object stream shapes for the associated picture of other one or more groups the known visual object that is used for discerning.
Need to prove that a stream shape can be constructed by different way, as atlas, subsidize, zero of a function collection etc., when using, those skilled in the art can select suitable method as required, the present invention need not make restriction to this.
Step 102: the subspace from the higher-dimension dimensionality reduction to low-dimensional with described object to be identified stream shape and known object stream shape;
In the visual object identifying, often face huge high dimensional data amount.These high dimensional datas provide extremely abundant, the detailed information of relevant description object on the one hand, and but then, these data comprise many redundancies usually.Under normal conditions, at first the dimension of data is reduced to a reasonably size, the original information of reservation as much as possible simultaneously, and then the data behind the dimensionality reduction are handled is an efficient ways.But will be under the condition that guarantees the data message sufficiently complete yojan data set reasonably, be a stern challenge.Used linear method (as the linear principal component analysis (PCA) (PCA) in the dimension yojan etc.) in the past mostly, came dimensionality reduction by the linear combination of feature, its essence be data projection to linear subspaces, this method is simple relatively and calculate easily.But because the stream shape in the visual object identification is non-linear, so linear method can not be handled the data of huge high dimensional nonlinear stream shape effectively.Local linear embedding inlay technique (Locally Linear Embedding, LLE) be a kind of in the multiple Nonlinear Dimension Reduction method, this method mainly utilizes local linearity to approach the non-linear of the overall situation, keep the geometry of part constant, provide whole information by overlapped local neighborhood, thereby keep whole geometric properties.The LLE method adopts the matrix X that is made up of N D dimensional vector D * NAs input, output then be one by N d dimensional vector (d＜＜the matrix Y that D) forms D * NThe k of matrix Y row are corresponding is k row among the matrix X.
Because the LLE algorithm can calculate low-dimensional linear subspaces preferably, present embodiment uses the method that described object to be identified stream shape and known object are flowed the subspace of shape from the higher-dimension dimensionality reduction to low-dimensional.Certainly, those skilled in the art also can adopt other Nonlinear Dimension Reduction method apart from reflection method, laplacian eigenmaps method etc. manifold of higher dimension to be carried out dimensionality reduction as protecting, and the present invention need not make restriction to this.
Step 103: extract the characteristic of each subspace of described object to be identified stream shape, and the characteristic of each subspace of described known object stream shape;
The characteristic of one sub spaces has reflected the feature of a sub spaces preferably, can be used as the representative data of this sub spaces.The mode of the orthonormal basis of present embodiment by calculating each subspace is extracted the characteristic of each subspace of described object to be identified stream shape and the characteristic of each subspace of known object stream shape.Certainly, those skilled in the art also can adopt other mode that is suitable for as required.
Step 104: the distance of calculating described object to be identified stream each subspace of shape and each subspace of known object stream shape;
Present embodiment calculates the distance of subspace according to the orthonormal basis of each subspace that obtains in the step 103.With a sub spaces of object to be identified stream shape and a sub spaces of a known object stream shape is example, supposes object to be identified stream shape M 1With known object stream shape M 2Two sub spaces be respectively U and V, wherein U is
Figure A200810225525D00101
In the m n-dimensional subspace n, V is
Figure A200810225525D00102
In the n n-dimensional subspace n, the orthonormal basis of U is u 1, u 2..., u m, the orthonormal basis of V is v 1, v 2..., v n, then the distance between subspace U and the V is the distance between their orthonormal basis, its computing method are:
d ( U , V ) = max ( m , n ) - Σ i = 1 m Σ j = 1 n ( u i T v j )
If known object stream shape is a plurality of, can adopt the distance between same procedure calculating object to be identified stream shape and the different known object stream shape subspace.
Need to prove that those skilled in the art can adopt the distance between other method calculating subspace, as kernel method, the present invention need not make restriction to this.
Step 105:, obtain the distance of object to be identified stream shape and known object stream shape according to the distance of described object to be identified stream each subspace of shape and each subspace of known object stream shape;
This step is with d (U p, V q) represent that the object to be identified that is calculated by step 104 flows shape M 1Subspace and known object stream shape M 2The distance of subspace, U wherein pAnd V qBe respectively from described M 1And M 2The p of two flow patterns and q sub spaces are used Hao Siduofu (Hausdorff) distance analysis method convection M 1And M 2Carry out distance calculation.
Hausdorff distance is a kind of two minimax (max-min) distances on the point set that are defined in, given limited two point set A and B, and the Hausdorff distance definition between A and the B is:
H(A，B)＝max[h(A，B)，h(B，A)]
In the formula,
h ( A , B ) = max a ∈ A min b ∈ B | | a - b | |
(A B) claims oriented Hausdorff distance to h.If h (A, B)=d, then the distance of point at least one point in B among each A is not more than d, and for some (at least one) point among the A, this distance is d just, and these points are exactly " least match point ".So Hausdorff must make it it minimize as similarity measurement apart from the least similarity degree that has characterized between two point sets.
Present embodiment uses Hausdorff distance analysis method, apart from replacing the point set distance, calculates described M with the subspace 1And M 2The distance of two flow patterns, concrete grammar is as follows:
D ( M 1 , M 2 ) = max U p ∈ M 1 min V q ∈ M 2 d ( U p , V q )
Object to be identified stream shape M 1With other known object stream shape M tDistance calculation can carry out according to said method.Wherein, the span of t is from 2 to n+1, and n is the number of known object stream shape.
Those skilled in the art also can calculate the distance that flows between the shape according to its other method of being familiar with, and the present invention need not make restriction to this.
Step 106: whether the distance of judging described object to be identified stream shape and known object stream shape satisfies condition for identification, if the object that then known object that satisfies condition is flowed the shape correspondence is defined as visual object to be identified; If not, then the object of known object stream shape correspondence is not a visual object to be identified.
Present embodiment is with the standard of the distance between the stream shape as stream of measurements shape similarity, and the more little then similarity of distance is high more between two stream shapes; Otherwise distance is big more, and similarity is low more.When similarity meets some requirements, promptly flow distance between the shape when meeting certain condition for identification, think that then known object stream shape is similar with object to be identified stream shape, the object that known object flows the shape correspondence is a visual object to be identified; If not, then the object of known object stream shape correspondence is not a visual object to be identified.Wherein, condition for identification situation according to actual needs rationally is provided with by those skilled in the art, and the present invention need not make restriction to this.
Need to prove, this method embodiment is for simple description, it is expressed as a series of combination of actions, but those skilled in the art should know, the present invention is not subjected to the restriction of described sequence of movement, because according to the present invention, some step can adopt other orders or carry out simultaneously.Secondly, those skilled in the art also should know, the embodiment described in the instructions all belongs to preferred embodiment, and related action and module might not be that the present invention is necessary.
With reference to figure 2, show the flow chart of steps of a kind of visual object recognition methods embodiment 2 based on manifold distance analysis of the present invention, this embodiment is applied to the CMUMobo database with visual object recognition methods of the present invention and carries out recognition of face.MoBo (Motionof Body) database of CMU (Carnegie Mellon University) is the database that is used for recognition of face research that CMU university robot research is provided, this database has comprised 96 sections video sequences of 24 objects, and each video sequence comprises 300 two field pictures.The size of each facial image is 30 * 30.For everyone face object, use one of them as searching object, other all data are as test data.
Present embodiment specifically can may further comprise the steps:
Step 201: will become nonlinear people's face stream shape to be identified with one group of relevant associated picture data configuration of people's face to be identified, other every group of relevant people face data configuration becomes a plurality of nonlinear known person face stream shapes;
Step 202: use the LLE method with described people's face stream shape to be identified and the subspace of known person face stream shape from the higher-dimension dimensionality reduction to low-dimensional;
Step 203: calculate the orthonormal basis of each subspace of described people's face stream shape to be identified, and the orthonormal basis of each subspace of described known person face stream shape;
This step is extracted its characteristic by the orthonormal basis that calculates each subspace.
Step 204: calculate the distance between the orthonormal basis that people's face to be identified stream each subspace of shape and known person face flow each subspace of shape;
Computing method are:
d ( U p , V q ) = max ( m , n ) - Σ i = 1 m Σ j = 1 n ( u i T v j )
Wherein, d (U p, V q) subspace of expression people's face stream to be identified shape and the distance that certain known person face flows the subspace of shape, wherein U pAnd V qBe respectively p and q sub spaces from described people's face stream shape to be identified and known person face flow pattern, U pFor
Figure A200810225525D00122
In the m n-dimensional subspace n, V qFor
Figure A200810225525D00123
In the n n-dimensional subspace n, U pOrthonormal basis be u 1, u 2..., u m, V qOrthonormal basis be v 1, v 2..., v n
Step 205:, obtain the Hausdorff distance of people's face stream shape to be identified and known person face stream shape according to the distance of described people's face to be identified stream each subspace of shape and each subspace of known person face stream shape;
Use Hausdorf distance analysis method, with the distance replacement point set distance of subspace, calculate the Hausdorff distance between the flow pattern, concrete grammar is as follows:
D ( M 1 , M t ) = max U p ∈ M 1 - min V q ∈ M t d ( U p , V q ) , t = ( 2 , . . . , n + 1 )
Wherein, the span of t is from 2 to n+1, and n is the number of known object stream shape.
Step 206: whether the Hausdorff distance of judging described people's face stream shape to be identified and known person face stream shape satisfies condition for identification, if the people's face that then the known person face that satisfies condition is flowed the shape correspondence is defined as people's face object to be identified; If not, then people's face of known person face stream shape correspondence is not people's face object to be identified.
Present embodiment be applied to CMU MoBo database recognition of face test the result and with being compared as follows shown in the table of traditional face recognition algorithms:
Method of the present invention Eigenface Fisherface
Discrimination 89.3％ 80.5％ 86.7％
As can be seen, present embodiment is compared with classic method on the average recognition rate of recognition of face, can obtain better effect.
Need to prove that what present embodiment was described is the concrete application of visual object recognition methods in the CMUMoBo database of embodiment 1, therefore description is comparatively simple, and relevant portion gets final product referring to embodiment 1.
With reference to figure 3, show the structured flowchart of a kind of visual object recognition device embodiment based on manifold distance analysis of the present invention, can comprise:
Stream shape constructing module 301 is used for according to nonlinear object to be identified stream shape of visual object data configuration to be identified, and, according to the one or more nonlinear known object stream shapes of known visual object data configuration;
Subspace constructing module 302 is used for the subspace from the higher-dimension dimensionality reduction to low-dimensional with described object to be identified stream shape and known object stream shape;
Preferably, present embodiment carries out dimensionality reduction by using local linear embedding grammar convection current shape.
Characteristic extraction module 303 is used to extract the characteristic of each subspace of described object to be identified stream shape, and, the characteristic of each subspace of described known object stream shape;
Preferably, the characteristic of described each subspace is the orthonormal basis of each subspace.
Distance calculation module 304 is used to calculate the distance of described object to be identified stream each subspace of shape and each subspace of known object stream shape, and obtains the distance of described object to be identified stream shape and known object stream shape;
Preferably, described object to be identified stream each subspace of shape and known object flow the distance between the orthonormal basis that the distance of each subspace of shape is described subspace.
Preferably, the distance of described object to be identified stream shape and known object stream shape is the Hausdorff distance based on the distance of subspace.
Condition for identification judge module 305 is used to judge whether the distance of described object to be identified stream shape and known object stream shape satisfies condition for identification;
Object identification module 306 is if the result who is used for the condition for identification judge module is for being that then the object with described known object stream shape correspondence is defined as waiting to look other visual object.
Preferably, distance calculation module 304 can further include:
Subspace distance calculation submodule 3041 is used to calculate the distance of described object to be identified stream each subspace of shape and each subspace of known object stream shape;
Manifold distance calculating sub module 3042 is used for the subspace distance according to the acquisition of subspace distance calculation submodule, calculates the distance of object to be identified stream shape and known object stream shape.
With reference to figure 4, show the present invention and use the flow chart of steps that device embodiment shown in Figure 3 carries out visual object identification, specifically can may further comprise the steps:
Step 401: stream shape constructing module is according to nonlinear object to be identified stream shape of visual object data configuration to be identified, according to the one or more nonlinear known object stream shapes of known visual object data configuration;
Stream shape constructing module is constructed not homogeneous turbulence shape according to the associated picture of different objects.
Step 402: the subspace constructing module is the subspace from the higher-dimension dimensionality reduction to low-dimensional with described object to be identified stream shape and known object stream shape;
In the present embodiment, the subspace constructing module is used local linear embedding grammar convection current shape and is carried out dimensionality reduction, with the subspace of stream shape from the higher-dimension dimensionality reduction to low-dimensional.Those skilled in the art also can adopt other dimension reduction method according to actual needs, and the present invention need not make restriction to this.
Step 403: the characteristic extraction module extracts the characteristic of each subspace of described object to be identified stream shape, and the characteristic of each subspace of known object stream shape;
The characteristic extraction module of present embodiment extracts the characteristic of subspace by the orthonormal basis that calculates each subspace, and those skilled in the art also can according to circumstances use other method to extract the characteristic of subspace.
Step 404: the subspace distance calculation submodule of distance calculation module calculates the distance of described object to be identified stream each subspace of shape and each subspace of known object stream shape;
Calculate the orthonormal basis of the subspace that subspace distance calculation submodule calculates according to the characteristic extraction module described object to be identified stream each subspace of shape and known object flow each subspace of shape distance, and give the manifold distance calculating sub module and handle, concrete account form can for:
d ( U p , V q ) = max ( m , n ) - Σ i = 1 m Σ j = 1 n ( u i T v j )
Wherein, d (U p, V q) subspace of expression object to be identified stream shape and the distance that a certain known object flows the subspace of shape, wherein U pAnd V qBe respectively p and q sub spaces from described object to be identified stream shape and known object flow pattern, U pFor
Figure A200810225525D00152
In the m n-dimensional subspace n, V qFor
Figure A200810225525D00153
In the n n-dimensional subspace n, U pOrthonormal basis be u 1, u 2..., u m, V qOrthonormal basis be v 1, v 2..., v n
Step 405: the subspace distance that the manifold distance calculating sub module of distance calculation module obtains according to subspace distance calculation submodule, calculate the distance that object to be identified stream shape and known object flow shape;
In the present embodiment, the manifold distance calculating sub module is used Hausdorf distance analysis method, apart from replacing the point set distance, calculates the distance between the flow pattern with the subspace, and concrete grammar is as follows:
D ( M 1 , M t ) = max U p ∈ M 1 - min V q ∈ M t d ( U p , V q ) , t = ( 2 , . . . , n + 1 )
Wherein, the span of t is from 2 to n+1, and n is the number of known object stream shape, d (U p, V q) be the distance between the subspace.
Step 406: the condition for identification judge module judges whether the distance of described object to be identified stream shape and known object stream shape satisfies condition for identification;
The condition for identification judge module is judged the result of calculation D (M of manifold distance calculating sub module 1, M t) whether satisfy condition for identification.
Step 407: when the result of condition for identification judge module when being, the object identification module is identified as visual object to be identified with the object of the known object stream shape correspondence that satisfies condition; When the result of condition for identification judge module for not the time, the object of object identification module identification known object stream shape correspondence is not a visual object to be identified.
Because embodiment shown in Figure 4 can correspondence be applicable among the aforesaid visual object recognition methods embodiment that so description is comparatively simple, not detailed part can be referring to the description of this instructions front appropriate section.
More than a kind of visual object recognition methods and device based on manifold distance analysis provided by the present invention is described in detail, used specific case in the literary composition core concept of the present invention and embodiment are set forth, the explanation of above embodiment just is used for helping to understand method of the present invention and core concept thereof; Simultaneously, for one of ordinary skill in the art, according to thought of the present invention, the part that all can change in specific embodiments and applications, in sum, this description should not be construed as limitation of the present invention.

Claims (11)

1, a kind of visual object recognition methods based on manifold distance analysis is characterized in that, may further comprise the steps:
According to nonlinear object to be identified stream shape of visual object data configuration to be identified, and, according to the one or more nonlinear known object stream shapes of known visual object data configuration;
The subspace with described object to be identified stream shape and known object stream shape from the higher-dimension dimensionality reduction to low-dimensional;
Extract the characteristic of each subspace of described object to be identified stream shape, and, the characteristic of each subspace of described known object stream shape;
Calculate the distance of described object to be identified stream each subspace of shape and each subspace of known object stream shape, and then obtain the distance of described object to be identified stream shape and known object stream shape;
When the distance of described object to be identified stream shape and known object stream shape satisfied condition for identification, then the object with described known object stream shape correspondence was defined as waiting to look other visual object.
2, method according to claim 1 is characterized in that, the described subspace of shape from the higher-dimension dimensionality reduction to low-dimensional of will flowing is for to carry out dimensionality reduction by using local linear embedding grammar.
3, method according to claim 1 and 2 is characterized in that, the characteristic of described each subspace is the orthonormal basis of each subspace.
4, method according to claim 3 is characterized in that, the distance that described object to be identified stream each subspace of shape and known object flow each subspace of shape is the distance between the orthonormal basis of described subspace.
5, method according to claim 4 is characterized in that, the distance of described object to be identified stream shape and known object stream shape is the Hausdorff distance based on the distance of subspace.
6, a kind of visual object recognition device based on manifold distance analysis is characterized in that, comprising:
Stream shape constructing module is used for according to nonlinear object to be identified stream shape of visual object data configuration to be identified, and, according to the one or more nonlinear known object stream shapes of known visual object data configuration;
The subspace constructing module is used for the subspace from the higher-dimension dimensionality reduction to low-dimensional with described object to be identified stream shape and known object stream shape;
The characteristic extraction module is used to extract the characteristic of each subspace of described object to be identified stream shape, and, the characteristic of each subspace of described known object stream shape;
Distance calculation module is used to calculate the distance of described object to be identified stream each subspace of shape and each subspace of known object stream shape, and then obtains the distance of described object to be identified stream shape and known object stream shape;
The condition for identification judge module is used to judge whether the distance of described object to be identified stream shape and known object stream shape satisfies condition for identification;
The object identification module, be used for when the result of condition for identification judge module when being, then the object with described known object stream shape correspondence is defined as waiting to look other visual object.
7, device according to claim 6 is characterized in that, described distance calculation module further comprises:
Subspace distance calculation submodule is used to calculate the distance of described object to be identified stream each subspace of shape and each subspace of known object stream shape;
The manifold distance calculating sub module is used for the subspace distance according to the acquisition of subspace distance calculation submodule, calculates the distance of object to be identified stream shape and known object stream shape.
According to claim 6 or 7 described devices, it is characterized in that 8, the described subspace of shape from the higher-dimension dimensionality reduction to low-dimensional of will flowing is for to carry out dimensionality reduction by using local linear embedding grammar.
According to claim 6 or 7 described devices, it is characterized in that 9, the characteristic of described each subspace is the orthonormal basis of each subspace.
According to claim 6 or 7 described devices, it is characterized in that 10, the distance that described object to be identified stream each subspace of shape and known object flow each subspace of shape is the distance between the orthonormal basis of described subspace.
According to claim 6 or 7 described devices, it is characterized in that 11, the distance of described object to be identified stream shape and known object stream shape is the Hausdorff distance based on the distance of subspace.
CNA200810225525XA 2008-11-04 2008-11-04 Visual object recognition method and apparatus based on manifold distance analysis Pending CN101393608A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CNA200810225525XA CN101393608A (en) 2008-11-04 2008-11-04 Visual object recognition method and apparatus based on manifold distance analysis

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CNA200810225525XA CN101393608A (en) 2008-11-04 2008-11-04 Visual object recognition method and apparatus based on manifold distance analysis

Publications (1)

Publication Number Publication Date
CN101393608A true CN101393608A (en) 2009-03-25

Family

ID=40493896

Family Applications (1)

Application Number Title Priority Date Filing Date
CNA200810225525XA Pending CN101393608A (en) 2008-11-04 2008-11-04 Visual object recognition method and apparatus based on manifold distance analysis

Country Status (1)

Country Link
CN (1) CN101393608A (en)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103824058A (en) * 2014-02-26 2014-05-28 杨勇 Face recognition system and method based on locally distributed linear embedding algorithm
CN104156698A (en) * 2014-07-25 2014-11-19 清华大学 Human face recognition method and device
CN106127112A (en) * 2016-06-15 2016-11-16 北京工业大学 Data Dimensionality Reduction based on DLLE model and feature understanding method
CN106780462A (en) * 2016-12-14 2017-05-31 广州大学 A kind of rapid copy image detecting method based on SPD manifolds
CN106980840A (en) * 2017-03-31 2017-07-25 北京小米移动软件有限公司 Shape of face matching process, device and storage medium
CN107066943A (en) * 2017-03-06 2017-08-18 中国科学院信息工程研究所 A kind of method for detecting human face and device
WO2020052480A1 (en) * 2018-09-12 2020-03-19 北京三快在线科技有限公司 Unmanned driving behaviour decision making and model training

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103824058A (en) * 2014-02-26 2014-05-28 杨勇 Face recognition system and method based on locally distributed linear embedding algorithm
CN104156698A (en) * 2014-07-25 2014-11-19 清华大学 Human face recognition method and device
CN104156698B (en) * 2014-07-25 2017-10-03 清华大学 Face identification method and device
CN106127112A (en) * 2016-06-15 2016-11-16 北京工业大学 Data Dimensionality Reduction based on DLLE model and feature understanding method
CN106780462A (en) * 2016-12-14 2017-05-31 广州大学 A kind of rapid copy image detecting method based on SPD manifolds
CN107066943A (en) * 2017-03-06 2017-08-18 中国科学院信息工程研究所 A kind of method for detecting human face and device
CN106980840A (en) * 2017-03-31 2017-07-25 北京小米移动软件有限公司 Shape of face matching process, device and storage medium
WO2020052480A1 (en) * 2018-09-12 2020-03-19 北京三快在线科技有限公司 Unmanned driving behaviour decision making and model training
US11983245B2 (en) 2018-09-12 2024-05-14 Beijing Sankuai Online Technology Co., Ltd Unmanned driving behavior decision-making and model training

Similar Documents

Publication Publication Date Title
Gao et al. Deep label distribution learning with label ambiguity
Liu et al. $ p $-Laplacian regularized sparse coding for human activity recognition
Kalsum et al. Emotion recognition from facial expressions using hybrid feature descriptors
Kumar et al. Extraction of informative regions of a face for facial expression recognition
CN101393608A (en) Visual object recognition method and apparatus based on manifold distance analysis
Shi et al. Training DCNN by combining max-margin, max-correlation objectives, and correntropy loss for multilabel image classification
CN102968635B (en) Image visual characteristic extraction method based on sparse coding
Perakis et al. Feature fusion for facial landmark detection
Elpeltagy et al. Multi‐modality‐based Arabic sign language recognition
Huang et al. DropRegion training of inception font network for high-performance Chinese font recognition
Zhou et al. Learning-based scientific chart recognition
Tan et al. L1-norm latent SVM for compact features in object detection
Yang et al. Parsing 3D motion trajectory for gesture recognition
Wang et al. Neural koopman pooling: Control-inspired temporal dynamics encoding for skeleton-based action recognition
Xie et al. Progressive shape-distribution-encoder for learning 3D shape representation
Lin et al. Kernel-based representation for 2D/3D motion trajectory retrieval and classification
Henderson Analysis of engineering drawings and raster map images
Gou et al. Representation-based classification methods with enhanced linear reconstruction measures for face recognition
Yao [Retracted] Application of Higher Education Management in Colleges and Universities by Deep Learning
Yahia et al. Off-lexicon online Arabic handwriting recognition using neural network
CN111144469B (en) End-to-end multi-sequence text recognition method based on multi-dimensional associated time sequence classification neural network
Uddin et al. Recognition of printed Urdu ligatures using convolutional neural networks
CN105740878A (en) Symbol recognition method and device of geographic information system
CN115830637B (en) Method for re-identifying blocked pedestrians based on attitude estimation and background suppression
Cao Face recognition robot system based on intelligent machine vision image recognition

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
C02 Deemed withdrawal of patent application after publication (patent law 2001)
WD01 Invention patent application deemed withdrawn after publication

Open date: 20090325